Tensor Network Theory: A formalization blueprint

19 Matrix Product Unitaries

This chapter develops the algebraic prerequisites for matrix product unitaries. The first result supplies the corrected nil-matrix bound used in the blocking argument of [ CPGSV17 ] .

Theorem 19.1 Dimension bound for a nilpotent Lie action

Let a Lie algebra \(L\) act nilpotently on a finite-dimensional vector space \(V\) over a field \(K\). Write \(C_r(L,V)\) for the \(r\)th term of the lower central series of this action. Then

\begin{align} C_{\dim _K(V)}(L,V) & = 0. \end{align}
Proof

As long as \(C_r(L,V)\) is nonzero, nilpotency prevents two consecutive terms from being equal: equality would make every later term equal, in contradiction with eventual vanishing. Thus each nonzero step strictly lowers dimension, and at most \(\dim _K(V)\) such steps are possible.

Theorem 19.2 Finite-dimensional nil-matrix theorem

Let \(V\) be a finite-dimensional vector space over a field \(K\), and let \(A\subseteq \operatorname {End}_K(V)\) be a not necessarily unital associative subalgebra. If every element of \(A\) is nilpotent, then every ordered product of \(\dim _K(V)\) elements of \(A\) is zero. The empty product in dimension zero is also zero because the identity endomorphism of the zero space is zero.

Proof

Regard \(A\) as a Lie algebra under the commutator bracket. Engel’s theorem makes \(V\) a nilpotent Lie module, so the dimension bound gives

\begin{align} C_{\dim _K(V)}(A,V) & = 0. \end{align}

For every nonnegative integer \(r\), every ordered family \(a_1,\ldots ,a_r\) in \(A\), and every vector \(v\) in \(V\), induction on \(r\) gives

\begin{align} (a_1\cdots a_r)v & \in C_r(A,V). \end{align}

Taking \(r=\dim _K(V)\) proves that the product annihilates every vector and is therefore zero.

Theorem 19.3 Nil matrix subalgebras

Let \(A\) be a not necessarily unital associative subalgebra of \(n\times n\) matrices over a field. If every matrix in \(A\) is nilpotent, then every ordered product of \(n\) matrices from \(A\) vanishes.

Proof

Transport the matrix algebra to the endomorphisms of the coordinate space through the standard algebra equivalence. Nilpotency, membership, and the ordered product are preserved, so the finite-dimensional nil-matrix theorem applies; injectivity of the equivalence returns the zero matrix.

The length \(n\) is a weak bound, not a strict bound below \(n\); the latter is false for strictly upper-triangular matrices. This corrects the intermediate strict inequality in [ CPGSV17 ] without changing its final strict \(D^4\) estimate.

19.1 The matrix product unitary condition

Let \(\mathcal U=\{ U^{ij}\} _{i,j=0}^{d-1}\) be a matrix product operator tensor of bond dimension \(D\), and write \(U^{(N)}\) for the periodic operator obtained by closing a chain of \(N\) copies of \(\mathcal U\).

Theorem 19.1.1 Unitarity under simultaneous reindexing
#

Let \(I\) and \(J\) be finite sets, let \(e\colon I\to J\) be a bijection, and let \(A\) be a unitary matrix indexed by \(I\). The matrix indexed by \(J\) whose \((j,k)\) entry is

\begin{align} A_{e^{-1}(j),e^{-1}(k)} \end{align}

is unitary.

Proof

Simultaneous reindexing preserves the identity and matrix multiplication, and it commutes with the conjugate transpose. Reindex the equation \(AA^\dagger =\mathbb {1}\).

Definition 19.1.2 Matrix product unitary tensor

The tensor \(\mathcal U\) generates matrix product unitaries if \(U^{(N)}\) is unitary for every integer \(N{\gt}1\). This is the matrix product unitary condition in [ CPGSV17 , Section III, equation (10) ] .

Lemma 19.1.3 First matrix product unitarity equation

If \(\mathcal U\) generates matrix product unitaries and \(N{\gt}1\), then

\begin{align} U^{(N)}\bigl(U^{(N)}\bigr)^\dagger & = \mathbb {1}. \end{align}
Proof

Expand the unitary-group condition for \(U^{(N)}\).

Lemma 19.1.4 Second matrix product unitarity equation

If \(\mathcal U\) generates matrix product unitaries and \(N{\gt}1\), then

\begin{align} \bigl(U^{(N)}\bigr)^\dagger U^{(N)} & = \mathbb {1}. \end{align}

This is the unitarity equation displayed in [ CPGSV17 , Section III, equation (12) ] .

Proof

Use the equivalent left-multiplied characterization of a unitary matrix.

Theorem 19.1.5 Composition preserves matrix product unitarity

Let \(\mathcal U\) and \(\mathcal V\) be matrix product unitary tensors with the same physical dimension. Contracting the output index of \(\mathcal U\) with the input index of \(\mathcal V\) gives a matrix product unitary tensor whose periodic operator is \(U^{(N)}V^{(N)}\), in that order. This is the concatenation tensor \(\mathcal W\) in the proof of [ CPGSV17 , Theorem IV.6(ii) ] .

Proof

For every \(N{\gt}1\), both \(U^{(N)}\) and \(V^{(N)}\) are unitary. The defining contraction of the product tensor gives the periodic operator \(U^{(N)}V^{(N)}\), which is unitary because the unitary matrices form a group under multiplication.

Lemma 19.1.6 Physical blocking preserves matrix product unitarity

Suppose that \(\mathcal U\) generates matrix product unitaries. For every positive integer \(L\), the tensor \(\mathcal U^{[L]}\) obtained by blocking \(L\) adjacent sites also generates matrix product unitaries. This is the blocking operation of [ CPGSV17 , Definition II.5 and equation (9) ] ; preservation of the MPU property under blocking is used in the proof of [ CPGSV17 , Proposition III.3 ] .

Proof

At system size \(N\), closing the blocked tensor gives \(U^{(NL)}\), up to the canonical bijection between \(N\) blocked configurations and \(NL\) unblocked configurations. If \(N{\gt}1\) and \(L{\gt}0\), then \(NL{\gt}1\), so \(U^{(NL)}\) is unitary. Simultaneous reindexing preserves unitarity.

19.2 Simple matrix product unitary tensors

Definition 19.2.1 Double layer

For an MPO tensor \(\mathcal U\), let \(\mathcal W\) be the double-layer tensor obtained by placing the physical adjoint above \(\mathcal U\) and contracting the adjacent physical legs:

\begin{align} \mathcal W^{ik} & =\sum _j (\mathcal U^\sharp )^{ij}\otimes \mathcal U^{jk}. \end{align}

The bond space of \(\mathcal W\) is the product of the two original bond spaces. The physical adjoint exchanges and conjugates the physical indices without reversing the order of the virtual chain, exactly as in the barred upper layer of Figures II_Simple1.png and II_Simple2.png of [ CPGSV17 ] .

For every chain length \(N\), closing the double layer gives

\begin{align} W^{(N)} & =\bigl(U^{(N)}\bigr)^\dagger U^{(N)}. \end{align}
Proof

The operator family is multiplicative under the local product tensor, and the physical adjoint generates the conjugate transpose without reflecting the chain.

The physical adjoint is involutive. If \(\mathcal U\) generates matrix product unitaries, then so does \(\mathcal U^\sharp \), and for every chain length \(N\),

\begin{align} (\mathcal U^\sharp )^\sharp & =\mathcal U, & (\mathcal U^\sharp )^{(N)} & =\bigl(\mathcal U^{(N)}\bigr)^\dagger . \end{align}
Proof

Applying the physical adjoint twice conjugates every entry twice and exchanges the physical indices twice. Closing a physical-adjoint tensor exchanges the two physical configurations and conjugates the resulting scalar. Hence its periodic operator is the conjugate transpose. The first MPU unitarity equation for \(U^{(N)}\) is therefore the second unitarity equation for \((U^\sharp )^{(N)}\).

Order the left doubled bond as \((\alpha ,\gamma )\) and the right doubled bond as \((\beta ,\delta )\). Then the double layer whose lower tensor is \(\mathcal U^\sharp \) has entries

\begin{align} \mathcal W(\mathcal U^\sharp )^{ik}_{(\alpha ,\gamma ),(\beta ,\delta )} & =\sum _j \mathcal U^{ij}_{\alpha ,\beta }\, \overline{\mathcal U^{kj}_{\gamma ,\delta }}. \end{align}
Proof

Expand the double layer and then apply the physical adjoint to its upper factor. The two exchanges of physical indices leave the contracted index \(j\) in the second physical position of both factors. The product-index convention keeps \((\alpha ,\gamma )\) and \((\beta ,\delta )\) in the displayed order.

Definition 19.2.5 Simple matrix product unitary tensor

The tensor \(\mathcal U\) is simple if there are vectors \(a,b\) in the double-layer bond space such that, for all physical indices \(i,j,k,\ell \),

\begin{align} (a|W^{ij}|b) & =\delta _{ij}, \tag {simple1}\label{eq:mpu_simple1} \\ W^{ij}W^{k\ell } & =W^{ij}|b)(a|W^{k\ell }. \tag {simple2}\label{eq:mpu_simple2} \end{align}

Thus (11) closes one double-layer letter to the physical identity, while (12) inserts the rank-one operator \(|b)(a|\) between adjacent letters. These are precisely the two diagrams in [ CPGSV17 , Definition III.2 ] .

Theorem 19.2.6 Sequential contraction of a simple double layer

If \(\mathcal U\) is simple and \(N{\gt}1\), then

\begin{align} W^{(N)} & =\mathbb {1}. \end{align}
Proof

Apply (12) sequentially between neighboring letters. The inserted rank-one operators split the periodic contraction into the product of the one-letter contractions \((a|W^{\sigma _x\tau _x}|b)\). Identity (11) turns this product into \(\prod _x\delta _{\sigma _x,\tau _x}\), which is the matrix element of the physical identity.

Theorem 19.2.7 A simple tensor generates matrix product unitaries

Every simple tensor generates matrix product unitaries. This is [ CPGSV17 , Proposition III.3(i) ] .

Proof

For every \(N{\gt}1\), the closed-double-layer identity and the sequential contraction give

\begin{align} \bigl(U^{(N)}\bigr)^\dagger U^{(N)} & =\mathbb {1}. \end{align}

Hence \(U^{(N)}\) is unitary.

19.3 Double-layer blocking and mixed physical traces

Theorem 19.3.1 Physical adjoints and double layers commute with blocking

For every blocking length \(L\), physical adjunction commutes with blocking, and consequently

\begin{align} \mathcal W(\mathcal U^{[L]}) & =\mathcal W(\mathcal U)^{[L]}. \end{align}

The virtual order is unchanged on both sides.

Proof

Physical adjunction conjugates each matrix entry without reversing the virtual word. The first identity follows entrywise. The second follows because physical blocking commutes with the local operator product.

Assume \(d{\gt}0\), and let \(E\) be the normalized transfer matrix of \(\mathcal U\). After blocking \(L\) sites, the normalized physical diagonal of the double layer is exactly \(E^L\). More precisely, if

\begin{align} q:\{ 0,\ldots ,D{-}1\} \times \{ 0,\ldots ,D{-}1\} & \longrightarrow \{ 0,\ldots ,D^2{-}1\} \end{align}

is the standard product encoding, then its entries are

\begin{align} \bigl(E^L\bigr)_{q^{-1}(a),q^{-1}(b)}. \end{align}

Thus the row and column reindexings have the same explicit orientation. If \(\mathcal U\) is an MPU and \(D{\gt}1\), then at the stabilization exponent \(J\) this diagonal is the same reindexing of the rank-one projector \(|\rho )(\Phi |\).

Proof

On one site, expanding the diagonal double-layer contraction gives the transfer matrix after the stated product reindexing. The factor \(d^{-1}\) is the square of the local factor \(d^{-1/2}\). Summing blocked diagonal letters over words and distributing the ordered product shows that normalized diagonals turn blocking into powers. Simultaneous reindexing preserves multiplication, giving the displayed formula. Transfer stabilization gives the final rank-one specialization.

Theorem 19.3.3 Normalized diagonal as a diagonal-word sum

Let \(\mathcal W\) be an MPO tensor of physical dimension \(d\), and let

\begin{align} R & =\frac1d\sum _{i=0}^{d-1}W^{ii} \end{align}

be its normalized physical diagonal. Here and in the two diagonal-tail identities below, scalar inverses are the totalized complex-field inverses, so \(0^{-1}=0\). For every \(K\geq 0\),

\begin{align} R^K & =d^{-K}\sum _{\tau \in \{ 0,\ldots ,d-1\} ^K} W^{\tau _0\tau _0}\cdots W^{\tau _{K-1}\tau _{K-1}}. \end{align}

At \(K=0\), the word product is the identity matrix. This is the algebraic diagonal-tail expansion used in the source-factor networks of [ CPGSV17 , Theorem III.8, Section III.B, equations (31)–(32) ] .

Proof

Expand the \(K\)th power of the finite sum defining \(R\). Distributivity indexes the resulting terms by physical words \(\tau \), and each term is the ordered diagonal word displayed above.

Theorem 19.3.4 A matrix basis adapted to the trace
#

If \(d{\gt}0\), the identity matrix extends to a basis of \(M_{d}(\mathbb {C})\) whose other elements are traceless.

Proof

The trace functional is nonzero on the identity. Extend the identity by a basis of the kernel of the trace.

Theorem 19.3.5 Identity plus traceless physical decomposition

Assume \(d{\gt}0\). For every double-layer tensor \(\mathcal W\), there is a basis \(\{ \sigma _0,\sigma _\alpha \} \) of the physical matrix algebra and virtual matrices \(E,S_\alpha \) such that \(\sigma _0=\mathbb {1}\) and, for every nonzero index \(\alpha \), \(\operatorname{tr}(\sigma _\alpha )=0\), while

\begin{align} \mathcal W & =E\otimes \mathbb {1}+ \sum _{\alpha \ne 0}S_\alpha \otimes \sigma _\alpha . \end{align}

Here \(E\) is the normalized physical diagonal. Equivalently, contraction against a physical matrix \(X\) separates as

\begin{align} \operatorname {contr}_{\rm phys}(\mathcal W,X) & =\operatorname{tr}(X)E+R_{\mathcal W}(X), \\ R_{\mathcal W}(X) & =\operatorname {contr}_{\rm phys} \left(\mathcal W, X-\frac{\operatorname{tr}(X)}d\mathbb {1}\right), \end{align}

where the second physical argument is traceless. This is equation WIsom in [ CPGSV17 , Proposition III.3(ii) ] .

Proof

The trace functional is nonzero on the identity. Extend the identity to a basis adapted to the kernel of the trace; all remaining basis vectors are therefore traceless. The bilinear pairing \((X,Y)\mapsto \operatorname{tr}(XY)\) is nondegenerate. Its dual basis gives the virtual coefficients by physical contraction. The dual vector paired with the identity is \(d^{-1}\mathbb {1}\), so its virtual coefficient is precisely \(E\). Expanding matrix units in the dual basis yields the decomposition. Subtracting the trace component gives the equivalent residual formula.

Suppose a length-\(N\) closed tensor equals the physical identity. For arbitrary physical matrices \(X_0,\ldots ,X_{N-1}\),

\begin{align} \operatorname{tr}\! \left(\prod _{k=0}^{N-1} \operatorname {contr}_{\rm phys}(\mathcal W,X_k)\right) & =\prod _{k=0}^{N-1}\operatorname{tr}(X_k). \end{align}

In particular this applies to the double layer of an MPU whenever \(N{\gt}1\).

Proof

Expand each contraction and distribute the ordered product. The closed tensor coefficient is a Kronecker delta between the two physical words, so only equal words survive. The remaining independent diagonal sums factor into the product of the physical traces.

Let \(\mathcal U\) be an MPU with physical dimension \(d{\gt}0\). Every closed product of positive length that contains at least one residual coefficient has zero trace. In particular, this includes a single residual coefficient. This is the mixed-trace assertion preceding equation ESE=0 in [ CPGSV17 , Proposition III.3(ii) ] .

Proof

At lengths greater than one, represent identity factors by \(d^{-1}\mathbb {1}\) and residual factors by traceless physical arguments. The multilinear contraction identity then vanishes. For a single residual coefficient \(A\), repeat the same coefficient at every length \(q{\gt}1\) to obtain \(\operatorname{tr}(A^q)=0\). Thus the one-letter case is recovered after the longer words: the shifted zero-trace-power theorem makes \(A\) nilpotent, and therefore \(\operatorname{tr}(A)=0\).

Let \(\mathcal U\) be an MPU with physical dimension \(d{\gt}0\), and let \(N{\gt}1\). Every closed product of \(N\) identity and residual coefficients that contains at least one residual coefficient has zero trace. If the normalized diagonal is rank one, then every nonempty residual word \(P\) satisfies

\begin{align} EPE & =0. \end{align}

The same conclusion holds after blocking \(J{\gt}0\) sites whenever the \(J\)th normalized transfer power is \(|\rho )(\Phi |\), with both virtual indices reindexed by \(q^{-1}:\{ 0,\ldots ,D^2{-}1\} \longrightarrow \{ 0,\ldots ,D{-}1\} \times \{ 0,\ldots ,D{-}1\} \).

Proof

Represent identity factors by \(d^{-1}\mathbb {1}\) and residual factors by their traceless physical arguments. Since \(N{\gt}1\), the multilinear contraction identity applies, and one trace factor vanishes. For a nonempty residual word, adjoin \(E\) as one extra factor; its closed trace has length at least two and therefore vanishes. Cyclicity of trace and

\begin{align} |\rho )(\Phi |P|\rho )(\Phi | & =\operatorname{tr}\! \left(P|\rho )(\Phi |\right)|\rho )(\Phi | \end{align}

give \(EPE=0\). The blocked statement follows from the transfer-power formula with the same product-index reindexing in both matrix indices.

Definition 19.3.9 Residual nonunital algebra

Let \(\mathcal U\) have bond dimension \(D\). Contract the double layer against arbitrary physical matrices, subtract the normalized diagonal coefficient, and let \(\mathcal S\) be the resulting set of residual matrices in \(M_{D^2}(\mathbb C)\). The residual algebra \(\mathfrak S\) is the not necessarily unital algebra generated by \(\mathcal S\). This is the algebra generated by the matrices \(S'_\alpha \) in [ CPGSV17 , Proposition III.3(ii) ] .

If \(\mathcal U\) is an MPU with positive physical dimension, then every nonempty product of residual generators has zero trace. Consequently, for every \(A\in \mathfrak S\) and every integer \(N{\gt}1\),

\begin{align} \operatorname{tr}(A) & =0, & \operatorname{tr}(A^N) & =0, \end{align}

and every element of \(\mathfrak S\) is nilpotent.

Proof

A nonempty product in the multiplicative semigroup generated by \(\mathcal S\) is a nonempty residual word, so its trace vanishes by the positive-length mixed residual identity. The generated nonunital algebra is the linear span of this semigroup. Linearity of the trace therefore gives \(\operatorname{tr}(A)=0\) for every \(A\in \mathfrak S\). Since every positive power of \(A\) remains in \(\mathfrak S\), the same argument gives \(\operatorname{tr}(A^N)=0\) for \(N{\gt}1\). The shifted zero-trace-power theorem now implies that \(A\) is nilpotent.

Let \(\mathcal U\) be an MPU of bond dimension \(D\) and positive physical dimension. Every ordered product of exactly \(D^2\) elements of \(\mathfrak S\) vanishes. In particular, for arbitrary residual generators,

\begin{align} S_{\alpha _1}\cdots S_{\alpha _{D^2}} & =0. \end{align}

This is equation Sprime=0 in [ CPGSV17 , Proposition III.3(ii) ] , with the corrected exact length \(D^2\) in place of the unsupported strict intermediate bound.

Proof

The algebra \(\mathfrak S\) acts on a space of dimension \(D^2\), and every element is nilpotent. The finite-dimensional nil-matrix theorem therefore annihilates every ordered product of \(D^2\) elements. Each residual generator belongs to \(\mathfrak S\), giving the stated specialization.

Let \(E\) be the normalized diagonal after the first blocking and let \(\mathfrak S\) be its residual nonunital algebra. Define

\begin{align} \mathcal T={} & \operatorname {span}\{ EA:A\in \mathfrak S\} +\operatorname {span}\{ AE:A\in \mathfrak S\} \\ & +\operatorname {span}\{ AEB:A,B\in \mathfrak S\} . \end{align}

This is the basis-independent sum of the three forms in equation sprimeforms of [ CPGSV17 , Proposition III.3(ii) ] .

Theorem 19.3.13 Ordered second-block expansion

Suppose \(E^2=E\), \(EAE=0\) for every \(A\in \mathfrak S\), and every ordered product of \(N{\gt}0\) elements of \(\mathfrak S\) vanishes. For scalars \(c_k\) and elements \(S_k\in \mathfrak S\),

\begin{align} \prod _{k=0}^{N-1}(c_kE+S_k) -\left(\prod _{k=0}^{N-1}c_k\right)E & \in \mathcal T. \end{align}

Products are taken in increasing site order and need not commute.

Proof

Expand the product as a sum over binary choices of \(E\) or \(S_k\), retaining the original order. The all-\(E\) choice is the subtracted term because \(E^q=E\) for \(q{\gt}0\). The all-residual choice vanishes by the length-\(N\) hypothesis. In every remaining word, contiguous copies of \(E\) collapse. If two resulting copies of \(E\) are separated by a nonempty residual word, the factor \(EAE\) makes the word zero. Thus each surviving word has one contiguous projector block and belongs to one of the three summands of \(\mathcal T\).

Let the first-block tensor be an MPU of bond dimension \(D{\gt}0\). Assume its normalized double-layer diagonal \(E\) is a rank-one idempotent. After a second blocking of exactly \(D^2\) sites, every residual coefficient belongs to \(\mathcal T\).

The conclusion is membership in the sum of the three subspaces. It does not assign one exclusive type to a coefficient in an arbitrary physical basis.

Proof

First extend \(EPE=0\) from nonempty residual words \(P\) to every element of \(\mathfrak S\) by writing the generated algebra as the linear span of the multiplicative semigroup of residual generators. For a blocked matrix unit, each local double-layer entry is

\begin{align} \delta _{ij}E+R_{ij}. \end{align}

The ordered expansion theorem applies with \(N=D^2\). Its all-\(E\) coefficient is the blocked Kronecker delta and is exactly removed by the residual slice. Hence every blocked matrix-unit residual lies in \(\mathcal T\). Expanding an arbitrary physical matrix in matrix units and using linearity proves the result.

Suppose \(E^2=E\) and \(EAE=0\) for every \(A\in \mathfrak S\). For all \(X,Y\in \mathcal T\),

\begin{align} XY & =XEY, & \operatorname{tr}(XE) & =0. \end{align}
Proof

It is enough by bilinearity and linearity of trace to consider the generators \(EA\), \(AE\), and \(AEB\) of the three summands. Representative products reduce as

\begin{align} (EA)(BE) & =E(AB)E=0, \\ (AE)(EB) & =AEB=(AE)E(EB), \\ \operatorname{tr}((AEB)E) & =\operatorname{tr}(A(EBE))=0. \end{align}

The remaining generator pairs follow by the same use of associativity, \(E^2=E\), and \(EAE=0\) for \(A\in \mathfrak S\).

Let \(\mathcal V\) be an MPU of positive physical and bond dimensions whose normalized double-layer diagonal is

\begin{align} E=|\rho )(\Phi |, & & (\Phi |\rho )=1. \end{align}

Then blocking \(\mathcal V\) by exactly \(D^2\) sites produces a simple tensor with witnesses \(a=\Phi \) and \(b=\rho \).

Proof

The pairing normalization makes \(E\) idempotent. Write every blocked double-layer coefficient as \(\delta _{ij}E+R_{ij}\), where \(R_{ij}\in \mathcal T\). The trace identity for \(R_{ij}E\) gives the diagonal contraction, and \(R_{ij}R_{k\ell }=R_{ij}ER_{k\ell }\), together with \(E^2=E\), gives the rank-one insertion identity after expanding both coefficients.

Definition 19.3.17 Physical-basis reindexing

For a bijection \(e\) between two physical index sets, the reindexed tensor and its action on length-\(N\) configurations are

\begin{align} (e^*\mathcal U)^{ij} & =\mathcal U^{e(i)e(j)}, & (e_*\sigma )(n) & =e(\sigma (n)). \end{align}

For ket and bra words \(i,j\), physical reindexing gives

\begin{align} (e^*\mathcal U)[i,j] & =\mathcal U[e_*i,e_*j]. \end{align}
Proof

Induct simultaneously along the two words. The empty pair gives the identity matrix, mismatched lengths give zero, and the nonempty case applies the defining equality to the first letters and the induction hypothesis to the tails.

At every length \(N\), the sitewise configuration bijection identifies

\begin{align} (e^*\mathcal U)^{(N)} & \cong \mathcal U^{(N)}. \end{align}
Proof

Apply the word identity to every pair of configurations and take the virtual trace. This is simultaneous row and column reindexing by \(e_*\).

For every physical-index bijection \(e\), taking the double layer commutes with physical reindexing:

\begin{align} W(e^*\mathcal U)^{ij} & =W(\mathcal U)^{e(i)e(j)}. \end{align}
Proof

This follows entrywise from the definitions of physical reindexing and the double layer.

Let \(e\) be the canonical bijection which groups a word of length \(mn\) into \(n\) consecutive words of length \(m\). Iterating blocking at lengths \(m\) and \(n\), then applying this physical reindexing, agrees with direct blocking at length \(mn\):

\begin{align} e^*((\mathcal U_m)_n) & =\mathcal U_{mn}. \end{align}
Proof

Expand the two blocked words. The canonical regrouping concatenates their \(n\) words of length \(m\) into the direct word of length \(mn\).

Every MPU of positive physical and bond dimensions has a positive blocking length \(k\leq D^4\) for which the blocked tensor is simple. If \(D{\gt}1\), one may choose

\begin{align} k=JD^2{\lt}D^4, & & 0{\lt}J\leq D^2-1. \end{align}

If \(D=1\), the length-one block is simple. Simplicity is preserved under every further iterated blocking by a positive factor and under a bijective relabeling of the physical basis.

Proof

At the stabilization exponent \(J\), the normalized double-layer diagonal is the rank-one fixed projector \(|\rho )(\Phi |\). The preceding theorem makes the subsequent length-\(D^2\) block simple. If \(e\) is the canonical grouping bijection, the two blocking stages satisfy

\begin{align} e^*((\mathcal U_J)_{D^2})=\mathcal U_{JD^2}. \end{align}

For \(D{\gt}1\),

\begin{align} JD^2 & {\lt}(D^2)D^2=D^4. \end{align}

The one-dimensional transfer matrix is already the identity. For a further block with letters \(W_1,\ldots ,W_L\) and \(R=|\rho )(\Phi |\), repeated insertion gives

\begin{align} W_1\cdots W_L & =W_1R W_2R\cdots R W_L, \\ (\Phi |W_1\cdots W_L|\rho ) & =\prod _{q=1}^{L}(\Phi |W_q|\rho ). \end{align}

The second identity multiplies the sitewise Kronecker deltas.

Theorem 19.3.23 Matching contractions at the direct stabilized block

Let \(\mathcal U\) be an MPU of positive physical and bond dimensions. Suppose that \(\rho \) has trace one and that, for some positive integer \(J\), the normalized transfer matrix satisfies

\begin{align} E^J & =|\rho )(\mathbb {1}|. \end{align}

Put \(K=JD^2\). After the standard product reindexing of the physical coordinates, the direct length-\(K\) block satisfies the contractions (11) and (12) with the specific witnesses \(\rho \) and \(\mathbb {1}\) from this transfer-power identity.

Proof

First block \(J\) sites. The normalized diagonal of this block is the reindexed rank-one matrix \(|\rho )(\mathbb {1}|\), and the trace-one condition is exactly \((\mathbb {1}|\rho )=1\). The exact second-block contraction theorem then supplies both identities after a further \(D^2\) sites. The canonical equivalence between iterated blocking and direct length-\(JD^2\) blocking transports each double-layer coefficient without changing either virtual witness.

Theorem 19.3.24 Normalized output-tail coisometry

Let \(d{\gt}0\), let \(\mathcal U\) be an MPU, and let \(K\geq 0\). For \(p,q\in \{ 0,\ldots ,d{-}1\} \times \{ 0,\ldots ,d{-}1\} \), separate the first two output sites from a tail \(\tau \in \{ 0,\ldots ,K{-}1\} \to \{ 0,\ldots ,d{-}1\} \). Then

\begin{align} d^{-K}\sum _{\tau }\sum _{\eta } U^{(K+2)}_{(q,\tau ),\eta } \overline{U^{(K+2)}_{(p,\tau ),\eta }} & =\delta _{p,q}. \end{align}

The retained matrix entry has the reversed order \((q,p)\), and the global contraction is the output-first equation \(U^{(K+2)}(U^{(K+2)})^\dagger =\mathbb {1}\).

Proof

Apply output-first unitarity before performing any source-factor rearrangement. For each fixed tail, the inner sum is the corresponding identity-matrix entry, hence it is \(\delta _{p,q}\). There are exactly

\begin{align} |\{ 0,\ldots ,K{-}1\} \to \{ 0,\ldots ,d{-}1\} |=d^K \end{align}

tails, and \(d^{-K}d^K=1\) because \(d{\gt}0\).

Theorem 19.3.25 Normalized input-tail isometry

Let \(d{\gt}0\), let \(\mathcal U\) be an MPU, and let \(K\geq 0\). For \(p,q\in \{ 0,\ldots ,d{-}1\} \times \{ 0,\ldots ,d{-}1\} \), separate the first two input sites from a common input tail \(\tau \in \{ 0,\ldots ,K{-}1\} \to \{ 0,\ldots ,d{-}1\} \). Then

\begin{align} d^{-K}\sum _{\tau }\sum _{\eta } \overline{U^{(K+2)}_{\eta ,(p,\tau )}} U^{(K+2)}_{\eta ,(q,\tau )} & =\delta _{p,q}. \end{align}

The global contraction is the input-first equation \((U^{(K+2)})^\dagger U^{(K+2)}=\mathbb {1}\). This is the input-oriented tail contraction supporting the network calculation in [ CPGSV17 , Theorem III.8, Section III.B, equations (31)–(32) ] .

Proof

Apply input-first unitarity for each fixed tail. The inner output sum is \(\delta _{p,q}\). Summing over the \(d^K\) tails and multiplying by \(d^{-K}\) leaves the same Kronecker delta.

Let \(\mathcal U\) generate an MPU, and suppose its direct length-\(k\) block is simple for some \(k{\gt}0\). Then its direct length-\(k'\) block is simple for every \(k'\geq k\), without any divisibility condition on \(k'\) by \(k\). This is the corollary following [ CPGSV17 , Proposition III.3 ] .

Proof

First suppose vectors \(a,b\) satisfy the rank-one insertion identity

\begin{align} W^{ij}W^{k\ell } & =W^{ij}|b)(a|W^{k\ell }. \end{align}

Set \(x_{ij}=(a|W^{ij}|b)\). Applying this identity around the constant two-site and three-site words, and then using MPU unitarity, gives

\begin{align} x_{ij}^2 & =\delta _{ij}, \\ x_{ij}^3 & =\delta _{ij}. \end{align}

Hence \(x_{ii}=1\), while \(x_{ij}=0\) for \(i\ne j\), so the same vectors obey the one-letter contraction \((a|W^{ij}|b)=\delta _{ij}\).

For the successor step, take four words \(I,J,K,L\) of length \(k+1\) and write \(W^{I,J}=W^{i_0j_0}\cdots W^{i_kj_k}\), with the analogous notation for \(W^{K,L}\). Apply the length-\(k\) rank-one insertion identity to the suffixes \(I_+=(i_1,\ldots ,i_k)\), \(J_+=(j_1,\ldots ,j_k)\) and the prefixes \(K_-=(k_0,\ldots ,k_{k-1})\), \(L_-=(\ell _0,\ldots ,\ell _{k-1})\). Then

\begin{align} W^{I,J}W^{K,L} & =W^{i_0j_0}\bigl(W^{I_+,J_+}W^{K_-,L_-}\bigr)W^{k_k\ell _k} \\ & =W^{I,J}|b)(a|W^{K,L}. \end{align}

Thus the length-\((k+1)\) block has the same insertion witnesses, and the two- and three-site argument gives their one-letter contraction. Induction on \(k'-k\) proves the claim.

Let \(d,D{\gt}0\). Every bond-\(D\) MPU tensor becomes simple after blocking exactly \(D^4\) sites. For each fixed \(L\), the map

\begin{align} \mathcal U & \longmapsto \mathcal U_L \end{align}

is continuous. Consequently, if \(x\mapsto \mathcal W(x)\) is a continuous family of bond-\(D\) MPU tensors, then \(x\mapsto \mathcal W(x)_{D^4}\) is a continuous family of simple tensors. This is only the common-blocking step in the proof of Proposition 19.9.59; it does not choose reduced representatives or assert constancy of their source ranks.

Proof

Choose for each tensor a positive simple blocking length \(k\leq D^4\). Simplicity persists at every later direct blocking, hence at \(D^4\). For fixed blocked ket and bra words, every matrix entry of the blocked tensor is a finite sum of products of entries of the original tensor. These entries are continuous, and the finite product and sum operations preserve continuity. Applying this fixed blocking map to the original path proves the path statement.

Theorem 19.3.28 Supplied-projector simple2 contraction

Let \(\mathcal U\) be simple with positive physical dimension, let \(W\) be its double layer, and let \(E\) be the normalized diagonal of \(W\). Suppose that for some \(J{\gt}0\),

\begin{align} E^J & =|\rho )(\Phi |. \end{align}

Then the simple insertion identity holds with these same supplied witnesses:

\begin{align} W^{ij}W^{k\ell } & =W^{ij}|\rho )(\Phi |W^{k\ell }. \end{align}
Proof

Choose the witnesses \(a,b\) from simplicity and put \(R=|b)(a|\). Expanding the normalized diagonal as the average of the diagonal letters, the two defining simple contractions give

\begin{align} W^{ij} E W^{k\ell } & =W^{ij}W^{k\ell }. \end{align}

Indeed, insert \(R\) on both sides of each diagonal letter; its middle sandwich is \(R W^{qq}R=R\) by the first contraction, and the normalized sum contains \(d\) equal terms. The same identity makes \(W^{ij}E^N=W^{ij}E\) for every \(N{\gt}0\). Hence

\begin{align} W^{ij}E^J W^{k\ell } & =W^{ij}W^{k\ell }. \end{align}

Substituting the supplied equality \(E^J=|\rho )(\Phi |\) proves the claim. No uniqueness of rank-one factorizations is used.

Theorem 19.3.29 Fixed boundaries of a normalized rank-one power

Let \(R\) be the normalized physical diagonal of the double layer of an MPO tensor. Suppose

\begin{align} (\Phi \rvert \rho ) & =1, & R^J & =\lvert \rho )(\Phi \rvert . \end{align}

Then the two boundaries of the stabilized diagonal tail are fixed:

\begin{align} R^J\lvert \rho ) & =\lvert \rho ), & (\Phi \rvert R^J & =(\Phi \rvert . \end{align}

This is the fixed-pair normalization from the canonical-form-II fixed-point equations in [ CPGSV17 , equations (6a)–(6b) ] and underlies the tail contractions in [ CPGSV17 , Theorem III.8, Section III.B, equations (31)–(32) ] .

Proof

Substitute \(R^J=\lvert \rho )(\Phi \rvert \) in each expression and use \((\Phi \rvert \rho )=1\).

Theorem 19.3.30 Rank-one insertion as a normalized diagonal tail

Let \(W\) be the double layer of an MPO tensor, and let \(R\) be its normalized physical diagonal. If

\begin{align} R^J & =\lvert \rho )(\Phi \rvert , \end{align}

then, for all physical indices \(i,j,k,\ell \),

\begin{align} W^{ij}\lvert \rho )(\Phi \rvert W^{k\ell } =d^{-J}\sum _{\tau \in \{ 0,\ldots ,d-1\} ^J} W^{ij}W^{\tau _0\tau _0}\cdots W^{\tau _{J-1}\tau _{J-1}}W^{k\ell }. \end{align}

This is the exact diagonal-word tail between the two retained letters in [ CPGSV17 , Theorem III.8, Section III.B, equations (31)–(32) ] .

Proof

Replace the rank-one insertion by \(R^J\), apply the diagonal-word expansion, and distribute the two retained letters across the normalized finite sum.

Theorem 19.3.31 Simple two-letter product as a normalized diagonal tail

In the setting of the preceding theorem, suppose in addition that the same fixed pair satisfies the supplied two-letter contraction

\begin{align} W^{ij}W^{k\ell } & =W^{ij}\lvert \rho )(\Phi \rvert W^{k\ell } \end{align}

for all physical indices. Then

\begin{align} W^{ij}W^{k\ell } =d^{-J}\sum _{\tau \in \{ 0,\ldots ,d-1\} ^J} W^{ij}W^{\tau _0\tau _0}\cdots W^{\tau _{J-1}\tau _{J-1}}W^{k\ell }. \end{align}

Thus the supplied insertion and its stabilized transfer power occur in one aligned network, as required in the calculation of [ CPGSV17 , Theorem III.8, Section III.B, equations (31)–(32) ] .

Proof

Apply the supplied two-letter contraction and then replace its rank-one insertion by the normalized diagonal-word tail.

Theorem 19.3.32 The fixed-point simple1 contraction

For every MPU of positive physical and bond dimensions there are normalized left and right fixed vectors \(\Phi \) and \(\rho \) of the transfer matrix such that the unblocked double layer satisfies

\begin{align} (\Phi |W^{ij}|\rho ) & =\delta _{ij}. \end{align}

The unblocked tensor need not satisfy the rank-one insertion identity.

Proof

Let \(E^J=|\rho )(\Phi |\) be a stabilized transfer power. Decompose \(W^{ij}=\delta _{ij}E+R_{ij}\). Fixedness and normalization give

\begin{align} \operatorname{tr}(E|\rho )(\Phi |) & =1. \end{align}

The mixed residual trace identity applied to one residual factor followed by \(J\) diagonal factors gives

\begin{align} \operatorname{tr}(R_{ij}E^J) & =0. \end{align}

Since \(E^J=|\rho )(\Phi |\), these two trace identities are exactly the displayed contraction.

19.4 Trace-power lemma

For the transfer-matrix spectrum argument of [ CPGSV17 , Proposition 1 ] we first record the algebraic fact that a matrix whose positive powers all have trace \(1\) must have characteristic polynomial \(X^{n-1}(X-1)\). The source assumes only the powers with exponent strictly greater than one; this weaker hypothesis determines the nonzero spectrum as a set.

Definition 19.4.1 Rank-one diagonal idempotent
#

For an index \(i_0\) in a finite set \(I\) of cardinality \(n\), let \(M_{i_0}\) be the \(n\times n\) diagonal matrix with a \(1\) at position \(i_0\) and zeros elsewhere.

Lemma 19.4.2 Properties of the rank-one diagonal idempotent

The matrix \(M_{i_0}\) from definition 19.4.1 satisfies

  • \(M_{i_0}^k = M_{i_0}\) for every \(k\ge 1\),

  • \(\operatorname{tr}(M_{i_0}^k) = 1\) for every \(k\ge 1\),

  • \(\chi _{M_{i_0}}(X) = X^{\, n-1}(X-1)\).

Proof

The first two items follow from \(1^k=1\) and \(0^k=0\) for \(k\ge 1\). The characteristic polynomial is computed from \(\chi _{\operatorname{diag}(d)}(X) = \prod _i (X - d_i)\).

Theorem 19.4.3 Finite-range trace-power lemma

Let \(A\) be an \(n\times n\) complex matrix with \(n\ge 1\). If \(\operatorname{tr}(A^k)=1\) for every \(k\) with \(1\le k\le n\), then

\begin{align} \chi _A(X) & = X^{\, n-1}(X-1). \end{align}
Proof

Choose any index \(i_0\) (the hypothesis \(n\ge 1\) makes this possible) and let \(B\) be the rank-one diagonal idempotent from Lemma 19.4.2. For \(1\le k\le n\) we have \(\operatorname{tr}(A^k)=1=\operatorname{tr}(B^k)\). The finite-range half of the Newton–Girard the corresponding theorem in the companion quantum-channel volume  [ LTC26 , “Newton–Girard trace recursion” ] then yields \(\chi _A = \chi _B = X^{\, n-1}(X-1)\).

Theorem 19.4.4 All-positive-powers trace-power lemma

Let \(A\) be an \(n\times n\) complex matrix. If \(\operatorname{tr}(A^k)=1\) for every \(k\ge 1\), then

\begin{align} \chi _A(X) & = X^{\, n-1}(X-1). \end{align}

(The hypothesis already forces \(n\ge 1\) because the trace of a \(0\times 0\) matrix is \(0\), contradicting \(\operatorname{tr}(A)=1\).)

Proof

If \(n=0\) the hypothesis \(\operatorname{tr}(A)=1\) contradicts the fact that the trace of an empty matrix is \(0\); the conclusion follows vacuously. When \(n\ge 1\), the hypothesis on all positive powers implies the finite-range hypothesis of Theorem 19.4.3.

Theorem 19.4.5 Shifted moments under matrix powers

Let \(A\) be an \(n\times n\) complex matrix such that \(\operatorname{tr}(A^k)=1\) for every \(k{\gt}1\), and let \(m{\gt}1\). Then \(\operatorname{tr}((A^m)^r)=1\) for every \(r\ge 1\).

Proof

For \(r\ge 1\), one has \((A^m)^r=A^{mr}\) and \(mr{\gt}1\), so the hypothesis at exponent \(mr\) gives the result.

Let \(A\) be an \(n\times n\) complex matrix. If \(\operatorname{tr}(A^k)=1\) for every integer \(k{\gt}1\), then

\begin{align} \sigma (A)\setminus \{ 0\} & =\{ 1\} . \end{align}

Thus \(1\) occurs and every nonzero spectral value equals \(1\). This is the set-spectrum consequence of [ CPGSV17 , Proposition III.1 ] ; set equality alone does not assert algebraic multiplicity.

Proof

For every positive integer \(r\), the shifted moments give

\begin{align} \operatorname{tr}((A^2)^r) & =\operatorname{tr}(A^{2r})=1, & \operatorname{tr}((A^3)^r) & =\operatorname{tr}(A^{3r})=1. \end{align}

Hence

\begin{align} \chi _{A^2}(X) & =\chi _{A^3}(X)=X^{n-1}(X-1). \end{align}

If \(\mu \in \sigma (A)\) is nonzero, spectral mapping places \(\mu ^2\) in \(\sigma (A^2)\) and \(\mu ^3\) in \(\sigma (A^3)\), so

\begin{align} \mu ^2 & =\mu ^3=1, & \mu & =1. \end{align}

Conversely, spectral mapping from \(1\in \sigma (A^2)\) supplies a nonzero spectral value of \(A\), which must be \(1\).

Theorem 19.4.7 Generalized eigenspace inclusion under squaring
#

For a complex square matrix \(A\), write \(G_1(A)\) for its maximal generalized eigenspace at the eigenvalue \(1\). Then

\begin{align} G_1(A) & \subseteq G_1(A^2). \end{align}
Proof

The commuting factorization

\begin{align} A^2-\mathbb {1}& =(A+\mathbb {1})(A-\mathbb {1}) \end{align}

gives, for every nonnegative integer \(k\),

\begin{align} (A^2-\mathbb {1})^k & =(A+\mathbb {1})^k(A-\mathbb {1})^k. \end{align}

Therefore every vector annihilated by a power of \(A-\mathbb {1}\) is annihilated by the corresponding power of \(A^2-\mathbb {1}\).

Theorem 19.4.8 Multiplicity at one does not decrease under squaring

For every complex square matrix \(A\),

\begin{align} \operatorname {mult}_{1}(\chi _A) & \le \operatorname {mult}_{1}(\chi _{A^2}). \end{align}
Proof

With \(G_1\) as above, algebraic multiplicity equals the dimension of the maximal generalized eigenspace. Hence

\begin{align} \operatorname {mult}_{1}(\chi _A) & =\dim G_1(A) \\ & \le \dim G_1(A^2) =\operatorname {mult}_{1}(\chi _{A^2}). \end{align}
Theorem 19.4.9 Trace powers above one determine the characteristic polynomial

Let \(A\) be an \(n\times n\) complex matrix. If \(\operatorname{tr}(A^k)=1\) for every integer \(k{\gt}1\), then

\begin{align} \operatorname {mult}_{1}(\chi _A) & =1, & \chi _A(X) & =X^{n-1}(X-1). \end{align}

Thus \(1\) is the sole nonzero eigenvalue and has algebraic multiplicity one. This is the exact shifted-moment conclusion of [ CPGSV17 , Proposition III.1 ] .

Proof

Applying the all-positive trace-power theorem to \(A^2\) gives

\begin{align} \chi _{A^2}(X) & =X^{n-1}(X-1), & \operatorname {mult}_{1}(\chi _{A^2}) & =1. \end{align}

The multiplicity inequality therefore bounds \(\operatorname {mult}_{1}(\chi _A)\) by one. The set-spectrum theorem shows that \(1\) occurs, so this multiplicity is positive and hence equals one. The same set-spectrum theorem excludes every other nonzero root. Splitting the monic characteristic polynomial now yields

\begin{align} \operatorname {mult}_{1}(\chi _A) & =1, & \chi _A(X) & =X^{n-1}(X-1). \end{align}
Theorem 19.4.10 Vanishing positive trace powers determine the characteristic polynomial

Let \(A\) be an \(n\times n\) complex matrix. If \(\operatorname{tr}(A^k)=0\) for every integer \(k\geq 1\), then

\begin{align} \chi _A(X) & =X^n. \end{align}
Proof

The zero matrix has the same positive trace powers as \(A\) and characteristic polynomial \(X^n\). Newton–Girard therefore identifies the two characteristic polynomials. This is the characteristic-polynomial consequence of the all-positive trace condition obtained in [ CPGSV17 , proof of Proposition III.3 ] .

Theorem 19.4.11 Shifted vanishing moments under matrix powers

Let \(A\) be an \(n\times n\) complex matrix such that \(\operatorname{tr}(A^k)=0\) for every \(k{\gt}1\), and let \(m{\gt}1\). Then \(\operatorname{tr}((A^m)^r)=0\) for every \(r\geq 1\). This auxiliary result supports the generalization below to a weaker hypothesis; the source itself supplies vanishing for every positive exponent.

Proof

For \(r\geq 1\), one has \((A^m)^r=A^{mr}\) and \(mr{\gt}1\), so the hypothesis at exponent \(mr\) gives the result.

Let \(A\) be an \(n\times n\) complex matrix. If \(\operatorname{tr}(A^N)=0\) for every integer \(N{\gt}1\), then \(A\) is nilpotent and

\begin{align} \chi _A(X) & =X^n. \end{align}

The source obtains the all-positive trace hypothesis in [ CPGSV17 , proof of Proposition III.3 ] ; its specialization is the elementwise nilpotence assertion used before the nil-matrix theorem. The result above is a stronger generalization because it assumes vanishing only for \(N{\gt}1\). It is useful independently and requires no positivity, normality, or diagonalizability.

Proof

Every positive power of \(A^2\) has vanishing trace. Hence

\begin{align} \chi _{A^2}(X) & =X^n. \end{align}

If \(\mu \) is a spectral value of \(A\), spectral mapping shows that \(\mu ^2\) is a root of this characteristic polynomial. Therefore \(\mu ^2=0\), and hence \(\mu =0\). Thus every root of the split monic polynomial \(\chi _A\) is zero, so \(\chi _A(X)=X^n\). Cayley–Hamilton now proves that \(A\) is nilpotent.

The standard-form construction uses the compact singular-value decomposition in which the intermediate dimension is the rank rather than the smaller ambient dimension.

Theorem 19.4.13 Compact rectangular singular-value decomposition
#

Let \(M\in M_{m,n}(\mathbb {C})\) be a complex matrix of rank \(r\). There are matrices \(V\in M_{r,m}(\mathbb {C})\) and \(U\in M_{r,n}(\mathbb {C})\) and strictly positive real numbers \(s_0,\ldots ,s_{r-1}\) such that, for \(D=\operatorname{diag}(s_0,\ldots ,s_{r-1})\),

\begin{align} M & =V^\dagger D U, & VV^\dagger & =\mathbb {1}_r, & UU^\dagger & =\mathbb {1}_r. \label{eq:mpu_compact_svd} \end{align}

Thus no zero singular values occur in the intermediate space. This is the compact decomposition used in the standard-form construction of [ CPGSV17 , Section III ] .

Proof

Apply the spectral theorem to \(MM^\dagger \). Its nonzero eigenvalues are positive, and their number is \(\operatorname {rank}(MM^\dagger )=\operatorname {rank}(M)=r\). For an orthonormal eigenvector \(e_j\) with eigenvalue \(\lambda _j{\gt}0\), put \(w_j=M^\dagger e_j\). Then \(\langle w_j,w_k\rangle =\lambda _j\delta _{jk}\). Hence the vectors \(\lambda _j^{-1/2}w_j\) are orthonormal. Expanding each column of \(M\) in the eigenbasis and discarding the zero-eigenvalue terms gives

\begin{align} M & =\sum _{\lambda _j{\gt}0} \sqrt{\lambda _j}\, e_j \bigl(\lambda _j^{-1/2}w_j\bigr)^\dagger , \end{align}

which has the form in (88).

We now introduce the two matrices used for the singular-value decompositions of a Matrix Product Unitary tensor in Section III of [ CPGSV17 ] . Write \(\mathcal U^i_{j,\alpha ,\beta }\) for the entries of the tensor, where \(i,j\) are physical indices and \(\alpha ,\beta \) are auxiliary indices.

The matrix \(M_1\) is obtained by combining the left auxiliary index with the lower physical index, and the upper physical index with the right auxiliary index. Thus

\begin{align} (M_1)_{(\alpha ,j),(i,\beta )} & = \mathcal U^i_{j,\alpha ,\beta }. \end{align}

Transporting the row and column indices along the standard product enumerations gives a matrix \(\widetilde M_1\) indexed by \(\{ 0,\ldots ,Dd-1\} \times \{ 0,\ldots ,dD-1\} \). Its entries are obtained by applying \(M_1\) to the inverse images of the enumerated indices, and

\begin{align} \operatorname {rank}(\widetilde M_1) & = \operatorname {rank}(M_1). \end{align}

The matrix \(M_2\) is obtained by combining the left auxiliary index with the upper physical index, and the lower physical index with the right auxiliary index. Thus

\begin{align} (M_2)_{(\alpha ,i),(j,\beta )} & = \mathcal U^i_{j,\alpha ,\beta }. \end{align}

Transporting the row and column indices along the standard product enumerations gives a matrix \(\widetilde M_2\) indexed by \(\{ 0,\ldots ,Dd-1\} \times \{ 0,\ldots ,dD-1\} \). Its entries are obtained by applying \(M_2\) to the inverse images of the enumerated indices, and

\begin{align} \operatorname {rank}(\widetilde M_2) & = \operatorname {rank}(M_2). \end{align}
Definition 19.4.16 Right rank

The right rank is

\begin{align} r & = \operatorname {rank}(M_1). \end{align}
Definition 19.4.17 Left rank

The left rank is

\begin{align} \ell & = \operatorname {rank}(M_2). \end{align}

Physical adjunction exchanges the two source cuts and conjugates every entry:

\begin{align} M_1(\mathcal U^\sharp ) & =\overline{M_2(\mathcal U)}, \\ M_2(\mathcal U^\sharp ) & =\overline{M_1(\mathcal U)}. \end{align}
Proof

The physical adjoint exchanges the upper and lower physical indices while leaving both auxiliary indices fixed. Thus the first regrouping becomes the entrywise conjugate of the second, and conversely.

Definition 19.4.19 Physical-pair swap

On the flattened physical alphabet, define the involution

\begin{align} s(i,j) & =(j,i). \end{align}

This acts on physical coordinates and is distinct from the doubled-bond exchange used in reflected transfer calculations.

The normalized flattening of the physical adjoint satisfies

\begin{align} \widetilde{\mathcal U^\sharp }^{\, (i,j)} & =\overline{\widetilde{\mathcal U}^{\, (j,i)}}. \end{align}
Proof

For every pair of physical indices \((i,j)\),

\begin{align} \widetilde{\mathcal U^\sharp }^{\, (i,j)} & =\frac{1}{\sqrt d}\, (\mathcal U^\sharp )^{(i,j)} =\frac{1}{\sqrt d}\, \overline{\mathcal U^{(j,i)}} =\overline{\widetilde{\mathcal U}^{\, (j,i)}}. \end{align}

The first equality is the normalized flattening, the second is physical adjunction, and the third uses the reality of \(d^{-1/2}\).

Definition 19.4.21 Canonical-form-II data under physical adjunction

Given canonical-form-II data for \(\widetilde{\mathcal U}\), define the transformed tuple by

\begin{align} r’ & =r, \\ D’_k & =D_k, \\ \mu ’_k & =\overline{\mu _k}, \\ (A’_k)^{(i,j)} & =\overline{A_k^{(j,i)}}, \\ V’ & =\overline V. \end{align}

These data reconstruct \(\widetilde{\mathcal U^\sharp }\).

If \(\widetilde{\mathcal U}\) is in canonical form II, then \(\widetilde{\mathcal U^\sharp }\) is in canonical form II.

Proof

Theorem 8.6.4 independently preserves normality and left-canonical normalization under entrywise conjugation, and Theorem 8.6.8 preserves these properties under the physical-pair swap. Conjugating the coisometry equation gives

\begin{align} V’V’^\dagger & =\overline{VV^\dagger }=\mathbb {1}. \end{align}

Conjugating the reconstruction at the swapped physical index gives

\begin{align} \widetilde{\mathcal U^\sharp }^{\, (i,j)} & =(V’)^\dagger \left(\bigoplus _k\mu ’_k(A’_k)^{(i,j)}\right)V’. \end{align}

Finally, each diagonal positive matrix \(\Lambda _k\) is real, and therefore

\begin{align} \mathcal{E}_{\overline{A_k}}(\Lambda _k) & =\overline{\mathcal{E}_{A_k}(\Lambda _k)} =\overline{\Lambda _k} =\Lambda _k. \end{align}

Physical relabeling leaves this transfer map unchanged, which proves the transformed fixed-point equation.

If canonical-form-II data for \(\widetilde{\mathcal U}\) have full support, then the data from Definition 19.4.21 also have full support.

Proof

Complex conjugation of the weights leaves the block indices and their bond dimensions unchanged, so the full-support sum is unchanged.

Physical adjunction exchanges the two source ranks:

\begin{align} r(\mathcal U^\sharp ) & =\ell (\mathcal U), \\ \ell (\mathcal U^\sharp ) & =r(\mathcal U). \end{align}
Proof

Apply Theorem 19.4.18 and use invariance of matrix rank under entrywise conjugation.

A bijective relabeling of the physical basis simultaneously relabels the physical component of every source-cut row and column. Hence

\begin{align} M_1(e^*\mathcal U) & \cong M_1(\mathcal U), & M_2(e^*\mathcal U) & \cong M_2(\mathcal U), \end{align}

with the right and left orientations unchanged.

Proof

Evaluate each cut after relabeling. In \(M_1\), the lower physical index lies in the row and the upper physical index lies in the column; in \(M_2\) these roles are exchanged. Applying \(e\) to both entries gives the two displayed row-and-column equivalences.

Bijective physical relabeling preserves both source ranks:

\begin{align} r(e^*\mathcal U) & =r(\mathcal U), & \ell (e^*\mathcal U) & =\ell (\mathcal U). \end{align}
Proof

Matrix rank is unchanged by bijective row and column relabelings. Apply this fact to the two source-cut identities.

Definition 19.4.27 Independent tensor product of matrix product tensors

Let \(A=\{ A^i\} _{i\in \{ 0,\ldots ,d{-}1\} }\) and \(B=\{ B^k\} _{k\in \{ 0,\ldots ,e{-}1\} }\) have auxiliary dimensions \(D\) and \(E\), respectively. Write

\begin{align} \pi _{m,n}:\{ 0,\ldots ,m{-}1\} \times \{ 0,\ldots ,n{-}1\} & \longrightarrow \{ 0,\ldots ,mn{-}1\} \end{align}

for the standard finite-product coordinate bijection. Their independent tensor product \(A\boxtimes B\) has physical dimension \(de\), auxiliary dimension \(DE\), and local matrices determined by

\begin{align} (A\boxtimes B)^{\pi _{d,e}(i,k)}_{ \pi _{D,E}(\alpha ,\gamma ),\pi _{D,E}(\beta ,\delta )} & =A^i_{\alpha ,\beta }B^k_{\gamma ,\delta }. \label{eq:mps_independent_tensor_product_entries} \end{align}

For every \(N\), the same coordinates give the pointwise splitting

\begin{align} (\{ 0,\ldots ,N{-}1\} \to \{ 0,\ldots ,de{-}1\} ) & \simeq (\{ 0,\ldots ,N{-}1\} \to \{ 0,\ldots ,d{-}1\} )\times (\{ 0,\ldots ,N{-}1\} \to \{ 0,\ldots ,e{-}1\} ), \\ \sigma & \longmapsto (\sigma _d,\sigma _e), & \sigma (n) & =\pi _{d,e}(\sigma _d(n),\sigma _e(n)). \label{eq:mps_tensor_product_tuple_split} \end{align}

This is infrastructure for the tensoring clause in the proof of the MPU Index Theorem IV.6(ii) [ CPGSV17 , lines 824–845 ] , not a separately stated theorem of that paper. The notation \(A=\{ A^i\} _i\) agrees with the canonical-form block notation in [ CPGSV16 , Section 2.3, lines 214–245 ] .

Theorem 19.4.28 Word evaluation of an independent tensor product

If \(w\) has length \(n\) and

\begin{align} w & =\bigl(\pi _{d,e}(i_1,k_1),\ldots , \pi _{d,e}(i_n,k_n)\bigr), \\ w_d & =(i_1,\ldots ,i_n), & w_e & =(k_1,\ldots ,k_n), \end{align}

then

\begin{align} (A\boxtimes B)^w & =\operatorname {reind}_{\pi _{D,E}} \left(A^{w_d}\otimes B^{w_e}\right). \label{eq:mps_eval_word_independent_tensor_product} \end{align}
Proof

Induct on the length of \(w\). For the empty word, use \(\mathbb {1}_D\otimes \mathbb {1}_E=\mathbb {1}_{DE}\) and invariance of the identity under the coordinate bijection. For a leading letter \(\pi _{d,e}(i,k)\), equation (114) and the induction hypothesis reduce the claim to

\begin{align} (A^i\otimes B^k)(X\otimes Y) & =(A^iX)\otimes (B^kY), \end{align}

after transporting multiplication through the same coordinate bijection.

Let \(S_N(A)=\operatorname{span}_{\mathbb {C}}\{ A^w:|w|=N\} \). The pointwise splitting of length-\(N\) words gives

\begin{align} S_N(A\boxtimes B) & =\operatorname{span}_{\mathbb {C}}\{ \operatorname {reind}_{\pi _{D,E}}(A^u\otimes B^v): |u|=|v|=N\} . \label{eq:mpu_tensor_word_span} \end{align}

In particular, if \(S_N(A)=M_{D}(\mathbb {C})\) and \(S_N(B)=M_{E}(\mathbb {C})\), then \(S_N(A\boxtimes B)=M_{DE}(\mathbb {C})\). If \(A\) and \(B\) are algebraically normal, then \(A\boxtimes B\) is algebraically normal.

This is the homogeneous-word infrastructure for the tensoring sentence in the proof of the MPU Index Theorem IV.6(ii) [ CPGSV17 , lines 824–845 ] , not a separate theorem stated there. It applies to the retained blocks in the canonical-form decomposition [ CPGSV16 , Section 2.3, lines 214–245 ] .

Proof

For every product word \(w\), let \((u,v)\) be its pointwise component words. The bijection between \(w\) and \((u,v)\) and Theorem 19.4.28 give

\begin{align} (A\boxtimes B)^w & =\operatorname {reind}_{\pi _{D,E}}(A^u\otimes B^v), \end{align}

which proves (121) after taking linear spans. Pure tensors span \(M_{D}(\mathbb {C})\otimes M_{E}(\mathbb {C})\). The Kronecker linear equivalence, followed by simultaneous reindexing along \(\pi _{D,E}\), is an isomorphism onto \(M_{DE}(\mathbb {C})\); this proves the fixed-length consequence. If \(S_{N_A}(A)=M_{D}(\mathbb {C})\) and \(S_{N_B}(B)=M_{E}(\mathbb {C})\) for positive \(N_A,N_B\), then persistence at positive multiples gives

\begin{align} S_{N_A N_B}(A) & =M_{D}(\mathbb {C}), & S_{N_A N_B}(B) & =M_{E}(\mathbb {C}). \end{align}

Applying the fixed-length consequence at \(N_A N_B\) proves algebraic normality of \(A\boxtimes B\).

Let \(A\) and \(B\) be CPSV normal tensors, and suppose that

\begin{align} \sum _i(A^i)^\dagger A^i & =\mathbb {1}_D, & \sum _k(B^k)^\dagger B^k & =\mathbb {1}_E. \label{eq:mpu_tensor_left_canonical_inputs} \end{align}

Then \(A\boxtimes B\) is a CPSV normal tensor and is left-canonical:

\begin{align} \sum _{i,k} \left((A\boxtimes B)^{\pi _{d,e}(i,k)}\right)^\dagger (A\boxtimes B)^{\pi _{d,e}(i,k)} & =\mathbb {1}_{DE}. \label{eq:mpu_tensor_left_canonical_output} \end{align}

This is the retained-normal-block infrastructure for the tensoring sentence in the proof of the MPU Index Theorem IV.6(ii) [ CPGSV17 , lines 824–845 ] , not a separate theorem stated there. The retained-normal-block terminology is that of the canonical-form decomposition in [ CPGSV16 , Section 2.3, lines 214–245 ] ; the two left-canonical identities are explicit infrastructure hypotheses, not an additional assertion attributed to that passage.

Proof

The adjoint and multiplication laws for Kronecker products, together with the fact that simultaneous reindexing preserves adjoints, products, and finite sums, give

\begin{align} & \sum _{i,k} \left((A\boxtimes B)^{\pi _{d,e}(i,k)}\right)^\dagger (A\boxtimes B)^{\pi _{d,e}(i,k)} \\ & \quad =\operatorname {reind}_{\pi _{D,E}} \left(\sum _{i,k} (A^i\otimes B^k)^\dagger (A^i\otimes B^k)\right) \\ & \quad =\operatorname {reind}_{\pi _{D,E}} \left( \left(\sum _i(A^i)^\dagger A^i\right)\otimes \left(\sum _k(B^k)^\dagger B^k\right) \right) \\ & \quad =\operatorname {reind}_{\pi _{D,E}} (\mathbb {1}_D\otimes \mathbb {1}_E)=\mathbb {1}_{DE}, \end{align}

proving (125). Spectral normality gives positive lengths \(N_A,N_B\) at which the two homogeneous word spans are full. Theorem 19.4.29 therefore makes \(A\boxtimes B\) algebraically normal at the common length \(N_A N_B\). The algebraically normal, left-canonical tensor is CPSV normal by Theorem 8.6.3.

Definition 19.4.31 Independent tensor product of matrix product operator tensors

Let \(\mathcal U\) and \(\mathcal V\) have physical dimensions \(d,e\) and auxiliary dimensions \(D,E\), respectively. Their independent tensor product \(\mathcal U\boxtimes \mathcal V\) has dimensions \(de\) and \(DE\) and entries

\begin{align} (\mathcal U\boxtimes \mathcal V)^{(i,k),(j,l)}_{(\alpha ,\gamma ),(\beta ,\delta )} & =\mathcal U^{i,j}_{\alpha ,\beta }\mathcal V^{k,l}_{\gamma ,\delta }. \end{align}

A product-valued physical configuration splits sitewise into one configuration for \(\mathcal U\) and one for \(\mathcal V\).

Definition 19.4.32 Blocked product-alphabet equivalence
#

A blocked word of \(L\) product letters separates canonically into two blocked words:

\begin{align} \{ 0,\ldots ,(de)^L{-}1\} & \simeq \{ 0,\ldots ,d^L e^L{-}1\} . \end{align}

The map decodes the length-\(L\) word, splits every letter into its \(d\) and \(e\) components, and re-encodes the two resulting words.

If \(q\) denotes the blocked product-alphabet equivalence, then each site \(s\) satisfies

\begin{align} \operatorname {dec}_d(q(I)_{1},s) & =\operatorname {dec}_{de}(I,s)_{1}, \\ \operatorname {dec}_e(q(I)_{2},s) & =\operatorname {dec}_{de}(I,s)_{2}. \end{align}
Proof

Each equality follows by composing the inverse product encoding at every site with the corresponding projection.

Theorem 19.4.34 Finite-size operator of an independent tensor product

For every system size \(N\), the sitewise configuration bijection identifies the finite-size operator with

\begin{align} (\mathcal U\boxtimes \mathcal V)^{(N)} & \cong \mathcal U^{(N)}\otimes \mathcal V^{(N)}. \end{align}
Proof

Grouping the two auxiliary indices makes every closed-word matrix a Kronecker product. The identity

\begin{align} \operatorname{Tr}(A\otimes B) & =\operatorname{Tr}(A)\operatorname{Tr}(B) \end{align}

then gives the stated finite-size operator after the sitewise configuration bijection.

Theorem 19.4.35 Independent tensor products preserve matrix product unitarity

The independent tensor product of two matrix product unitary tensors is a matrix product unitary tensor. This is the tensoring operation in the proof of [ CPGSV17 , Theorem IV.6(ii) ] .

Proof

The multiplication rule \((A\otimes B)(C\otimes D)=AC\otimes BD\) gives

\begin{align} & (\mathcal U^{(N)}\otimes \mathcal V^{(N)}) (\mathcal U^{(N)}\otimes \mathcal V^{(N)})^\dagger \\ & \quad =\mathcal U^{(N)}(\mathcal U^{(N)})^\dagger \otimes \mathcal V^{(N)}(\mathcal V^{(N)})^\dagger =\mathbb {1}\otimes \mathbb {1}=\mathbb {1}. \end{align}

Simultaneous reindexing preserves this unitarity equation.

For every blocking length \(L\), separating each blocked product letter by the canonical equivalence \(q\) gives the exact identity

\begin{align} (\mathcal U\boxtimes \mathcal V)^{[L]} & =q^*\bigl(\mathcal U^{[L]}\boxtimes \mathcal V^{[L]}\bigr). \end{align}

The ordering of the two component words agrees with the independent tensor-product convention.

Proof

Evaluate both sides on blocked ket and bra letters. Word multiplication of local Kronecker products gives

\begin{align} (\mathcal U\boxtimes \mathcal V)[I,J] & =\mathcal U[I_1,J_1]\otimes \mathcal V[I_2,J_2]. \end{align}

The two decoding identities identify \(I_1,I_2\) and \(J_1,J_2\) with the corresponding letters of \(q(I)\) and \(q(J)\).

Definition 19.4.37 Source-cut shuffle for an independent tensor product
#

The source-cut shuffle sends

\begin{align} ((\alpha ,j),(\alpha ’,j’)) & \longmapsto ((\alpha ,\alpha ’),(j,j’)). \end{align}

The analogous map applies to column indices.

The source cuts satisfy

\begin{align} M_1(\mathcal U\boxtimes \mathcal V) & \cong M_1(\mathcal U)\otimes M_1(\mathcal V), \\ M_2(\mathcal U\boxtimes \mathcal V) & \cong M_2(\mathcal U)\otimes M_2(\mathcal V). \end{align}

Thus the paper’s right and left orientations are unchanged.

Proof

Evaluate both source cuts entry by entry after the displayed row and column shuffles. Each entry is the product of the corresponding entries of the two source cuts.

The right and left source-cut ranks are multiplicative:

\begin{align} r(\mathcal U\boxtimes \mathcal V) & =r(\mathcal U)r(\mathcal V), \\ \ell (\mathcal U\boxtimes \mathcal V) & =\ell (\mathcal U)\ell (\mathcal V). \end{align}

This is the rank-multiplication ingredient for the tensoring case in the proof of [ CPGSV17 , Theorem IV.6(ii) ] .

Proof

Matrix rank is unchanged by row and column bijections and satisfies \(\operatorname {rank}(A\otimes B)=\operatorname {rank}(A) \operatorname {rank}(B)\).

Let \(\mathcal U^{[L]}\) denote the direct blocking of \(L\) sites of a tensor \(\mathcal U\) with physical dimension \(d\), and let \(r_L\) and \(\ell _L\) be its right and left source-cut ranks. Then

\begin{align} r_{L+1} & \leq d\, r_L, \\ \ell _{L+1} & \leq d\, \ell _L. \end{align}

These are the one-site forms of the two upper bounds in the proof of [ CPGSV17 , Proposition IV.2 ] .

Proof

For the first cut, choose a compact factorization of the source matrix of \(\mathcal U^{[L]}\) through a space of dimension \(r_L\). Split each physical word of length \(L+1\) into its first letter and its remaining word of length \(L\). For suitable rectangular matrices \(A_{1,L}\) and \(B_{1,L}\), the successor source matrix factors as

\begin{align} M_{1,L+1} & =A_{1,L}B_{1,L}. \end{align}

The intermediate space is \(\mathbb C^d\otimes \mathbb C^{r_L}\), and hence

\begin{align} \operatorname {rank}(M_{1,L+1}) & \leq \dim \bigl(\mathbb C^d\otimes \mathbb C^{r_L}\bigr)=d r_L. \end{align}

Interchanging the upper and lower physical indices gives matrices \(A_{2,L}\) and \(B_{2,L}\) with

\begin{align} M_{2,L+1} & =A_{2,L}B_{2,L}. \end{align}

This factorization passes through \(\mathbb C^d\otimes \mathbb C^{\ell _L}\), so

\begin{align} \operatorname {rank}(M_{2,L+1}) & \leq \dim \bigl(\mathbb C^d\otimes \mathbb C^{\ell _L}\bigr)=d\ell _L. \end{align}
Theorem 19.4.41 Source-cut rank bounds between two blocking lengths

If \(L_0\leq L\), then

\begin{align} r_L & \leq d^{L-L_0}r_{L_0}, \\ \ell _L & \leq d^{L-L_0}\ell _{L_0}. \end{align}

These are the two blocking upper bounds used in the proof of [ CPGSV17 , Proposition IV.2 ] .

Proof

Induct on \(L-L_0\). The case \(L=L_0\) is equality. For the right rank, the successor bound and the induction hypothesis give

\begin{align} r_{L+1} & \leq d r_L \leq d\bigl(d^{L-L_0}r_{L_0}\bigr) =d^{L+1-L_0}r_{L_0}. \end{align}

The same multiplication by \(d\) gives

\begin{align} \ell _{L+1} & \leq d\ell _L \leq d\bigl(d^{L-L_0}\ell _{L_0}\bigr) =d^{L+1-L_0}\ell _{L_0}. \end{align}

Let \(d{\gt}0\) and \(L_0\leq L\). Suppose the endpoint source-cut ranks satisfy the supplied product identities

\begin{align} r_{L_0}\ell _{L_0} & =d^{2L_0}, & r_L\ell _L & =d^{2L}. \end{align}

Then the blocking bounds are exact:

\begin{align} r_L & =d^{L-L_0}r_{L_0}, & \ell _L & =d^{L-L_0}\ell _{L_0}. \end{align}

This is the saturation step in the proof of [ CPGSV17 , Proposition IV.2 ] .

Proof

The blocking bounds give

\begin{align} r_L & \leq d^{L-L_0}r_{L_0}, & \ell _L & \leq d^{L-L_0}\ell _{L_0}. \end{align}

Their upper bounds have product

\begin{align} \bigl(d^{L-L_0}r_{L_0}\bigr) \bigl(d^{L-L_0}\ell _{L_0}\bigr) & =d^{2(L-L_0)}d^{2L_0}=d^{2L}=r_L\ell _L. \end{align}

Since \(d{\gt}0\), the latter product is positive, so both \(r_L\) and \(\ell _L\) are positive. Hence neither upper bound can be strict, and both equalities follow.

Definition 19.4.43 Product-index source weight
#

For \(\rho \in M_{D}(\mathbb {C})\), let

\begin{align} \chi \colon \mathbb C^D\otimes \mathbb C^d & \longrightarrow \mathbb C^d\otimes \mathbb C^D, & \chi (e_\alpha \otimes e_j) & =e_j\otimes e_\alpha . \end{align}

In the virtual–physical row order \((\alpha ,j)\) of \(M_1\), define

\begin{align} W_\rho & =\rho \otimes \mathbb {1}_d =\chi ^\dagger (\mathbb {1}_d\otimes \rho )\chi . \end{align}

Thus the source expression \(\mathbb {1}_d\otimes \rho \) is first written in the physical–virtual order \((j,\alpha )\) and then reindexed by \(\chi \) to the virtual–physical order \((\alpha ,j)\). The two Kronecker products are not identified without this reindexing. This is the source weight in [ CPGSV17 , equation labelled Y1Y1X1X1 ] .

Theorem 19.4.44 Positivity of the product-index source weight

If \(\rho \) is positive definite, then \(W_\rho \) is positive definite.

Proof

The Kronecker product of the positive-definite matrices \(\rho \) and \(\mathbb {1}_d\) is positive definite.

Choose compact singular-value decompositions

\begin{align} M_i & =V_i^\dagger D_iU_i, & V_iV_i^\dagger & =\mathbb {1}, & U_iU_i^\dagger & =\mathbb {1}, \end{align}

with intermediate dimensions \(r\) and \(\ell \), respectively. For \(\rho \in M_{D}(\mathbb {C})\), define the compressed positive metric

\begin{align} A & =V_1W_\rho V_1^\dagger . \end{align}

The inverse square root \(A^{-1/2}\), rather than \(A\) itself, is the normalization factor used in the first weighted source factor. This is the source-factor construction of [ CPGSV17 , Section III ] .

Theorem 19.4.46 Positivity of the first compressed source metric

If \(\rho \) is positive definite, then \(A\) is positive definite.

Proof

The map \(V_1^\dagger \) is injective because \(V_1V_1^\dagger =\mathbb {1}_r\). Positive definiteness of \(W_\rho \) is therefore preserved by this congruence.

Let \(\rho \in M_{D}(\mathbb {C})\) be positive definite. Define

\begin{align} X_1 & =V_1^\dagger A^{-1/2}, & Y_1 & =A^{1/2}D_1U_1, & Z_1 & =U_1^\dagger D_1^{-1}A^{-1/2}, \\ X_2 & =V_2^\dagger , & Y_2 & =D_2U_2, & Z_2 & =U_2^\dagger D_2^{-1}. \end{align}

These are the factors in [ CPGSV17 , Section III ] .

Proof

Since \(A\) is positive definite, its support projection is the identity and

\begin{align} A^{-1/2}A^{1/2} & =\mathbb {1}_r, & A^{-1/2}AA^{-1/2} & =\mathbb {1}_r. \end{align}

Therefore

\begin{align} X_1Y_1 & =V_1^\dagger A^{-1/2}A^{1/2}D_1U_1 =V_1^\dagger D_1U_1=M_1, \\ X_1^\dagger W_\rho X_1 & =A^{-1/2}V_1W_\rho V_1^\dagger A^{-1/2} =A^{-1/2}AA^{-1/2}=\mathbb {1}_r, \\ Y_1Z_1 & =A^{1/2}D_1U_1U_1^\dagger D_1^{-1}A^{-1/2} =A^{1/2}A^{-1/2}=\mathbb {1}_r. \end{align}

For the second cut, \(V_2V_2^\dagger =U_2U_2^\dagger =\mathbb {1}_\ell \) and \(D_2D_2^{-1}=\mathbb {1}_\ell \) give

\begin{align} X_2Y_2 & =V_2^\dagger D_2U_2=M_2, \\ X_2^\dagger X_2 & =V_2V_2^\dagger =\mathbb {1}_\ell , \\ Y_2Z_2 & =D_2U_2U_2^\dagger D_2^{-1}=\mathbb {1}_\ell . \end{align}

The two factorizations are

\begin{align} M_1 & =X_1Y_1, & M_2 & =X_2Y_2. \end{align}

These are the source decompositions in [ CPGSV17 , Section III ] .

Proof

The source-factor construction supplies \(M_1=X_1Y_1\). The second compact singular-value decomposition gives \(M_2=X_2Y_2\) directly.

The entries of the two products satisfy

\begin{align} (X_1Y_1)_{(\alpha ,j),(i,\beta )} & =\mathcal U^i_{j,\alpha ,\beta }, \\ (X_2Y_2)_{(\alpha ,i),(j,\beta )} & =\mathcal U^i_{j,\alpha ,\beta }. \end{align}

These are the two graphical entry identifications in [ CPGSV17 , Section III ] .

Proof

Evaluate the two matrix factorizations at \(((\alpha ,j),(i,\beta ))\) and \(((\alpha ,i),(j,\beta ))\), respectively.

The left factors satisfy

\begin{align} X_1^\dagger W_\rho X_1 & =\mathbb {1}_r, & X_2^\dagger X_2 & =\mathbb {1}_\ell , \\ \sum _{x,y}\overline{(X_1)_{x,r}}(W_\rho )_{x,y}(X_1)_{y,r'} & =\delta _{r,r'}, \\ \sum _{\beta ,i}\overline{(X_2)_{(\beta ,i),l}}(X_2)_{(\beta ,i),l'} & =\delta _{l,l'}. \end{align}

These are [ CPGSV17 , Section III ] .

Proof

The source-factor construction supplies \(X_1^\dagger W_\rho X_1=\mathbb {1}_r\); evaluating at \((r,r')\) gives the displayed weighted sum. Since \(X_2=V_2^\dagger \), the coisometry identity \(V_2V_2^\dagger =\mathbb {1}_\ell \) gives \(X_2^\dagger X_2=\mathbb {1}_\ell \); evaluating at \((l,l')\) gives the second displayed entry sum.

For every tensor \(\mathcal U\), physical indices \(p,q\), and auxiliary indices \(a,b\),

\begin{align} \sum _{\beta ,z} \overline{\mathcal U^z_{p,\beta ,a}}\, \mathcal U^z_{q,\beta ,b} & =\sum _l \overline{(Y_2)_{l,(p,a)}}(Y_2)_{l,(q,b)}. \end{align}

Thus the factor indexed by \((p,a)\) is starred and the factor indexed by \((q,b)\) is unstarred. This is the second-cut rotation used in [ CPGSV17 , Theorem III.8, Section III.B, equations (31)–(32) ] . The identity requires no Matrix Product Unitary, dimension, positivity, or simple-tensor hypothesis.

Proof

Substitute the entry factorization \(M_2=X_2Y_2\) into the left-hand side and reorder the finite sums. The contraction of the two \(X_2\) factors is \(X_2^\dagger X_2=\mathbb {1}_\ell \), so the two source indices coincide and the remaining sum is the displayed Gram matrix of \(Y_2\).

The right source factors are recovered from the two source matrices by

\begin{align} Y_1 & =X_1^\dagger W_\rho M_1, & Y_2 & =X_2^\dagger M_2. \end{align}

Here \(W_\rho =\rho \otimes I_d\) in the product-index order \((\text{left virtual},\text{physical})\).

Proof

Multiply \(M_1=X_1Y_1\) on the left by \(X_1^\dagger W_\rho \) and use \(X_1^\dagger W_\rho X_1=I_r\). Similarly, multiply \(M_2=X_2Y_2\) on the left by \(X_2^\dagger \) and use \(X_2^\dagger X_2=I_\ell \).

The right factors satisfy

\begin{align} Y_1Z_1 & =\mathbb {1}_r, & Y_2Z_2 & =\mathbb {1}_\ell . \end{align}

This is [ CPGSV17 , Section III ] .

Proof

The source-factor construction supplies \(Y_1Z_1=\mathbb {1}_r\). For the second cut,

\begin{align} Y_2Z_2 & =D_2U_2U_2^\dagger D_2^{-1} =D_2D_2^{-1}=\mathbb {1}_\ell . \end{align}

Given source-cut decompositions \(M_1=X_1Y_1\) and \(M_2=X_2Y_2\), define

\begin{align} u & : \mathbb C^d\otimes \mathbb C^d \longrightarrow \mathbb C^\ell \otimes \mathbb C^r, \\ v & : \mathbb C^r\otimes \mathbb C^\ell \longrightarrow \mathbb C^d\otimes \mathbb C^d \end{align}

by

\begin{align} u_{(l,r),(i_1,i_2)} & =\sum _\beta (Y_1)_{r,(i_1,\beta )}(X_2)_{(\beta ,i_2),l}, \\ v_{(j_1,j_2),(r,l)} & =\sum _\alpha (X_1)_{(\alpha ,j_1),r}(Y_2)_{l,(j_2,\alpha )}. \end{align}

The source legs of \(u\) have order \(\ell \times r\), whereas those of \(v\) have order \(r\times \ell \). These are the two source tensors introduced after the source-cut decompositions in [ CPGSV17 , Section III ] . The entry formulas also fix the conjugate-transpose orientation. A supplied source-factor witness provides these decompositions. The fixed compact-SVD factors specialize the formulas to the previously defined tensors.

Identify a two-site physical configuration with an ordered pair in \(\{ 0,\ldots ,d{-}1\} \times \{ 0,\ldots ,d{-}1\} \). Then

\begin{align} \operatorname {reindex}(U^{(2)}) & =u^{\mathsf T}\bigl(\operatorname {swap}_{r,\ell }v\bigr)^{\mathsf T}. \end{align}

The transpose records the periodic entry orientation. The map \(\operatorname {swap}_{r,\ell }\) is the explicit equivalence \(r\times \ell \simeq \ell \times r\); the two product orders are not silently identified. This is the bond swap, or one-site translation, in the source-tensor decomposition of [ CPGSV17 , Section III ] .

Proof

At physical pairs \((i_1,i_2)\) and \((j_1,j_2)\), expand the periodic trace as

\begin{align} \sum _{\alpha ,\beta } \mathcal U^{i_1}_{j_1,\alpha ,\beta } \mathcal U^{i_2}_{j_2,\beta ,\alpha }. \end{align}

Use the entry forms of \(M_1=X_1Y_1\) for the first letter and \(M_2=X_2Y_2\) for the second. Reordering the four scalar finite sums gives

\begin{align} \sum _{l,r}u_{(l,r),(i_1,i_2)}v_{(j_1,j_2),(r,l)}. \end{align}

Reindexing both physical configuration spaces and explicitly swapping \((r,l)\) to \((l,r)\) gives the matrix identity.

For \(q=(q_1,q_2)\) and \(j=(j_1,j_2)\), the untraced two-site product is

\begin{align} (\mathcal U^{q_1j_1}\mathcal U^{q_2j_2})_{\alpha ,\gamma } =\sum _{r,l}(X_1)_{(\alpha ,j_1),r} u_{(l,r),q}(Y_2)_{l,(j_2,\gamma )}. \end{align}

The intermediate partial contraction is

\begin{align} \sum _r (X_1)_{(\alpha ,j_1),r}u_{(l,r),(q_1,q_2)} =\sum _\beta \mathcal U^{q_1}_{j_1,\alpha ,\beta } (X_2)_{(\beta ,q_2),l}. \end{align}
Proof

Use the three identities

\begin{align} u_{(l,r),q} & =\sum _\beta (Y_1)_{r,(q_1,\beta )} (X_2)_{(\beta ,q_2),l}, \\ (X_1Y_1)_{(\alpha ,j_1),(q_1,\beta )} & =\mathcal U^{q_1}_{j_1,\alpha ,\beta }, \\ (X_2Y_2)_{(\beta ,q_2),(j_2,\gamma )} & =\mathcal U^{q_2}_{j_2,\beta ,\gamma }. \end{align}

Substitution and reordering of the scalar finite sums give the displayed partial contraction and the untraced two-site identity, with the virtual indices \(\alpha \) and \(\gamma \) left open. Encoding both physical pairs by the standard equivalence \(\{ 0,\ldots ,d{-}1\} \times \{ 0,\ldots ,d{-}1\} \simeq \{ 0,\ldots ,({-}1\} d^2)\) gives the corresponding formula for the concrete two-site block.

The reflected factorization of the same untraced two-site product is

\begin{align} (\mathcal U^{i_1j_1}\mathcal U^{i_2j_2})_{\alpha ,\gamma } =\sum _{l,r}(X_2)_{(\alpha ,i_1),l} v_{(j_2,j_1),(r,l)}(Y_1)_{r,(i_2,\gamma )}. \end{align}

Thus the physical row of \(v\) is \((j_2,j_1)\), while its source column is \((r,l)\). Both orders remain explicit in the formula for the concrete two-site block under \(\{ 0,\ldots ,d{-}1\} \times \{ 0,\ldots ,d{-}1\} \simeq \{ 0,\ldots ,({-}1\} d^2)\).

Proof

Expand the first tensor letter through \(M_2=X_2Y_2\) and the second through \(M_1=X_1Y_1\). Then

\begin{align} \sum _\beta (X_2)_{(\alpha ,i_1),l}(Y_2)_{l,(j_1,\beta )} (X_1)_{(\beta ,j_2),r}(Y_1)_{r,(i_2,\gamma )} = (X_2)_{(\alpha ,i_1),l} v_{(j_2,j_1),(r,l)}(Y_1)_{r,(i_2,\gamma )}. \end{align}

Summing over \(l\) and \(r\) gives the result. The two-site block statement follows by decoding its ket and bra indices into ordered pairs.

Separate the first two physical sites of a length-\((K+2)\) configuration. For an output pair \(q\), an input pair \(j\), and tails \(\tau ,\zeta \), the periodic entry first expands as

\begin{align} U^{(K+2)}_{(q,\tau ),(j,\zeta )} =\sum _{\alpha ,\gamma ,r,l} & (X_1)_{(\alpha ,j_1),r}u_{(l,r),q} (Y_2)_{l,(j_2,\gamma )} (T_{\tau ,\zeta })_{\gamma ,\alpha }, \end{align}

where \(T_{\tau ,\zeta }\) is the open \(K\)-site virtual word. Reconstructing \(Y_2\) gives

\begin{align} U^{(K+2)}_{(q,\tau ),(j,\zeta )} =\sum _{\alpha ,r,l,\beta ,i} & (X_1)_{(\alpha ,j_1),r}u_{(l,r),q} \overline{(X_2)_{(\beta ,i),l}} (\mathcal U^{ij_2}T_{\tau ,\zeta })_{\beta ,\alpha }. \end{align}

No power of \(d\) occurs in either algebraic expansion.

Proof

Reindexing each configuration into its first two coordinates and its tail gives

\begin{align} U^{(K+2)}_{(q,\tau ),(j,\zeta )} & =\operatorname{tr}\! \left[ (\mathcal U^{q_1j_1}\mathcal U^{q_2j_2}) T_{\tau ,\zeta }\right]. \end{align}

Insert the open two-site factorization from the preceding theorem. The recovery identity and its entrywise evaluation are

\begin{align} Y_2 & =X_2^\dagger M_2, \\ (Y_2)_{l,(j_2,\gamma )} & =\sum _{\beta ,i}\overline{(X_2)_{(\beta ,i),l}} \mathcal U^i_{j_2,\beta ,\gamma }. \end{align}

Substitution combines the last raw tensor letter with \(T_{\tau ,\zeta }\) by matrix multiplication and gives the second displayed expansion.

Definition 19.4.59 Normalized open-tail coefficient

For

\begin{align} A^{\tau ,\zeta }_{i,j_2} & =\mathcal U^{ij_2}T_{\tau ,\zeta }, \end{align}

define the coefficient indexed by \(l,r,l',r'\) to be

\begin{align} C^{(K)}_{lr,l'r'}=d^{-K} \sum _{\substack {\tau ,\zeta ,j_1,j_2,\alpha ,\alpha ’, \\ \beta ,\beta ’,i,i’}} & (X_1)_{(\alpha ,j_1),r} \overline{(X_1)_{(\alpha ',j_1),r'}} \overline{(X_2)_{(\beta ,i),l}} (X_2)_{(\beta ',i'),l'} \\ & \mathrel {\phantom{=}}\cdot (A^{\tau ,\zeta }_{i,j_2})_{\beta ,\alpha } \overline{(A^{\tau ,\zeta }_{i',j_2})_{\beta ',\alpha '}}. \end{align}

This is the coefficient obtained from the two open-tail expansions in the normalized Gram contraction in the proof of the source-tensor isometry lemma in [ CPGSV17 , Section III ] . The external factor \(u_{(l,r),q}\) is unstarred, while \(u_{(l',r'),p}\) is starred. Thus the displayed coefficient contains the unique normalization factor \(d^{-K}\).

Definition 19.4.60 Forward source-\(u\) kernel
#

For a tail of length \(K\), define

\begin{align} F^{(K)}_{ii';\beta \beta ',\alpha \alpha '} =d^{-K}\sum _{\tau ,\zeta ,j_2} & (\mathcal U^{ij_2}T_{\tau ,\zeta })_{\beta ,\alpha } \overline{(\mathcal U^{i'j_2}T_{\tau ,\zeta })_{\beta ',\alpha '}}. \end{align}

The expression contains exactly one normalization factor \(d^{-K}\).

Theorem 19.4.61 Normalized output-layer tail entry

Let \(W_{\mathrm{out}}\) be the output-first double layer. Then

\begin{align} (E_{\mathrm{out}}^K)_{(\gamma ,\gamma '),(\alpha ,\alpha ')} =d^{-K}\sum _{\tau ,\zeta } (T_{\tau ,\zeta })_{\gamma ,\alpha } \overline{(T_{\tau ,\zeta })_{\gamma ',\alpha '}}. \end{align}
Proof

Expand the normalized diagonal of the blocked output layer, reindex a blocked physical index by its decoded word, and evaluate the product double layer. The physical adjoint contributes the conjugated second word, while the blocked physical dimension supplies \(d^{-K}\).

The forward kernel is

\begin{align} F^{(K)}_{ii'} =W_{\mathrm{out}}^{ii'}E_{\mathrm{out}}^K. \end{align}

Both doubled bond indices put the unstarred component before the starred component.

Proof

Expand the matrix product over its doubled bond index. The first output-layer letter supplies the sum over \(j_2\), and the normalized diagonal power supplies the common output word and contracted input word. Reordering finite sums gives the defining forward kernel without reversing the virtual chain.

On the range of the first source cut,

\begin{align} X_1X_1^\dagger W_\rho M_1=M_1, \qquad W_\rho =\rho \otimes 1_d. \end{align}
Proof

Substitute \(X_1^\dagger W_\rho M_1=Y_1\) and then use \(M_1=X_1Y_1\).

Theorem 19.4.64 Second source-cut range projection

On the range of the second source cut,

\begin{align} X_2X_2^\dagger M_2=M_2. \end{align}

This does not assert \(X_2X_2^\dagger =1\) on the ambient product-index space.

Proof

Substitute \(X_2^\dagger M_2=Y_2\) and then use \(M_2=X_2Y_2\).

Theorem 19.4.65 Expanded normalized source-\(u\) metric

Assume \(d{\gt}0\), let \(\mathcal U\) be an MPU, and let \(\rho \) be positive definite. Expanding both \((K+2)\)-site periodic entries through \(X_1,u,X_2\) and the open tail gives

\begin{align} d^{-K}\sum _{\tau ,j,\zeta } \mathcal A_{q;\tau ,j,\zeta } \overline{\mathcal A_{p;\tau ,j,\zeta }} =\delta _{p,q}, \end{align}

where \(\mathcal A_{q;\tau ,j,\zeta }\) is the full source-\(u\) expansion with the \(q\) entry unstarred. This theorem evaluates the displayed expansion directly. It neither regroups the sum through \(C^{(K)}_{lr,l'r'}\) nor identifies it with the ordinary Gram matrix \(u^\dagger u\).

Proof

Replace each source-factor sum by its periodic MPO entry. Reindex the input configuration into its first two coordinates and its length-\(K\) tail, then apply normalized output-first MPU coisometry.

Let \(\mathcal U\) be an MPO tensor, let \(\rho \) be positive definite, and let \(p,q\in \{ 0,\ldots ,d{-}1\} \times \{ 0,\ldots ,d{-}1\} \). Write \(W_{\mathrm{out}}\) for the output-first double layer of \(\mathcal U\). Then

\begin{align} \sum _{l,r}u_{(l,r),q}\overline{u_{(l,r),p}} =\sum _{\beta ,\delta ,\gamma ,\alpha } & \rho _{\alpha ,\gamma } (W_{\mathrm{out}}^{q_1p_1})_{(\gamma ,\alpha ),(\beta ,\delta )} \\ & \qquad \cdot (X_2X_2^\dagger )_{(\beta ,q_2),(\delta ,p_2)}. \end{align}

Thus \(q\) is the unstarred ket index, \(p\) is the starred bra index, and each doubled bond is ordered as \((\text{ket},\text{bra})\). This is the first closed network in [ CPGSV17 , Eq. uUnitary and lines 545–556 ] .

Proof

Expand both source-\(u\) entries and contract their common source indices. Weighted recovery of \(Y_1\), followed by the first source-cut range identity, replaces \(Y_1^\dagger Y_1\) by \(M_1^\dagger W_\rho M_1\). Expanding \(M_1\) and the output-first double layer gives the displayed \(\rho \)-weighted letter, while the remaining \(X_2\) factors form the stated range projector.

Theorem 19.4.67 Closed double-layer trace for the normalized output tail

Assume \(d{\gt}0\). For any MPO tensor \(\mathcal U\), tail length \(K\), and \(p,q\in \{ 0,\ldots ,d{-}1\} \times \{ 0,\ldots ,d{-}1\} \), let \(W_{\mathrm{out}}\) be its output-first double layer and let \(E_{\mathrm{out}}\) be the normalized physical diagonal of \(W_{\mathrm{out}}\). Then

\begin{align} d^{-K}\sum _{\tau ,\eta } U^{(K+2)}_{(q,\tau ),\eta } \overline{U^{(K+2)}_{(p,\tau ),\eta }} =\operatorname{tr}\! \left[ W_{\mathrm{out}}^{q_1p_1} W_{\mathrm{out}}^{q_2p_2}E_{\mathrm{out}}^K\right]. \end{align}

The retained letters occur in \((q,p)\) order, so \(q\) is unstarred and \(p\) is starred; their doubled bonds are in \((\text{ket},\text{bra})\) order. This is the normalized closed output tail in [ CPGSV17 , Eq. uUnitary and lines 550–556 ] .

Proof

Close the contracted input configuration as the periodic MPO of the output-first double layer. Splitting off the two retained sites leaves a common diagonal word of length \(K\). Expanding the normalized diagonal power as the normalized sum of those words and using linearity of the trace gives the displayed closed trace.

Define the second-cut metric by \(H_2=Z_2Z_2^\dagger \). Then the range projector of the second source cut factors as

\begin{align} X_2X_2^\dagger =M_2H_2M_2^\dagger . \end{align}

This is a factorization through the chosen right inverse; it does not assert that \(H_2\), \(X_2X_2^\dagger \), or \(M_2M_2^\dagger \) is the identity.

Proof

From \(M_2=X_2Y_2\) and \(Y_2Z_2=1\) one obtains \(M_2Z_2=X_2\). Multiply this identity by its adjoint and reassociate the three factors.

Theorem 19.4.69 Source-\(u\) Gram entry with the second-cut metric

Let \(\mathcal U\) be an MPO tensor, let \(\rho \) be positive definite, and let \(p,q\in \{ 0,\ldots ,d{-}1\} \times \{ 0,\ldots ,d{-}1\} \) be the retained physical pairs. No MPU or positive-dimension hypothesis is required. With \(H_2=Z_2Z_2^\dagger \), the ordinary source-\(u\) Gram entry is

\begin{align} \sum _{l,r}u_{(l,r),q}\overline{u_{(l,r),p}} =\sum _{\beta ,\delta ,\gamma ,\alpha ,j,a,k,c} & \rho _{\alpha ,\gamma } (W_{\mathrm{out}}^{q_1p_1})_{(\gamma ,\alpha ),(\beta ,\delta )} \\ & \quad \cdot U^{q_2j}_{\beta ,a} (H_2)_{(j,a),(k,c)}\overline{U^{p_2k}_{\delta ,c}}. \end{align}

The \(q\) factors are unstarred and the \(p\) factors are starred.

Proof

Substitute the second-cut metric factorization into the closed output-letter formula and expand both matrix products. Reordering the finite sums gives the displayed eight-index expression.

Suppose a direct block of length \(K\) satisfies the two-letter contraction with specified vectors \((a|\) and \(|b)\). For an arbitrary MPO tensor, one suffix/prefix extension preserves the same insertion \(|b)(a|\) at direct length \(K+1\). For any MPU tensor, specified vectors satisfying this two-letter identity also satisfy the one-letter identity with the same vectors. Applying that implication after two extensions gives, at direct length \(K+2\),

\begin{align} (a|W^{IJ}|b) & =\delta _{I,J}, \\ W^{IJ}W^{LM} & =W^{IJ}|b)(a|W^{LM}. \end{align}

No positivity or nonzero-dimension hypothesis is required. This is the corollary following [ CPGSV17 , Proposition III.3 ] .

Proof

Extend the supplied two-letter identity twice by overlapping suffix and prefix windows. For \(x=(a|W^{IJ}|b)\), the constant configurations of lengths two and three give \(x^2=\delta _{I,J}\) and \(x^3=\delta _{I,J}\), which force \(x=\delta _{I,J}\). Physical blocking preserves MPU unitarity, so this one-letter identity and the twice-extended two-letter identity hold simultaneously at direct length \(K+2\) with the original supplied witnesses.

Definition 19.4.71 Reflected blocked-transfer coordinate maps

The doubled-bond coordinate reflection exchanges the two virtual factors,

\begin{align} (a,b) & \longmapsto (b,a). \end{align}

Independently, spatial reflection reverses a blocked word of length \(K\),

\begin{align} (i_0,\ldots ,i_{K-1}) & \longmapsto (i_{K-1},\ldots ,i_0). \end{align}

These are support coordinates for the proof of [ CPGSV17 , Lemma III.7 in Section III.B ] , rather than separate definitions printed there. The blocked-word reversal is not used in the transfer transport below; conjugate transposition there reverses only the order of the two blocked letters.

The doubled-bond equivalence and the blocked-word reversal are involutions. The latter sends the letter in position \(k\) to the letter in position \(K-1-k\) and reverses the encoded word.

Proof

Decode the doubled bond, apply the product swap, and encode it again. Regard a blocked index as a word indexed by the \(K\) ordered positions, precompose with \(k\mapsto K-1-k\), and reverse the corresponding finite list.

Let \(R(\mathcal U)\) denote the normalized physical diagonal of the double layer of an MPO tensor \(\mathcal U\), and let \(s(a,b)=(b,a)\) be the doubled-bond swap. Then

\begin{align} R(\mathcal U^\sharp )_{x,y}=R(\mathcal U)_{s(x),s(y)}. \end{align}

This identity requires neither the MPU property nor a positivity or positive-dimension hypothesis. It is a coordinate identity supporting the calculation in the proof of [ CPGSV17 , Lemma III.7 in Section III.B ] , not a separate theorem printed there.

Proof

Expand the two normalized diagonals. Physical adjunction exchanges the two physical indices, and exchanging the two finite physical sums leaves precisely the simultaneous doubled-bond swap on the row and column.

Let \(\mathcal U\) be an MPO tensor. Write \(E\) for the transfer matrix of its normalized flattening, \(E^\sharp \) for that of \(\mathcal U^\sharp \), and \(E_{\mathrm{out}}\) for the normalized diagonal of the double layer of \(\mathcal U^\sharp \). If \(d{\gt}0\) and

\begin{align} E^J=|\rho )(1|, \end{align}

then

\begin{align} (E^\sharp )^J=|\rho ^{\mathsf T})(1|. \end{align}

No MPU, positivity, trace, positive bond-dimension, or positivity of \(J\) is needed for this implication.

Suppose in addition that \(D{\gt}0\) and \(\operatorname{tr}(\rho )=1\), and set \(K=JD^2\). Then the normalized diagonal \(R_K\) of the double layer of the \(K\)-site block of \(\mathcal U\) satisfies

\begin{align} R_K=|\rho )(1|. \end{align}

More generally, if \(d{\gt}0\), \(\rho {\gt}0\), and the normalized diagonal \(R_L\) of an \(L\)-site block satisfies \(R_L=|\rho )(1|\), write \(R_L^\sharp \) for the normalized diagonal of the corresponding block of \(\mathcal U^\sharp \). Then

\begin{align} R_L^{\sharp \dagger }=|1)(\rho |. \end{align}

Consequently, under the combined assumptions \(d,D{\gt}0\), \(\rho {\gt}0\), \(\operatorname{tr}(\rho )=1\), and \(E^J=|\rho )(1|\),

\begin{align} R_K^{\sharp \dagger } & =|1)(\rho |, & E_{\mathrm{out}}^K & =|\rho ^{\mathsf T})(1|. \end{align}

Here all rank-one operators use the column-stacking identification \(M_D(\mathbb C)\cong \mathbb C^{D^2}\).

Proof

The blocked normalized diagonal is the corresponding transfer power. Reindex the reflected diagonal by the doubled-bond swap to obtain the transpose weight. When \(\rho \) is positive definite, its Hermitian symmetry converts the conjugate-transposed reflected diagonal to \(|1)(\rho |\). The trace-one condition makes \(|\rho )(1|\) idempotent, so the additional positive \(D^2\) power leaves it unchanged. Conjugating once more gives the stated output-layer orientation.

Let \(\mathcal U\) be an MPU with \(d,D{\gt}0\). Let \(E\) be the transfer matrix of its normalized flattening, and suppose

\begin{align} E^J=|\rho )(1|, \end{align}

where \(\rho {\gt}0\), \(\operatorname{tr}(\rho )=1\), and \(J{\gt}0\). Set \(K=JD^2\). For \(n\in \{ K,K+2\} \), let \(\widehat W_n\) be the double layer of the \(n\)-site block of \(\mathcal U^\sharp \). Then, for all blocked physical indices,

\begin{align} (\rho |\widehat W_n^{IJ\dagger }|1) & =\delta _{I,J}, \\ \widehat W_n^{LM\dagger }\widehat W_n^{IJ\dagger } & =\widehat W_n^{LM\dagger }|1)(\rho | \widehat W_n^{IJ\dagger }. \end{align}

The same supplied vectors occur at both lengths; the \(K+2\) identities do not replace them by existentially chosen witnesses.

Proof

Apply the direct-block simple contractions to the reflected transfer power \(|\rho ^{\mathsf T})(1|\). Positivity identifies the conjugate-transposed vectors with \((\rho |\) and \(|1)\), while conjugate transposition reverses the order of the two blocked letters. Overlapping-window reblocking preserves these witnesses through the two additional sites. These contractions are support identities for the graphical proof of [ CPGSV17 , Lemma III.7 and its proof in Section III.B ] ; they are not asserted as a separate printed theorem of that paper.

Remark 19.4.76 Scope of the closed source-\(u\) formulas
#

The first theorem rewrites the ordinary source-\(u\) Gram entry, while the second independently rewrites the normalized closed output tail. They do not identify these two expressions or justify inserting the source-cut range projections through the intervening chain. That range-restricted equality remains separate. Neither theorem asserts an ambient \(X_2X_2^\dagger =\mathbb {1}\), \(M_2M_2^\dagger =\mathbb {1}\), or \(Y_2Y_2^\dagger =\mathbb {1}\) identity.

Theorem 19.4.77 Terminal source-factor boundary contraction

If the normalized internal tail supplies

\begin{align} \rho _{\alpha ',\alpha }\, \delta _{\beta ,\beta '}\delta _{i,i'}, \end{align}

then the remaining boundary sum is

\begin{align} \sum _{\substack {j,\alpha ,\alpha ’,\beta ,i}} & (X_1)_{(\alpha ,j),r} \overline{(X_1)_{(\alpha ',j),r'}} \overline{(X_2)_{(\beta ,i),l}} (X_2)_{(\beta ,i),l'}\rho _{\alpha ',\alpha } =\delta _{l,l'}\delta _{r,r'}. \end{align}
Proof

Reorder the scalar factors so that the sum separates into the weighted \(X_1\) contraction and the unweighted \(X_2\) contraction. Evaluating

\begin{align} X_1^\dagger W_\rho X_1 & =\mathbb {1}_r, & X_2^\dagger X_2 & =\mathbb {1}_\ell \end{align}

at \((r',r)\) and \((l,l')\), respectively, gives the two Kronecker deltas. This is precisely the terminal boundary simplification in the proof of the source-tensor isometry lemma in [ CPGSV17 , Section III ] .

Remark 19.4.78 Scope of the open-tail algebra
#

The preceding results evaluate the fully expanded normalized metric, but do not prove that regrouping it through \(C^{(K)}_{lr,l'r'}\) gives the ordinary Gram matrix \(u^\dagger u\). The remaining independent equality is between that Gram matrix and the normalized closed \((K+2)\)-site output tail, with both external source-\(u\) tensors retained before moving the cut. No result uses \(X_2X_2^\dagger =\mathbb {1}\), \(M_2M_2^\dagger =\mathbb {1}\), or \(Y_2Y_2^\dagger =\mathbb {1}\).

Put

\begin{align} Y & =Y_1\otimes Y_2, & Z & =Z_1\otimes Z_2, \end{align}

with source-bond order \(r\times \ell \). Thus \(Y\) maps \((\mathbb {C}^d\otimes \mathbb {C}^D)^{\otimes 2}\) to \(\mathbb {C}^r\otimes \mathbb {C}^\ell \), and \(Z\) maps in the opposite direction. The regrouping of the open indices is

\begin{align} R\bigl((i,a),(j,b)\bigr) & =\bigl((i,j),\operatorname {flat}(a,b)\bigr), \end{align}

where \(\operatorname {flat}:\{ 0,\ldots ,D{-}1\} \times \{ 0,\ldots ,D{-}1\} \simeq \{ 0,\ldots ,({-}1\} D^2)\) is the canonical product enumeration. The dressed source-\(v\) Gram equation is

\begin{align} Y^\dagger (v^\dagger v)Y & =Y^\dagger Y. \end{align}

These are the tensorized factors surrounding \(v^\dagger v\) in [ CPGSV17 , Theorem III.8, Section III.B, equations (31)–(32) ] .

Let \(p=(p_1,p_2)\) be the starred physical pair, \(q=(q_1,q_2)\) the unstarred pair, and let \(a=(a_1,a_2)\) and \(b=(b_1,b_2)\) be virtual paths. The weighted first-cut contraction is

\begin{align} \sum _r \overline{(Y_1)_{r,(p_1,a_1)}}(Y_1)_{r,(q_1,b_1)} & =\sum _{\alpha ,\alpha ',j_1} \overline{U^{p_1j_1}_{\alpha a_1}} \rho _{\alpha \alpha '}U^{q_1j_1}_{\alpha 'b_1}. \end{align}

The entry of two ordinary double-layer letters from row \((a_1,b_1)\) to column \((a_2,b_2)\) is

\begin{align} \sum _{\alpha ,\beta ,j_1,j_2} \overline{U^{j_1p_1}_{a_1\alpha }} U^{j_1q_1}_{b_1\beta } \overline{U^{j_2p_2}_{\alpha a_2}} U^{j_2q_2}_{\beta b_2}. \end{align}

Set

\begin{align} \rho ’_x & =\operatorname {vec}(\rho )_{\operatorname {flat}^{-1}(x)}, & \Phi ’_x & =\operatorname {vec}(\mathbb {1}_D)_{\operatorname {flat}^{-1}(x)}. \end{align}

Inserting \(|\rho '\rangle \! \langle \Phi '|\) between the two letters gives

\begin{align} & \bigl(W^{p_1q_1}|\rho ’\rangle \! \langle \Phi ’|W^{p_2q_2}\bigr) _{(a_1,b_1),(a_2,b_2)} \\ & \quad =\sum _{\alpha ,\alpha ',\beta ,j_1,j_2} \overline{U^{j_1p_1}_{a_1\alpha }} U^{j_1q_1}_{b_1\alpha '}\rho _{\alpha '\alpha } \overline{U^{j_2p_2}_{\beta a_2}}U^{j_2q_2}_{\beta b_2}. \end{align}

Finally, the Gram entry of \(Y_1\otimes Y_2\) expands as

\begin{align} \sum _{\alpha ,\alpha ',\beta ,j_1,j_2} \overline{U^{p_1j_1}_{\alpha a_1}} \rho _{\alpha \alpha '}U^{q_1j_1}_{\alpha 'b_1} \overline{U^{j_2p_2}_{\beta a_2}} U^{j_2q_2}_{\beta b_2}. \end{align}
Proof

Expand the ordinary and rank-one-inserted matrix products entrywise. For the tensorized Gram entry, use \(X_1^\dagger (\rho \otimes \mathbb {1}_d)X_1=\mathbb {1}_r\) on the first cut and \(X_2^\dagger X_2=\mathbb {1}_\ell \) on the second cut, then reorder the finite sums. No ambient identity \(X_2X_2^\dagger =\mathbb {1}\) is used.

Let \(\rho {\gt}0\). For output pair \(z=(z_1,z_2)\), physical pair \(p=(p_1,p_2)\), and virtual pair \(a=(a_1,a_2)\), the entry of \(v(Y_1\otimes Y_2)\) is

\begin{align} \bigl(v(Y_1\otimes Y_2)\bigr)_{z,((p_1,a_1),(p_2,a_2))} =\sum _{\gamma ,l} U^{p_1z_1}_{\gamma a_1} (Y_2)_{l,(z_2,\gamma )}(Y_2)_{l,(p_2,a_2)}. \end{align}

The first source cut has contracted to one local tensor entry, while the two factors from the second cut remain in the same summand. No partial second-cut Gram matrix has been formed. This coupling is the source-cut ordering in the calculation of [ CPGSV17 , Theorem III.8, Section III.B, equations (31)–(32) ] .

Proof

Expand the matrix product and the tensor product, reorder the three finite sums, and contract \(X_1Y_1\) by the first source-cut entry factorization. Keep both \(Y_2\) factors together throughout.

Theorem 19.4.82 Tensorized source right inverse

The tensorized source factors satisfy

\begin{align} YZ & =\mathbb {1}_{r\ell }. \end{align}
Proof

Multiplicativity of the Kronecker product gives

\begin{align} YZ & =(Y_1Z_1)\otimes (Y_2Z_2) =\mathbb {1}_r\otimes \mathbb {1}_\ell =\mathbb {1}_{r\ell }. \end{align}
Theorem 19.4.83 Regrouped entries of the dressed Gram equation

For every positive definite source weight \(\rho \), and for \(x=R^{-1}(p,a)\) and \(y=R^{-1}(q,b)\), the dressed Gram equation is equivalent to

\begin{align} \sum _t\left(\sum _s\overline{Y_{s,x}} \sum _z\overline{v_{z,s}}v_{z,t}\right)Y_{t,y} & =\sum _t\overline{Y_{t,x}}Y_{t,y}. \end{align}

Here \(p,q\in \{ 0,\ldots ,d{-}1\} \times \{ 0,\ldots ,d{-}1\} \), \(a,b\in \{ 0,\ldots ,({-}1\} D^2)\), the indices \(s,t\) range over \(\{ 0,\ldots ,r{-}1\} \times \{ 0,\ldots ,\ell {-}1\} \), and \(z\) ranges over \(\{ 0,\ldots ,d{-}1\} \times \{ 0,\ldots ,d{-}1\} \).

Proof

Evaluate both sides of \(Y^\dagger (v^\dagger v)Y=Y^\dagger Y\) at \((x,y)\) and expand the three matrix products. Conversely, apply the displayed equality after regrouping each pair of open source-cut indices by \(R\).

Theorem 19.4.84 Right-inverse cancellation of a dressed Gram equation

If \(Y\) has a right inverse \(Z\), so that \(YZ=\mathbb {1}\), then for every square matrix \(G\) on the row space of \(Y\),

\begin{align} Y^\dagger GY=Y^\dagger Y & \quad \Longleftrightarrow \quad G=\mathbb {1}. \end{align}
Proof

Conjugate the first equality by \(Z^\dagger \) on the left and \(Z\) on the right. Since \(Z^\dagger Y^\dagger =(YZ)^\dagger =\mathbb {1}\), both outer factors cancel. The reverse implication follows by substituting \(G=\mathbb {1}\).

Theorem 19.4.85 Cancellation of the source-\(v\) dressing

For every positive definite source weight \(\rho \),

\begin{align} Y^\dagger (v^\dagger v)Y=Y^\dagger Y & \quad \Longleftrightarrow \quad v^\dagger v=\mathbb {1}_{r\ell }. \end{align}
Proof

Apply right-inverse cancellation with \(G=v^\dagger v\) and \(Z=Z_1\otimes Z_2\).

Theorem 19.4.86 Right-rank bound

If the physical and auxiliary dimensions are \(d\) and \(D\), respectively, then

\begin{align} r & \leq Dd. \end{align}
Proof

The matrix \(M_1\) has \(dD\) columns. Its rank is at most its number of columns, hence at most \(dD=Dd\).

Theorem 19.4.87 Left-rank bound

If the physical and auxiliary dimensions are \(d\) and \(D\), respectively, then

\begin{align} \ell & \leq Dd. \end{align}
Proof

The matrix \(M_2\) has \(dD\) columns. Its rank is at most its number of columns, hence at most \(dD=Dd\).

19.5 Full-support canonical representatives

Definition 19.5.1 Full support

Let

\begin{align} A^i = V^*\left(\bigoplus _{k=1}^r \mu _k A_k^i\right)V \end{align}

be CPSV canonical-form data in ambient bond dimension \(D\), where every \(A_k\) has positive bond dimension \(D_k\). The data have full support when

\begin{align} \sum _{k=1}^rD_k=D. \end{align}

Let \(d,D{\gt}0\), let \(U\) be an MPU tensor of physical dimension \(d\) and bond dimension \(D\), and let \(\widetilde U=d^{-1/2}U\) be its normalized doubled-index tensor. There are an integer \(D_{\mathrm{red}}{\gt}0\), a normal tensor \(A_{\mathrm{red}}\) of bond dimension \(D_{\mathrm{red}}\), and canonical-form-II data for \(A_{\mathrm{red}}\) such that, for every integer \(N{\gt}0\) and every word \(\sigma \) of length \(N\),

\begin{align} D_{\mathrm{red}} & \leq D, \\ V^{(N)}(A_{\mathrm{red}})_\sigma & =V^{(N)}(\widetilde U)_\sigma , \\ \sum _{k=1}^{r}D_k & =D_{\mathrm{red}}. \end{align}

This organizes the dimension-bounded canonical replacement used in [ CPGSV17 , lines 257–294 and 319–326 ] . Its irreducible reduction and canonical-form-II gauge are those of [ CPGSV16 , lines 195–255 and 1058–1077 ] . The shifted transfer argument is the one used in [ CPGSV17 , lines 344–356 ] .

Proof

Remove the zero irreducible blocks of \(\widetilde U\) and write

\begin{align} A_0^i & =\bigoplus _{k=1}^{r}B_k^i, & \sum _{k=1}^{r}D_k & \leq D, \\ V^{(N)}(A_0)_\sigma & =V^{(N)}(\widetilde U)_\sigma , \end{align}

for every \(N{\gt}0\) and every word \(\sigma \) of length \(N\). At \(N=2\) the common self-overlap is one, so the retained sum is nonempty and \(D_0=\sum _kD_k{\gt}0\). For every integer \(N{\gt}1\), overlap invariance gives

\begin{align} \operatorname{tr}(E_{A_0}^{\, N}) & =\sum _\sigma \left|V^{(N)}(A_0)_\sigma \right|^2 =\sum _\sigma \left|V^{(N)}(\widetilde U)_\sigma \right|^2 =1. \label{eq:mpu_reduced_trace} \end{align}

Let \(J_k:\mathbb C^{D_k}\to \mathbb C^{D_0}\) be the inclusion of the \(k\)th block. Then

\begin{align} J_k^\dagger J_k & =\mathbb {1}, & A_0^iJ_k & =J_kB_k^i, \\ \mathcal{E}_{A_0}(J_kXJ_k^\dagger ) & =J_k\mathcal{E}_{B_k}(X)J_k^\dagger . \end{align}

Thus every nonzero transfer eigenvalue of \(B_k\) is a nonzero eigenvalue of \(E_{A_0}\) and hence equals one by (257). Perron–Frobenius theory supplies \(\rho _k{\gt}0\) and \(s_k{\gt}0\) with \(\mathcal{E}_{B_k}(\rho _k)=s_k\rho _k\). Therefore \(s_k=1\), and every peripheral eigenvalue of \(B_k\) is also one. Each \(B_k\) is consequently normal.

The direct sum now has literal canonical-form data with full support. Apply the nonsingular blockwise canonical-form-II gauge. It preserves the block dimensions and all periodic vectors, so the resulting tensor \(A_{\mathrm{red}}\) still satisfies

\begin{align} \sum _{k=1}^{r}D_k & =D_{\mathrm{red}}, & \operatorname{tr}(E_{A_{\mathrm{red}}}^{\, N}) & =1, \end{align}

for every integer \(N{\gt}1\). The shifted traces force \(r=1\), spectral radius one, and primitivity. Full support identifies the unique block with the ambient bond space, so \(A_{\mathrm{red}}\) is normal.

Let \(d,D{\gt}0\) and let \(U\) be an MPU tensor. There are \(D_{\mathrm{red}}{\gt}0\) and an MPU tensor \(U_{\mathrm{red}}\) of bond dimension \(D_{\mathrm{red}}\) such that, for every integer \(N{\gt}0\),

\begin{align} D_{\mathrm{red}} & \leq D, \\ U_{\mathrm{red}}^{(N)} & =U^{(N)}. \end{align}

The normalized doubled-index tensor of \(U_{\mathrm{red}}\) has canonical-form-II data whose retained block dimensions sum to \(D_{\mathrm{red}}\). No equality is asserted at length zero, and no virtual gauge is asserted between the two potentially unequal-dimensional bond spaces.

Proof

Apply Theorem 19.5.2 and write its representative as \(A_{\mathrm{red}}\). Define

\begin{align} U_{\mathrm{red}}^{ij} & =\sqrt d\, A_{\mathrm{red}}^{(i,j)}. \end{align}

Since \(d{\gt}0\), normalization recovers \(A_{\mathrm{red}}\) exactly. For every \(N{\gt}0\) and every ket–bra pair \((\sigma ,\tau )\), the doubled-index coefficient identity gives

\begin{align} (U_{\mathrm{red}}^{(N)})_{\sigma ,\tau } & =d^{N/2}V^{(N)}(A_{\mathrm{red}})_{(\sigma ,\tau )} \\ & =d^{N/2}V^{(N)}(\widetilde U)_{(\sigma ,\tau )} \\ & =(U^{(N)})_{\sigma ,\tau }. \end{align}

Hence the periodic operators agree at every positive length. For \(N{\gt}1\) the right side is unitary, so \(U_{\mathrm{red}}\) is an MPU.

Suppose that canonical-form data have full support:

\begin{align} \sum _{k=1}^rD_k & =D. \end{align}

Then every positive physical blocking of these data also has full support.

Proof

Positive blocking supplies canonical-form data with the same block dimensions, so the full-support sum is unchanged.

A bijective relabeling of the physical alphabet preserves full support of canonical-form data.

Proof

Physical relabeling leaves the blocks and their bond dimensions unchanged, so the full-support sum is unchanged.

Theorem 19.5.6 An ambient-dimensional block is irreducible

If a block has bond dimension \(D\), then the ambient tensor \(A\) is irreducible.

Proof

If \(k\) is the block with \(D_k=D\), the total-dimension bound gives

\begin{align} D = D_k \leq \sum _{j=1}^r D_j \leq D. \end{align}

Since every retained block dimension is positive, these equalities force \(k\) to be the only retained block. After reindexing its bond coordinates, the ambient coisometry \(U\) is square, hence unitary, and reconstruction reads

\begin{align} A^i = U^*\bigl(\mu _k B^i\bigr)U, \end{align}

where \(B\) is the reindexed normal block and \(\mu _k\ne 0\). Thus Theorem D.3.4 transports irreducibility of \(B\) through the nonzero rescaling and unitary conjugation.

Theorem 19.5.7 Unique block with full support

Canonical-form data with one block and full support reconstruct an irreducible ambient tensor.

Proof

Let \(k\) be the unique block. The singleton sum and full-support identity give

\begin{align} \sum _{j=1}^{r} D_j = D_k = D. \end{align}

Thus \(k\) has ambient bond dimension, so Theorem 19.5.6 applies.

Theorem 19.5.8 Normality of a unique full-support block

Suppose that canonical-form data have one block and full support. If the transfer map has spectral radius one and is primitive, then the ambient tensor is a normal tensor.

Proof

Combine the preceding irreducibility conclusion with the supplied spectral-radius and primitivity properties.

19.6 Transfer-matrix identities

Let \(D\) be a positive integer and let \(T:M_D(\mathbb C)\to M_D(\mathbb C)\) be a linear map. The transfer matrix \(\widehat T\) of \(T\) under the column-stacking identification \(M_D(\mathbb C)\cong \mathbb C^{D^2}\) is defined in the companion quantum-channel volume  [ LTC26 , “Transfer matrix of a matrix endomorphism” ] , the matrix-units special case of the one-family transfer matrix of the corresponding definition in the companion quantum-channel volume  [ LTC26 , “Transfer matrix in one family” ] .

The corresponding generic result is proved in the companion quantum-channel volume  [ LTC26 , “Functoriality of transfer matrices” ] .

The corresponding generic result is proved in the companion quantum-channel volume  [ LTC26 , “Trace preservation fixes the vectorized identity on the left” ] .

The corresponding generic result is proved in the companion quantum-channel volume  [ LTC26 , “Powers and traces of transfer matrices” ] .

Theorem 19.6.1 Eigenvalues of a transfer matrix
#

A complex number \(\mu \) is an eigenvalue of \(T\) if and only if it is an eigenvalue of \(\widehat T\).

Proof

Column stacking is a linear bijection and intertwines \(T\) with multiplication by \(\widehat T\). It therefore carries nonzero eigenmatrices to nonzero eigenvectors and conversely.

For a matrix product tensor \(A\), write \(V^{(N)}(A)_\sigma \) for the coefficient of its length-\(N\) periodic vector at the configuration \(\sigma \).

Let \(A\) and \(B\) be matrix product tensors with the same positive bond dimension, and let \(F_{A,B}(X)=\sum _i A^iX(B^i)^\dagger \) be their mixed transfer map. Then

\begin{align} \operatorname{tr}\! \left(\widehat{F_{A,B}}^{\, N}\right) & =\sum _\sigma V^{(N)}(A)_\sigma \overline{V^{(N)}(B)_\sigma }. \end{align}

In particular, taking \(B=A\) identifies the trace of the \(N\)th power of the ordinary transfer matrix with the length-\(N\) self-overlap.

Proof

Replace the matrix power by the transfer matrix of the corresponding operator power, identify matrix trace with operator trace, and expand the latter over periodic words. For the ordinary self-transfer statement, specialize to \(B=A\) and use the identity \(F_{A,A}=\mathcal E_A\) from Theorem B.1.1.

Definition 19.6.3 Normalized flattening of an MPO tensor

For an MPO tensor \(U\) of physical dimension \(d{\gt}0\), its normalized flattening is the doubled-physical-index MPS tensor

\begin{align} \widetilde U^{ij} & =\frac{1}{\sqrt d}\, U^{ij}. \end{align}

This is the normalization used in the normal-tensor argument of [ CPGSV17 , Section III, equation (13) ] .

Lemma 19.6.4 Normalized flattening of a composition tensor

Let \(\mathcal U\) and \(\mathcal V\) be MPO tensors with the same positive physical dimension \(d\), and let

\begin{align} W^{ik} & =\sum _{j=0}^{d-1}U^{ij}\otimes V^{jk} \end{align}

be their composition tensor. Under the canonical identification of its auxiliary space with \(\mathbb {C}^{D_U}\otimes \mathbb {C}^{D_V}\), the normalized flattenings satisfy

\begin{align} \widetilde W^{ik} & =\sqrt d\sum _{j=0}^{d-1} \widetilde U^{ij}\otimes \widetilde V^{jk}. \end{align}

This identity combines the raw composition tensor used in the proof of [ CPGSV17 , Theorem IV.6(ii) ] with the normalization in [ CPGSV17 , Section III, equation (13) ] ; it is not a separately stated result of that proof.

Proof

Substitute \(U^{ij}=\sqrt d\, \widetilde U^{ij}\) and \(V^{jk}=\sqrt d\, \widetilde V^{jk}\) into the defining contraction for \(W^{ik}\), and then multiply the result by \(d^{-1/2}\).

Theorem 19.6.5 Normalized flattening under blocking

Let \(\mathcal U\) be an MPO tensor of physical dimension \(d\), and let \(L\geq 0\). For the normalized-tensor interpretation assume \(d{\gt}0\), as in Definition 19.6.3; with the convention \(0^{-1}=0\), the algebraic identity below also holds for \(d=0\). Pairing the ket and bra letters at each site gives a canonical bijection

\begin{align} \vartheta _L:\operatorname {Fin}(d^L)^2 & \longrightarrow \operatorname {Fin}(d^2)^L. \end{align}

If \(\widetilde{\mathcal U}=d^{-1/2}\mathcal U\) denotes normalized flattening, then

\begin{align} \widetilde{\mathcal U^{[L]}} & =\vartheta _L^*\bigl(\widetilde{\mathcal U}^{[L]}\bigr). \label{eq:mpu_normalized_flattening_blocking} \end{align}

Thus normalized flattening commutes with blocking, up to the canonical doubled-index relabeling. This is the normalized form of the blocking operation in [ CPGSV17 , Definition II.5 and equation (9) ] .

Proof

A blocked letter is a product of \(L\) original letters. The identity

\begin{align} \bigl(\sqrt{d^L}\bigr)^{-1} & =\bigl((\sqrt d)^{-1}\bigr)^L \end{align}

identifies its normalization with the product of the \(L\) one-site normalizations. Theorem 20.14.2 identifies the two physical indexings for every \(L\geq 0\); at \(L=0\) both blocked words are empty and both local matrix products are the identity.

Suppose that an MPS tensor \(A\) is left-canonical:

\begin{align} \sum _i (A^i)^\dagger A^i & =\mathbb {1}. \end{align}

Then for every \(L\geq 0\) its blocked tensor is left-canonical:

\begin{align} \sum _I \bigl((A^{[L]})^I\bigr)^\dagger (A^{[L]})^I & =\mathbb {1}. \label{eq:mpu_left_canonical_blocking} \end{align}

If \(e\) is a bijection of physical alphabets, then \(e^*A\) is left-canonical if and only if \(A\) is:

\begin{align} \sum _j \bigl((e^*A)^j\bigr)^\dagger (e^*A)^j=\mathbb {1}\quad \Longleftrightarrow \quad \sum _i(A^i)^\dagger A^i=\mathbb {1}. \label{eq:mpu_left_canonical_reindexing} \end{align}
Proof

Expand (279) as a sum over length-\(L\) words and sum one letter at a time, repeatedly applying the one-site left-canonical identity. For \(L=0\), the unique empty word evaluates to the identity. The equivalence (280) follows by reindexing the finite sum along \(e\).

Let \(A\) have canonical-form-II data with retained weights \(\mu _k\), blocks \(A_k\), and positive definite diagonal fixed matrices \(\Lambda _k\). For every positive integer \(L\), the blocked tensor \(A^{[L]}\) has canonical-form-II data with weights \(\mu _k^L\), blocks \(A_k^{[L]}\), and the same matrices \(\Lambda _k\). The retained blocks remain left-canonical:

\begin{align} \sum _I \bigl((A_k^{[L]})^I\bigr)^\dagger (A_k^{[L]})^I & =\mathbb {1}. \end{align}

A bijective relabeling of the physical alphabet likewise preserves canonical-form-II data, including both the matrices \(\Lambda _k\) and the left-canonical identities.

Consequently, if the normalized flattening \(\widetilde{\mathcal U}\) of an MPO tensor has canonical-form-II data, then for every \(L{\gt}0\) the normalized flattening of \(\mathcal U^{[L]}\) has the data obtained by first blocking and then relabeling by \(\vartheta _L\) from Theorem 19.6.5. Theorem 19.6.6 preserves the left-canonical identity under both operations. For each retained block, Lemma D.4.8 gives

\begin{align} \mathcal{E}_{A_k^{[L]}} & =\mathcal{E}_{A_k}^{L}, & \mathcal{E}_{A_k}^{L}(\Lambda _k) & =\Lambda _k. \end{align}

The second identity follows by iterating \(\mathcal{E}_{A_k}(\Lambda _k)=\Lambda _k\). Thus the same positive definite diagonal matrix is fixed after blocking. A physical bijection leaves the transfer map unchanged. The final construction applies these two operations to (276).

For a matrix product tensor \(A\), a complex number \(c\), and a nonnegative integer \(N\), scaling every local matrix by \(c\) gives

\begin{align} \sum _\sigma V^{(N)}(cA)_\sigma \overline{V^{(N)}(cA)_\sigma } & =(c\overline c)^N \sum _\sigma V^{(N)}(A)_\sigma \overline{V^{(N)}(A)_\sigma }. \end{align}

Consequently, if \(d{\gt}0\) and \(U\) is an MPO tensor of physical dimension \(d\), then

\begin{align} \sum _\sigma V^{(N)}(\widetilde U)_\sigma \overline{V^{(N)}(\widetilde U)_\sigma } & =d^{-N} \sum _\sigma V^{(N)}(U^{\mathrm{MPS}})_\sigma \overline{V^{(N)}(U^{\mathrm{MPS}})_\sigma }. \end{align}
Proof

Each length-\(N\) coefficient is multiplied by \(c^N\). Multiplication by its complex conjugate gives the first identity. Taking \(c=1/\sqrt d\) gives the second.

Let \(U\) be an MPU tensor with positive physical and bond dimensions, and let \(E\) be the transfer matrix of its normalized flattening. For every integer \(N{\gt}1\),

\begin{align} \operatorname{tr}(E^N) & =1. \end{align}

This is the trace identity in the proof of the normal-tensor result [ CPGSV17 , Proposition III.1 ] .

Proof

Scaling both tensor copies by \(1/\sqrt d\) multiplies the length-\(N\) self-overlap by \(d^{-N}\). The doubled-index overlap identity and MPU unitarity give the complete chain

\begin{align} \operatorname{tr}(E^N) & =\sum _\sigma V^{(N)}(\widetilde U)_\sigma \overline{V^{(N)}(\widetilde U)_\sigma } \\ & =d^{-N}\sum _\sigma V^{(N)}(U^{\mathrm{MPS}})_\sigma \overline{V^{(N)}(U^{\mathrm{MPS}})_\sigma } \\ & =d^{-N}\operatorname{tr}\! \left(U^{(N)}(U^{(N)})^\dagger \right) \\ & =d^{-N}\operatorname{tr}(\mathbb {1}) \\ & =d^{-N}d^N \\ & =1. \end{align}

Here the fourth equality uses \(N{\gt}1\), so that the MPU equation applies.

19.7 Block transfer multiplicity

Definition 19.7.1 Flattened dependent-block inclusion
#

Let the blocks be indexed by \(k\in \{ 0,\ldots ,r{-}1\} \), let \(n_k\) be their nonnegative dimensions, and put \(n=\sum _{k\in \{ 0,\ldots ,r{-}1\} }n_k\). The standard block-order bijection is

\begin{align} \phi \colon \bigsqcup _{k\in \{ 0,\ldots ,r{-}1\} }\{ 0,\ldots ,n_k{-}1\} & \longrightarrow \{ 0,\ldots ,n{-}1\} , \\ \phi (k,a) & =\sum _{j=0}^{k-1}n_j+a. \end{align}

For each \(k\in \{ 0,\ldots ,r{-}1\} \), define \(J_k:\mathbb C^{n_k}\to \mathbb C^n\) by

\begin{align} (J_k)_{x,a} & =\begin{cases} 1, & x=\phi (k,a), \\ 0, & x\ne \phi (k,a). \end{cases}\end{align}

The inclusions satisfy the following identities, with the second holding for distinct \(k\) and \(\ell \):

\begin{align} J_k^\dagger J_k & =\mathbb {1}, \\ J_k^\dagger J_\ell & =0. \end{align}

If \(B^i=\bigoplus _k\mu _k A_k^i\) is a weighted block-diagonal tensor, then

\begin{align} B^iJ_k & =J_k(\mu _kA_k^i). \end{align}
Proof

Reindex the canonical dependent-summand inclusion along the equivalence between dependent pairs and flattened coordinates. Transporting its isometry, cross-block orthogonality, and block-diagonal intertwining identities gives the three displayed formulas.

Definition 19.7.3 Ambient inclusion of a canonical block

If \(U\) is the reconstruction coisometry and \(J_k\) is the inclusion of the \(k\)th retained block, define

\begin{align} V_k & =U^\dagger J_k. \end{align}

For every retained block and physical index \(i\),

\begin{align} V_k^\dagger V_k & =\mathbb {1}, \\ A^iV_k & =V_k(\mu _k A_k^i). \end{align}
Proof

The coisometry \(UU^\dagger =\mathbb {1}\) and \(J_k^\dagger J_k=\mathbb {1}\) give

\begin{align} V_k^\dagger V_k & =J_k^\dagger UU^\dagger J_k =J_k^\dagger J_k =\mathbb {1}. \end{align}

The canonical reconstruction and block-diagonal intertwining give

\begin{align} A^iV_k & =U^\dagger \Bigl(\bigoplus _j\mu _jA_j^i\Bigr)J_k =U^\dagger J_k(\mu _kA_k^i) =V_k(\mu _kA_k^i). \end{align}
Definition 19.7.5 Weighted-block transfer eigenvalue

For a block \(k\), define its weighted transfer eigenvalue by

\begin{align} q_k & =\mu _k\overline{\mu _k}=\lvert \mu _k\rvert ^2. \end{align}

If the ambient transfer matrix satisfies \(\operatorname{tr}(E^N)=1\) for every integer \(N{\gt}1\), then every block satisfies

\begin{align} q_k & =1. \end{align}
Proof

Choose a nonzero fixed matrix \(X_k\) for the unweighted normal block and set \(Y_k=V_kX_kV_k^\dagger \). The isometry and intertwining identities give

\begin{align} Y_k & \ne 0, & E(Y_k) & =q_kY_k. \end{align}

Thus \(q_k\) lies in the ambient transfer spectrum; it is nonzero because \(\mu _k\ne 0\). The shifted trace-power spectrum is \(\{ 0,1\} \), hence \(q_k=1\).

If there is exactly one block \(k\) and the blocks fill the ambient bond space, then

\begin{align} D_k & =D, & V_kV_k^\dagger & =\mathbb {1}. \end{align}
Proof

The full-support sum reduces to the unique summand:

\begin{align} D & =\sum _{j=1}^{r}D_j=D_k. \end{align}

Therefore \(V_k\) is a square matrix. Since \(V_k^\dagger V_k=\mathbb {1}\), finite equal-dimensional inverse symmetry gives \(V_kV_k^\dagger =\mathbb {1}\).

Let \(A\) have literal CPSV canonical-form data with \(r\) blocks. If its transfer matrix \(E\) satisfies \(\operatorname{tr}(E^N)=1\) for every integer \(N{\gt}1\), then

\begin{align} r & =1. \end{align}

This is the block multiplicity consequence used in [ CPGSV17 , Proposition III.1 ] .

Proof

Each normal block \(A_k\) has a nonzero transfer fixed vector \(X_k\). Let \(U\) be the reconstruction coisometry and let \(J_k\) be the flattened inclusion of the \(k\)th block. Define the ambient inclusion and transported matrix by

\begin{align} V_k & =U^\dagger J_k, & Y_k & =V_kX_kV_k^\dagger . \end{align}

The matrix \(Y_k\) is nonzero and is an ambient transfer eigenmatrix with eigenvalue \(q_k=\mu _k\overline{\mu _k}\ne 0\). Theorem 19.4.6 gives \(q_k=1\).

For distinct \(k\) and \(\ell \), the inclusion identities give

\begin{align} V_k^\dagger Y_kV_k & =X_k, \\ V_k^\dagger Y_\ell V_k & =0. \end{align}

Hence the matrices \(Y_k\), for \(1\le k\le r\), are linearly independent. Vectorizing gives

\begin{align} \operatorname {vec}(Y_k) & \in \ker (E^2-\mathbb {1}). \end{align}

For every \(r\ge 1\), the shifted traces and Theorem 19.4.4 give

\begin{align} \operatorname{tr}((E^2)^r) & =1, \\ \chi _{E^2}(X) & =X^{D^2-1}(X-1). \end{align}

The eigenspace dimension is bounded by the algebraic multiplicity, so

\begin{align} r & \le \dim \ker (E^2-\mathbb {1}) \le \operatorname{mult}_{1}(\chi _{E^2}) =1. \end{align}

Finally, if \(r\) were zero, the retained direct sum and therefore \(A\) would vanish, contradicting \(\operatorname{tr}(E^2)=1\). Consequently \(r=1\).

19.8 Normalized transfer spectrum and conditional normality

Theorem 19.8.1 Shifted transfer traces imply radius one and primitivity

Let \(T:M_D(\mathbb C)\to M_D(\mathbb C)\) be a linear map with \(D{\gt}0\). If \(\operatorname{tr}(\widehat T^{\, N})=1\) for every integer \(N{\gt}1\), then

\begin{align} r(T) & =1, & \operatorname {peripheral}(T) & =\{ 1\} . \end{align}

This is the radius-and-primitivity part of [ CPGSV17 , Proposition III.1 ] .

Proof

The shifted traces give

\begin{align} \sigma (\widehat T)\setminus \{ 0\} & =\{ 1\} . \end{align}

Column stacking identifies the eigenvalues of \(T\) and \(\widehat T\). Consequently every spectral value of \(T\) is zero or one, and one occurs, so

\begin{align} r(T) & =\max _{\lambda \in \sigma (T)}|\lambda |=1. \end{align}

Restricting to eigenvalues of modulus one gives

\begin{align} \operatorname {peripheral}(T) & =\{ 1\} , \end{align}

which is primitivity.

Let \(U\) be an MPU tensor with positive physical and bond dimensions, and let \(E\) be the transfer matrix of its normalized flattening. Then

\begin{align} \sigma (E)\setminus \{ 0\} & =\{ 1\} , & \operatorname {mult}_{1}(\chi _E) & =1, & \chi _E(X) & =X^{D^2-1}(X-1). \end{align}

This is the transfer-matrix conclusion of [ CPGSV17 , Proposition III.1 ] . The first set equality alone does not assert algebraic multiplicity.

Proof

For every integer \(N{\gt}1\), the normalized transfer identity gives

\begin{align} \operatorname{tr}(E^N) & =1. \end{align}

Applying the exact shifted trace-power theorem in dimension \(D^2\) yields

\begin{align} \chi _E(X) & =X^{D^2-1}(X-1), & \operatorname {mult}_{1}(\chi _E) & =1, & \sigma (E)\setminus \{ 0\} & =\{ 1\} . \end{align}

Let \(U\) be an MPU tensor with physical dimension \(d{\gt}0\) and bond dimension \(D{\gt}1\), and let \(E\) be the transfer matrix of its normalized flattening. There is an integer \(J\) satisfying

\begin{align} 0{\lt}J\leq D^2-1 \end{align}

and vectors \(\rho ,\Phi \) satisfying

\begin{align} (\Phi \mid \rho ) & =1, & E\rho & =\rho , & (\Phi \mid E & =(\Phi \mid , \end{align}

such that

\begin{align} E^J & =\lvert \rho )(\Phi \rvert . \end{align}

Moreover, for every integer \(k\geq J\),

\begin{align} E^k & =E^J. \end{align}

Thus blocking \(J\) sites removes the entire generalized zero-eigenspace. The one-dimensional case is stated separately below.

This is the Jordan-elimination step in [ CPGSV17 , proof of Proposition III.2(ii) ] .

Proof

The exact characteristic polynomial is

\begin{align} \chi _E(X) & =X^{D^2-1}(X-1). \end{align}

Cayley–Hamilton gives \(E^{D^2}=E^{D^2-1}\). Hence every power from \(J=D^2-1\) onward is equal, and \(P=E^J\) is idempotent. Since \(J{\gt}1\), the normalized transfer trace identity gives \(\operatorname{tr}(P)=1\). The range of an idempotent has dimension equal to its trace, so \(P\) has rank one and factors as \(P=\lvert \rho )(\Phi \rvert \). Taking the factors with \((\Phi \mid \rho )=1\) gives the displayed fixed-point identities from \(EP=PE=P\).

Theorem 19.8.4 Uniqueness of the normalized fixed-pair factorization

Let \(P=E^J\) be a stabilized rank-one transfer power with normalized left and right factors. If \(\rho '\) and \(\Phi '\) are any other fixed vectors with \((\Phi '\mid \rho ')=1\), then

\begin{align} P=\lvert \rho ’)(\Phi ’\rvert . \end{align}
Proof

Applying \(P\) to \(\rho '\) and applying \(\Phi '\) to \(P\) shows that the two pairs of rank-one factors differ by reciprocal scalars. Their normalized pairings force the product of those scalars to be one.

Suppose \(d{\gt}0\), \(D{\gt}1\), and \(\mathcal U\) is an MPU whose normalized flattening is presented by chosen canonical-form-II data with blocks filling the bond space. There is a unique block \(k\). Normalize its diagonal positive definite fixed matrix \(\Lambda _k\) to have trace one, and let \(V_k\) be its inclusion into the ambient bond space. Then

\begin{align} \rho & =V_k\Lambda _kV_k^\dagger , & \rho & {\gt}0, & \operatorname{tr}(\rho ) & =1, \\ E\lvert \rho ) & =\lvert \rho ), & (\Phi \rvert E & =(\Phi \rvert , & (\Phi \mid \rho ) & =1, \end{align}

where \((\Phi \rvert \) is the vectorized identity. Consequently,

\begin{align} E^{D^2-1}=\lvert \rho )(\Phi \rvert . \end{align}

The diagonal matrix is the block matrix \(\Lambda _k\) in the canonical-form-II fixed-point equations in [ CPGSV17 , equations (6a)–(6b) ] ; its ambient transport \(\rho \) need not be diagonal. This is a statement about the chosen reduced representative and the blocking step in [ CPGSV17 , proof of Proposition III.2(ii) ] ; it does not assert that every ambient MPU tensor is itself in canonical form II.

Proof

The shifted trace identities satisfy \(\operatorname{tr}(E^N)=1\) for every integer \(N{\gt}1\) and give

\begin{align} \operatorname{tr}(E^N)=1 \quad \Longrightarrow \quad r=1. \end{align}

Let \(k\) be this block. Canonical form II supplies a diagonal \(\Lambda _0{\gt}0\) fixed by the unweighted block transfer map. Normalize it by

\begin{align} \Lambda & =\operatorname{tr}(\Lambda _0)^{-1}\Lambda _0, & \operatorname{tr}(\Lambda ) & =1. \end{align}

If \(\omega _k\) is its weight, then

\begin{align} \lvert \omega _k\rvert ^2 & =1, & E_k(\Lambda ) & =\Lambda , \end{align}

where \(E_k\) is the transfer map of the weighted block \(\widetilde B_k=\omega _k B_k\). Full support makes the block inclusion \(V=V_k\) both isometric and onto:

\begin{align} V^\dagger V & =I, & VV^\dagger & =I. \end{align}

Hence \(\rho =V\Lambda V^\dagger \) is positive definite with trace one, and the intertwining equations \(A_iV=V\widetilde B_k(i)\) give the fixedness transport

\begin{align} E(V\Lambda V^\dagger ) & =V E_k(\Lambda )V^\dagger =V\Lambda V^\dagger =\rho . \end{align}

The blockwise left-canonical identity transports similarly:

\begin{align} \sum _i \widetilde B_k(i)^\dagger \widetilde B_k(i)=I \quad \Longrightarrow \quad \sum _i A_i^\dagger A_i=I \quad \Longrightarrow \quad (\Phi \rvert E=(\Phi \rvert . \end{align}

Finally,

\begin{align} (\Phi \mid \rho ) & =\operatorname{tr}(\rho )=1, & E^{D^2-1} & =\lvert \rho )(\Phi \rvert , \end{align}

where the last equality is the uniqueness of the normalized fixed-pair factorization for the stabilized power.

Let \(d{\gt}0\), \(D{\gt}1\), and let \(\mathcal U\) be an MPU whose normalized flattening has chosen canonical-form-II data with full support. Put \(J=D^2-1\) and \(W=\mathcal U_{JD^2}\). There is a matrix \(\rho {\gt}0\) with \(\operatorname{tr}(\rho )=1\) such that

\begin{align} E\lvert \rho ) & =\lvert \rho ), & (\Phi \rvert E & =(\Phi \rvert , & E^J & =\lvert \rho )(\Phi \rvert . \end{align}

With the corresponding vectorized pair \(\rho ',\Phi '\), the same blocked tensor \(W\) satisfies the two aligned simple contractions

\begin{align} \Phi ’\mathbin {\boldsymbol \cdot }W^{ij}\rho ’ & =\delta _{ij}, \\ W^{ij}W^{k\ell } & =W^{ij}\lvert \rho ’)(\Phi ’\rvert W^{k\ell }. \end{align}

This theorem supplies only the fixed pair and the two simple contractions. It does not assert a source-cut factorization or source-\(v\) isometry.

Proof

Theorem 19.8.5 gives the positive trace-one fixed matrix and the identity \(E^J=\lvert \rho )(\Phi \rvert \). Applying Theorem 19.3.23 with this pair gives both contractions on the direct block \(W=\mathcal U_{JD^2}\).

Theorem 19.8.7 One-dimensional normalized transfer matrix

If the physical dimension satisfies \(d{\gt}0\) and the bond dimension is \(D=1\), then the normalized transfer matrix satisfies

\begin{align} E & =1. \end{align}

Consequently its powers stabilize at the positive exponent one.

Proof

In dimension one the characteristic polynomial identity reduces to \(\chi _E(X)=X-1\), and Cayley–Hamilton gives \(E=1\).

Let \(U\) be an MPU tensor with positive physical and bond dimensions. Suppose that the normalized flattening is equipped with CPSV canonical-form data whose blocks have full support in the ambient bond space. Then the normalized flattening is a normal tensor.

This is the full-support representative form of [ CPGSV17 , Proposition III.1 ] . The full-support hypothesis is essential for this ambient representative: adjoining a zero virtual direct summand preserves every periodic operator but introduces a nontrivial invariant projection. Thus the theorem does not assert that an arbitrary, possibly nonminimal MPU representative is normal.

Proof

Write \(\widetilde U\) for the normalized flattening, set \(T=\mathcal E_{\widetilde U}\) for its transfer map, and let \(E=\widehat T\) be the corresponding transfer matrix. Also write \(r\) for the number of blocks and \(D_k\) for the bond dimension of the block \(k\). For every integer \(N{\gt}1\), the normalized transfer identity gives

\begin{align} \operatorname{tr}(E^N) & =1. \end{align}

The block-multiplicity theorem and the transfer-spectrum theorem then give

\begin{align} r & =1, & r(T) & =1, & \operatorname {peripheral}(T) & =\{ 1\} . \end{align}

Full support supplies

\begin{align} \sum _{k=1}^{r}D_k & =D. \end{align}

Since \(r=1\), the unique block has ambient dimension \(D\) and therefore makes \(\widetilde U\) irreducible. The displayed radius and peripheral-spectrum identities provide its remaining normality fields. Consequently \(\widetilde U\) is a normal tensor.

19.9 Standard form, index, and symmetry classification

Definition 19.9.1 Reduced full-support source datum

Let \(\widetilde{\mathcal U}=d^{-1/2}\mathcal U\) be the normalized flattening, and let \(E\) be its transfer matrix. A reduced full-support source datum for \(\mathcal U\), with physical dimension \(d{\gt}0\) and positive bond dimension, consists of an integer \(J{\gt}0\), a positive definite matrix \(\rho \) with \(\operatorname{tr}(\rho )=1\), and \((\Phi \rvert =\operatorname {vec}(\mathbb {1}_D)^{\mathsf T}\) such that

\begin{align} E^J & =\lvert \rho )(\Phi \rvert . \end{align}

These data are supplied in one of two ways.

  1. If \(D=1\), take \(J=1\) and \(\rho =\mathbb {1}_1\) using Theorem 19.8.7.

  2. If \(D{\gt}1\), choose canonical-form-II data for the normalized flattening of this same tensor, with its blocks filling the ambient bond space. Theorem 19.8.5 supplies the fixed pair with \(J=D^2-1\).

For \(D{\gt}1\), this states the fixed-pair convention in [ CPGSV17 , equations labelled Erightleft and II_UTransfer, and proof of Proposition III.2(ii) ] . The \(D=1\) branch is the direct one-dimensional specialization. This definition does not assert that an arbitrary representative with \(D{\gt}1\) admits the required canonical-form-II data and does not replace \(\mathcal U\) by another tensor.

Let \(\mathcal U\) be an MPU tensor with physical dimension \(d{\gt}0\) and bond dimension \(D{\gt}0\). Let \(\rho {\gt}0\) satisfy \(\operatorname{tr}(\rho )=1\), put \((\Phi \rvert =\operatorname {vec}(\mathbb {1}_D)^{\mathsf T}\), and suppose that for some \(J{\gt}0\) the normalized transfer matrix satisfies

\begin{align} E^J & =\lvert \rho )(\Phi \rvert . \end{align}

Construct the source factors and \(u\) from \(\mathcal U\) using \(\rho \), and set \(K=JD^2\). For retained physical pairs \(p,q\in \{ 0,\ldots ,d-1\} ^2\), the complete source-\(u\) contraction satisfies

\begin{align} \sum _{l,r}u_{(l,r),q}\overline{u_{(l,r),p}} & =d^{-K}\sum _{\tau ,\eta } U^{(K+2)}_{(q,\tau ),\eta } \overline{U^{(K+2)}_{(p,\tau ),\eta }}. \end{align}

Here \(\tau \) ranges over input words of length \(K\), and \(\eta \) ranges over output words of length \(K+2\). This is the supplied-fixed-pair form of the range-insertion equality in [ CPGSV17 , Lemma III.7 ] .

Proof

Put \(W=\mathcal U^{[K]}\). The reflected transfer-power identity, the direct matching contractions, and supplied-witness reblocking give the aligned direct and reflected contractions at lengths \(K\) and \(K+2\), with the same fixed pair at both lengths. The block \(W\) is used only inside this contraction argument; the external source tensors are the source-\(u\) factors of \(\mathcal U\).

Expand the ordinary Gram entry while retaining both external source-\(u\) tensors and the full \(\rho \)-weighted boundary. Insert the first- and second-cut range projections, but contract them only inside the closed \((K+2)\)-site network. The reflected contractions move the second-cut metric through the overlapping windows. After every virtual and source index is closed, the terminal boundary contraction gives the normalized output-first trace displayed above. No separate identity for the second-cut metric or either ambient range projection is asserted.

Let \(\mathcal U\) be an MPU tensor equipped with a reduced full-support source datum. Let \(J\), \(\rho \), and \((\Phi \rvert \) be its stabilization exponent and normalized fixed pair, so that

\begin{align} E^J & =\lvert \rho )(\Phi \rvert . \end{align}

Compute both source cuts, their ranks \(r\) and \(\ell \), the source operator \(v\), and the right source factors \(Y_1,Y_2\) from this same tensor \(\mathcal U\) using \(\rho \). Let \(W\) be the double layer of \(\mathcal U\). Then the supplied contraction

\begin{align} W^{ij}W^{k\ell } & =W^{ij}\lvert \rho )(\Phi \rvert W^{k\ell } \end{align}

holds if and only if

\begin{align} (Y_1\otimes Y_2)^\dagger v^\dagger v(Y_1\otimes Y_2) & =(Y_1\otimes Y_2)^\dagger (Y_1\otimes Y_2). \end{align}

When \(\mathcal U\) is simple, Theorem 19.3.28 supplies the displayed contraction with this same fixed pair. This is the complete equivalence in [ CPGSV17 , proof of Theorem III.8 ] .

Proof

Insert \(E^J=\lvert \rho )(\Phi \rvert \) into the complete network while retaining the source factors of \(\mathcal U\). Expand both source cuts and rotate the second cut. Contract the resulting four-letter expression as one closed network to obtain the dressed Gram equation, and reverse the same calculation for the converse.

Lemma 19.9.4 The source operator \(u\) is an isometry

For any MPU tensor \(\mathcal U\) under the source’s standing canonical-form-II convention, let \(r\) and \(\ell \) be the two source-cut ranks and let \(u\) be the source operator. Then

\begin{align} u^\dagger u & =\mathbb {1}. \end{align}

Consequently \(u\) is an isometry and \(r\ell \geq d^2\). This is the unrestricted statement of [ CPGSV17 , Lemma III.7 ] .

Let \(\mathcal U\) be an MPU tensor with physical dimension \(d{\gt}0\) and bond dimension \(D{\gt}0\). Let \(\rho {\gt}0\) satisfy \(\operatorname{tr}(\rho )=1\), put \((\Phi \rvert =\operatorname {vec}(\mathbb {1}_D)^{\mathsf T}\), and suppose that for some \(J{\gt}0\) the normalized transfer matrix satisfies

\begin{align} E^J & =\lvert \rho )(\Phi \rvert . \end{align}

Let \(r\) and \(\ell \) be the two source-cut ranks, and construct the source factors and \(u\) from \(\mathcal U\) using \(\rho \). Then

\begin{align} u^\dagger u & =\mathbb {1}. \end{align}

Consequently \(u\) is an isometry and \(r\ell \geq d^2\). This is the reduced-source version of [ CPGSV17 , Lemma III.7 ] .

Proof

Set \(K=JD^2\). The complete-network identity and output-first MPU unitarity give, for all retained physical pairs \(p,q\),

\begin{align} (u^\dagger u)_{p,q} & =d^{-K}\sum _{\tau ,\eta } U^{(K+2)}_{(q,\tau ),\eta } \overline{U^{(K+2)}_{(p,\tau ),\eta }} \\ & =\delta _{p,q}. \end{align}

Hence \(u^\dagger u=\mathbb {1}\).

Theorem 19.9.6 Simple-tensor equivalence theorem

Let \(\mathcal U\) be an MPU tensor under the source’s standing canonical-form-II convention. Let \(r\) and \(\ell \) be its source-cut ranks, and let \(u\) and \(v\) be its source operators. Then the following conditions are equivalent:

  1. \(\mathcal U\) is simple;

  2. \(r\ell =d^2\);

  3. \(u\) is unitary;

  4. \(v\) is unitary.

This is the unrestricted statement of [ CPGSV17 , Theorem III.8 ] .

Let \(\mathcal U\) be an MPU tensor equipped with a reduced full-support source datum, and let \(\rho \) and \((\Phi \rvert \) be its normalized fixed pair. Compute the source-cut ranks \(r\) and \(\ell \) and the source operators \(u\) and \(v\) from this same tensor \(\mathcal U\) using \(\rho \). Then the following conditions are equivalent:

  1. \(\mathcal U\) is simple;

  2. \(r\ell =d^2\);

  3. \(u\) is unitary;

  4. \(v\) is unitary.

This is the reduced-source version of [ CPGSV17 , Theorem III.8 ] .

Proof

Lemma 19.9.5 gives \(u^\dagger u=\mathbb {1}\) before any of the four conditions is assumed. Thus \(u\) is a standing isometry and \(r\ell \geq d^2\) throughout the proof.

Suppose first that \(\mathcal U\) is simple. Theorem 19.3.28, applied to the stabilized transfer power, gives the second contraction with the supplied pair \(\lvert \rho )(\Phi \rvert \). By Theorem 19.9.3,

\begin{align} (Y_1\otimes Y_2)^\dagger v^\dagger v(Y_1\otimes Y_2) & =(Y_1\otimes Y_2)^\dagger (Y_1\otimes Y_2). \end{align}

Cancelling the dressing by the right inverses of \(Y_1\) and \(Y_2\) gives \(v^\dagger v=\mathbb {1}\), hence \(r\ell \leq d^2\). Together with the standing source-\(u\) isometry, this gives \(r\ell =d^2\).

If \(r\ell =d^2\), the standing isometry \(u\) is square and therefore unitary. If \(u\) is unitary, the exact two-site factorization of \(U^{(2)}\) and unitarity of \(U^{(2)}\) imply that \(v\) is unitary. Finally, suppose that \(v\) is unitary. Unitarity gives the dressed Gram identity, and Theorem 19.9.3 gives the supplied second contraction for \(\mathcal U\). The same-witness implication in Theorem 19.4.70 supplies the first contraction, so \(\mathcal U\) is simple.

Let \(\mathcal U\) be a simple tensor under the source’s standing canonical-form-II convention. Its standard form is the depth-two form of the two-site block \(\mathcal U_2\) determined by the source unitaries \(u\) and \(v\). This is the unrestricted definition of [ CPGSV17 , Definition III.9 ] .

Definition 19.9.9 Admissible standard form

Let \(\mathcal U\) be a simple tensor equipped with a reduced full-support source datum. The admissible standard form of the two-site blocking \(\mathcal U_2\) is the depth-two form determined by the source unitaries

\begin{align} u & \colon \mathbb C^d\otimes \mathbb C^d \longrightarrow \mathbb C^\ell \otimes \mathbb C^r, \\ v & \colon \mathbb C^r\otimes \mathbb C^\ell \longrightarrow \mathbb C^d\otimes \mathbb C^d. \end{align}

On four consecutive blocked legs its unitary is

\begin{align} U & =v_{23}v_{41}u_{12}u_{34}. \end{align}
Theorem 19.9.10 Fundamental Theorem of MPU

Given two simple tensors \(\mathcal U\) and \(\mathcal V\) with standard forms, they generate the same MPU for every \(N\) if and only if there are unitaries \(x\), \(y\), and \(z\) relating the two standard forms by the local gauge equations of [ CPGSV17 , Eqns. (SFuu) and (SFvv) ] . This is the unrestricted statement of the source’s Fundamental Theorem of MPU.

Theorem 19.9.11 Admissible Fundamental Theorem of MPU

Let \(\mathcal U\) and \(\mathcal V\) be simple tensors equipped with reduced full-support source data and written in admissible standard form, with source unitaries \((u_{\mathcal U},v_{\mathcal U})\) and \((u_{\mathcal V},v_{\mathcal V})\). They generate the same MPU for every \(N\),

\begin{align} U^{(N)} & =V^{(N)}, \end{align}

if and only if their admissible standard forms are related by the local unitary gauges of [ CPGSV17 , Eqns. (SFuu) and (SFvv) ] : there are unitaries \(x\), \(y\), and \(z\) on the three corresponding source and canonical-form-II legs. In particular, the source unitaries obey the local relation

\begin{align} u_{\mathcal V} & =(x\otimes y)u_{\mathcal U}, \end{align}

while the two canonical-form-II tensors are related on their virtual leg by \(z\) and on their source legs by \(x\) and \(y\).

Theorem 19.9.12 Rank of a Kronecker product
#

Let \(A\) and \(B\) be finite matrices over a field. Then

\begin{align} \operatorname {rank}(A\otimes B) & =\operatorname {rank}(A)\operatorname {rank}(B). \end{align}
Proof

Identify the image of the Kronecker product with the tensor product of the two images and multiply their dimensions.

Definition 19.9.13 Specified-tensor source-index value

For a specified tensor \(\mathcal U\) whose right and left source-cut ranks \(r\) and \(\ell \) are positive, define

\begin{align} \operatorname {ind}_{\mathrm{src}}(\mathcal U) & =\frac12(\log _2 r-\log _2\ell ). \end{align}

This definition makes no choice of a simple blocking.

Let \(\mathcal U\) and \(\mathcal V\) be specified tensors. Suppose that their source ranks obey

\begin{align} r_{\mathcal V} & =c r_{\mathcal U}, & \ell _{\mathcal V} & =c\ell _{\mathcal U}, \end{align}

where \(c\), \(r_{\mathcal U}\), and \(\ell _{\mathcal U}\) are positive. Then

\begin{align} \operatorname {ind}_{\mathrm{src}}(\mathcal V) & =\operatorname {ind}_{\mathrm{src}}(\mathcal U). \end{align}

Let \(\mathcal U\), \(\mathcal V\), and \(\mathcal W\) instead have positive source ranks and suppose that

\begin{align} r_{\mathcal W} & =c r_{\mathcal U}r_{\mathcal V}, \\ \ell _{\mathcal W} & =c\ell _{\mathcal U}\ell _{\mathcal V}, \end{align}

where \(c\) is a positive integer. Then

\begin{align} \operatorname {ind}_{\mathrm{src}}(\mathcal W) & =\operatorname {ind}_{\mathrm{src}}(\mathcal U) +\operatorname {ind}_{\mathrm{src}}(\mathcal V). \end{align}

If \(d\), \(r_{\mathcal U}\), and \(\ell _{\mathcal U}\) are positive and a specified tensor has \(r_{\mathcal U}\ell _{\mathcal U}=d^2\), then

\begin{align} \operatorname {ind}_{\mathrm{src}}(\mathcal U) & =\log _2\frac{r_{\mathcal U}}d =-\log _2\frac{\ell _{\mathcal U}}d. \end{align}
Proof

Expand the definition. For common scaling, apply \(\log _2(cx)=\log _2c+\log _2x\) to both ranks and cancel the two copies of \(\log _2c\). For the three-tensor identity, apply the product formula twice to each rank and cancel the common term. For the equivalent formulas, apply the same product identity to \(r_{\mathcal U}\ell _{\mathcal U}=d^2\) and use \(\log _2(x/y)=\log _2x-\log _2y\).

Theorem 19.9.15 Conditional blocking invariance of the specified source-index value

Let \(d{\gt}0\) and \(k_0\leq k\). Suppose that the endpoint source ranks satisfy

\begin{align} r_{k_0}\ell _{k_0} & =d^{2k_0}, \\ r_k\ell _k & =d^{2k}. \end{align}

Then the specified source-index values agree:

\begin{align} \operatorname {ind}_{\mathrm{src}}(\mathcal U_k) & =\operatorname {ind}_{\mathrm{src}}(\mathcal U_{k_0}). \end{align}

This statement assumes the endpoint products; it neither selects a simple blocking nor defines the public index.

Proof

Exact blocking growth gives the common positive factor \(d^{k-k_0}\):

\begin{align} r_k & =d^{k-k_0}r_{k_0}, \\ \ell _k & =d^{k-k_0}\ell _{k_0}. \end{align}

Its logarithm occurs once in each rank term and therefore cancels from their difference.

Definition 19.9.16 Index

Let \(k\) be any blocking length for which \(\mathcal U_k\) is simple, and let \(r\) and \(\ell \) be its right and left source-cut ranks. The source index is

\begin{align} \operatorname {ind}(\mathcal U) & =\frac12(\log _2 r-\log _2\ell ). \end{align}

This is the unrestricted definition of [ CPGSV17 , Definition IV.1 ] .

Let \(k\) be a blocking length for which \(\mathcal U_k\) is simple and is equipped with a reduced full-support source datum, and let \(r\) and \(\ell \) be its right and left source-cut ranks. Define

\begin{align} \operatorname {ind}(\mathcal U) & =\frac12(\log _2 r-\log _2\ell ). \end{align}

Since \(r\ell =d^{2k}\), this is equivalently

\begin{align} \operatorname {ind}(\mathcal U) & =\log _2\frac{r}{d^k} =-\log _2\frac{\ell }{d^k}. \end{align}

The source index is independent of the chosen simple blocking. More precisely, if \(k\geq k_0\) and both \(\mathcal U_{k_0}\) and \(\mathcal U_k\) are simple, then

\begin{align} r_k & =d^{k-k_0}r_{k_0}, & \ell _k & =d^{k-k_0}\ell _{k_0}. \end{align}

This is the unrestricted statement of [ CPGSV17 , Proposition IV.2 ] .

Proposition 19.9.19 Well-definedness of the admissible index

The admissible index is independent of the chosen admissible simple blocking. More precisely, if \(k\geq k_0\), both \(\mathcal U_{k_0}\) and \(\mathcal U_k\) are simple, and each is equipped with a reduced full-support source datum, then

\begin{align} r_k & =d^{k-k_0}r_{k_0}, & \ell _k & =d^{k-k_0}\ell _{k_0}. \end{align}
Lemma 19.9.20 Tensor-product additivity

For MPU tensors under the source’s standing canonical-form convention,

\begin{align} \operatorname {ind}(\mathcal U\otimes \mathcal V) & =\operatorname {ind}(\mathcal U)+\operatorname {ind}(\mathcal V). \end{align}

This is the unrestricted tensor-product statement in [ CPGSV17 , Theorem IV.6(ii) ] .

Lemma 19.9.21 Admissible tensor-product additivity

Suppose there is one positive blocking length \(k\) such that \(\mathcal U_k\), \(\mathcal V_k\), and \((\mathcal U\otimes \mathcal V)_k\) are simple and each carries its own reduced full-support source datum. Define all three indices from these blocks. Then

\begin{align} \operatorname {ind}(\mathcal U\otimes \mathcal V) & =\operatorname {ind}(\mathcal U)+\operatorname {ind}(\mathcal V). \end{align}

This is the common-block reduced-source version of [ CPGSV17 , Theorem IV.6(ii) ] .

Let \(\mathcal U\) and \(\mathcal V\) be arbitrary matrix product operator tensors of common physical dimension \(d\). For their product tensor, four-site blocking satisfies

\begin{align} r[(\mathcal U\mathbin {\cdot }\mathcal V)_4] & \leq d^2 r[\mathcal U]r[\mathcal V], \\ \ell [(\mathcal U\mathbin {\cdot }\mathcal V)_4] & \leq d^2 \ell [\mathcal U]\ell [\mathcal V]. \end{align}

No unitarity or simplicity assumption is required.

Proof

Use the canonical physical reindexing to identify iterated two-site blocking with direct four-site blocking. For the right cut, factor the two-site block of each tensor through its source tensor \(u\). The diagonal cut through the two blocked factors retains one physical leg and one right source leg from each factor. Thus, for suitable rectangular matrices \(A_{1,4}\) and \(B_{1,4}\),

\begin{align} M_1((\mathcal U\mathbin {\cdot }\mathcal V)_4) & =A_{1,4}B_{1,4}, \end{align}

where the columns of \(A_{1,4}\) and the rows of \(B_{1,4}\) are indexed by

\begin{align} \mathcal K_1 & =(\mathbb C^d\otimes \mathbb C^{r[\mathcal U]}) \otimes (\mathbb C^d\otimes \mathbb C^{r[\mathcal V]}). \end{align}

Consequently,

\begin{align} r[(\mathcal U\mathbin {\cdot }\mathcal V)_4] & =\operatorname {rank}(A_{1,4}B_{1,4}) \leq \dim (\mathcal K_1) =d^2r[\mathcal U]r[\mathcal V]. \end{align}

For the left cut, use the reflected \(v\) factorization, with its physical order \((j_2,j_1)\) and source order \((r,\ell )\). For suitable rectangular matrices \(A_{2,4}\) and \(B_{2,4}\),

\begin{align} M_2((\mathcal U\mathbin {\cdot }\mathcal V)_4) & =A_{2,4}B_{2,4}, \end{align}

where the intermediate space is

\begin{align} \mathcal K_2 & =(\mathbb C^d\otimes \mathbb C^{\ell [\mathcal U]}) \otimes (\mathbb C^d\otimes \mathbb C^{\ell [\mathcal V]}). \end{align}

Hence

\begin{align} \ell [(\mathcal U\mathbin {\cdot }\mathcal V)_4] & =\operatorname {rank}(A_{2,4}B_{2,4}) \leq \dim (\mathcal K_2) =d^2\ell [\mathcal U]\ell [\mathcal V]. \end{align}

Both source-cut ranks are unchanged by the canonical physical relabeling.

Let \(\mathcal W\) generate the composition of the MPUs generated by \(\mathcal U\) and \(\mathcal U'\). Under the source’s standing canonical-form convention,

\begin{align} \operatorname {ind}(\mathcal W) & =\operatorname {ind}(\mathcal U)+\operatorname {ind}(\mathcal U’). \end{align}

This is the unrestricted composition statement in [ CPGSV17 , Theorem IV.6(ii) ] .

Suppose there is a positive integer \(k\) such that \(\mathcal U_k\), \(\mathcal U'_k\), and the composition tensor \(\mathcal W_k\) are simple and each carries its own reduced full-support source datum. Suppose in addition that \(\mathcal U_{2k}\), \(\mathcal U'_{2k}\), and \(\mathcal W_{4k}\) are simple and each carries its own reduced full-support source datum. Define the three indices initially from \(\mathcal U_k\), \(\mathcal U'_k\), and \(\mathcal W_k\); the supplied data at \(2k\) and \(4k\) make the later comparison admissible. If \(\mathcal W\) generates the composition of the MPUs generated by \(\mathcal U\) and \(\mathcal U'\), then

\begin{align} \operatorname {ind}(\mathcal W) & =\operatorname {ind}(\mathcal U)+\operatorname {ind}(\mathcal U’). \end{align}

This is the six-block reduced-source version of [ CPGSV17 , Theorem IV.6(ii) ] .

Definition 19.9.25 MPU canonical form

A tensor \(A\) of physical dimension \(d\) and bond dimension \(D\) is in MPU canonical form if there are positive block dimensions \(D_k\), weights \(\mu _k\in \mathbb C\), tensors \(A_k\) of bond dimension \(D_k\), and a coisometry

\begin{align} C\colon \mathbb C^D & \longrightarrow \bigoplus _{k=1}^r\mathbb C^{D_k}, & CC^\dagger & =\mathbb {1}, \end{align}

such that

\begin{align} A^i=C^\dagger \left(\bigoplus _{k=1}^r\mu _k A_k^i\right)C. \end{align}

Every \(A_k\) is irreducible and the spectral radius of its transfer map is one. The blocks may be periodic, so no uniqueness condition is imposed on their peripheral eigenvalues. This is the canonical form of [ CPGSV17 , lines 259–267 ] ; normal tensors form a strictly stronger class.

Definition 19.9.26 Strict equivalence

Fix a virtual bond dimension \(D\). Let \(\mathcal U\) and \(\mathcal V\) be canonical-form MPU tensors with physical dimensions \(d_a\) and \(d_b\). Given an equality \(h\colon d_a=d_b\), reindex \(\mathcal U\) from \(d_a\) to \(d_b\) along \(h^{-1}\). The tensors are strictly equivalent with respect to \(h\) if there is a continuous path \(\mathcal W(p)\), \(p\in [0,1]\), in the fixed-bond MPU locus such that

\begin{align} \mathcal W(0) & =\operatorname {reind}_{h^{-1}}(\mathcal U), \\ \mathcal W(1) & =\mathcal V. \end{align}

Only the canonical-form requirement is relaxed along the path. This is the fixed-bond interpretation of [ CPGSV17 , lines 708–714 ] ; it does not compare unequal raw bond dimensions.

Let \(x{\gt}0\). After adjoining an identity ancilla of dimension \(x\), the normalized flattening is the physical reindexing of the original flattening with a normalized diagonal ancilla alphabet. Its transfer map is exactly the original transfer map. Consequently, chosen canonical-form-II data with full support determine canonical-form-II data with full support for \(\mathcal U^{(x)}=\mathcal U\otimes \mathbb {1}_x\).

The product-index equivalence has the orientation

\begin{align} ((i,a),(j,b)) & \longmapsto ((i,j),(a,b)). \end{align}

This result preserves the normalized transfer data. It does not identify the old and new source cuts, ranks, or chosen compact-SVD factors.

Proof

Only the \(x\) diagonal ancilla letters are nonzero, each with coefficient \(x^{-1/2}\). Hence their total transfer contribution is

\begin{align} x\left(x^{-1/2}\right)^2 & =1. \end{align}

Thus the transfer map and each diagonal positive fixed matrix are unchanged. The block dimensions, weights, and ambient coisometry are unchanged as well, so full support follows from the same block-dimension sum. The displayed physical reindexing gives the normalized flattening of the enlarged MPO.

Definition 19.9.28 Equivalence after blocking and ancillas

Fix a virtual bond dimension \(D\). Let \(\mathcal U\) and \(\mathcal V\) be MPU tensors with physical dimensions \(d_a\) and \(d_b\). They are equivalent if there are a positive blocking length \(k\) and coprime positive integers \(p_a,p_b\) with

\begin{align} p_a d_a & =p_b d_b \end{align}

such that \((\mathcal U^{(p_a)})_k\) and \((\mathcal V^{(p_b)})_k\) are strictly equivalent. Here the positive-dimensional identity ancillas are attached before the common blocking, and

\begin{align} \mathcal U^{(p)} & =\mathcal U\otimes \mathbb {1}_p. \end{align}

This is the fixed-bond interpretation of [ CPGSV17 , lines 716–724 ] ; it does not introduce a stabilization relation for unequal raw bond dimensions.

Definition 19.9.29 Admissible equivalence and path datum

An admissible strict-equivalence datum for a path \([0,1]\ni p\mapsto \mathcal W(p)\) consists of a common positive blocking length \(m\) such that the actual blocked path tensor \(\mathcal W(p)_m\) is simple and carries a reduced full-support source datum for every \(p\). The blocked tensors and their source data vary continuously with \(p\), after fixed identifications of the finite-dimensional spaces involved. No equivalent or reduced representative may replace \(\mathcal W(p)_m\). In particular, the actual blocked endpoint tensors carry their own source data.

When the path is used under a specified symmetry \(\mathcal S\), the datum also includes continuously varying reduced full-support source data for the fixed family of comparison paths

\begin{align} \overline{\mathcal W(p)_m},\qquad \mathcal W(p)_m^\sharp ,\qquad \mathcal W(p)_m^{\mathsf T},\qquad \mathcal S[\mathcal W(p)_m]. \end{align}

Every tensor in this family is simple. This requirement is part of the datum independently of which comparisons a later proof uses; no preservation under conjugation, physical adjoint, transposition, or \(\mathcal S\) is assumed.

Two tensors are admissibly strictly equivalent if their strict equivalence is witnessed by such a path. They are admissibly equivalent if the blocked, ancilla-enlarged endpoints in Definition 19.9.28 are joined by an admissible strict-equivalence path. The source arguments motivating these additional path hypotheses are Proposition IV.5, Theorem IV.6, Corollary IV.7, and the lemma following the symmetry-equivalence definitions in [ CPGSV17 ] ; the datum itself is introduced here and has no source-labelled counterpart.

Let \(P\) and \(Q\) be orthogonal projections on a complex inner-product space. The compression of \(Q\) to \(\operatorname {ran}P\) is self-adjoint. If a unit vector \(x\in \operatorname {ran}P\) satisfies

\begin{align} PQx & =\mu x, & 0 & {\lt}\mu {\lt}1, \end{align}

then there is a unit vector \(w\perp x\) such that

\begin{align} Px & =x, & Pw & =0, \\ Qx & =\mu x+\sqrt{\mu (1-\mu )}\, w, & Qw & =\sqrt{\mu (1-\mu )}\, x+(1-\mu )w. \end{align}

Thus \(\operatorname {span}\{ x,w\} \) is the generic two-dimensional angle block. This lemma concerns one eigenvalue in \((0,1)\); it does not include the endpoint eigenspaces or an orthogonal direct-sum decomposition of the ambient space.

Proof

Set \(d=Qx-\mu x\). Symmetry and idempotence give

\begin{align} Pd & =0, & \langle x,d\rangle & =0, & \lVert d\rVert ^2 & =\mu (1-\mu ), \\ Qd & =\mu (1-\mu )x+(1-\mu )d. \end{align}

Since \(0{\lt}\mu {\lt}1\), the vector \(w=d/\sqrt{\mu (1-\mu )}\) is well defined and has the stated properties.

For orthogonal projections \(P\) and \(Q\) on a finite-dimensional complex inner-product space, let \((x_i)_i\) be an orthonormal eigenbasis of the compression of \(Q\) to \(\operatorname {ran}P\), with ordered real eigenvalues \((\mu _i)_i\).

Let \(P\) and \(Q\) be orthogonal projections on a finite-dimensional complex inner-product space, with compression spectral data \((x_i,\mu _i)\). Then, for every \(i\), \(\mu _i\in [0,1]\) and exactly one of the following alternatives holds:

  1. \(\mu _i=0\) and \(Qx_i=0\);

  2. \(\mu _i=1\) and \(Qx_i=x_i\);

  3. \(0{\lt}\mu _i{\lt}1\), and \(x_i\) belongs to the two-dimensional angle block described in Theorem 19.9.30.

This theorem classifies the compression eigenvectors.

Proof

Apply the finite-dimensional spectral theorem to the self-adjoint compression of \(Q\) to \(\operatorname {ran}P\). For a unit compression eigenvector \(x\) with eigenvalue \(\mu \), the angle-defect identity gives

\begin{align} \lVert Qx-\mu x\rVert ^2 & =\mu (1-\mu ), \end{align}

hence \(0\leq \mu \leq 1\). At \(\mu =0\) or \(1\) the defect vanishes, giving the two common one-dimensional blocks. For \(0{\lt}\mu {\lt}1\), apply Theorem 19.9.30.

Definition 19.9.33 Compression defect block

For compression spectral data \((x_i,\mu _i)\), define

\begin{align} d_i & =Qx_i-\mu _i x_i, \\ B_i & =\operatorname {span}\{ x_i,d_i\} . \end{align}

For all indices \(i,j\),

\begin{align} \langle x_i,d_j\rangle & =0, \\ \langle d_i,x_j\rangle & =0, \\ \langle d_i,d_j\rangle & =\mu _j(1-\mu _j)\langle x_i,x_j\rangle . \end{align}

In particular, \(\langle d_i,d_j\rangle =0\) when \(i\ne j\), including when \(\mu _i=\mu _j\). The block \(B_i\) has dimension at most two, is invariant under both \(P\) and \(Q\), and satisfies \(B_i\perp B_j\) for distinct indices.

Proof

Since \(Px_i=x_i\), symmetry of \(P\) and the compression eigenvalue equation give

\begin{align} \langle x_i,d_j\rangle & =\langle x_i,Qx_j\rangle -\mu _j\langle x_i,x_j\rangle \\ & =\langle Px_i,Qx_j\rangle -\mu _j\langle x_i,x_j\rangle \\ & =\langle x_i,PQx_j\rangle -\mu _j\langle x_i,x_j\rangle =0. \end{align}

Conjugate symmetry gives \(\langle d_i,x_j\rangle =0\). For the exact defect pairing, symmetry of \(Q\) and the local action formula give

\begin{align} Qd_j & =\mu _j(1-\mu _j)x_j+(1-\mu _j)d_j, \\ \langle d_i,d_j\rangle & =\langle Qx_i,d_j\rangle -\mu _i\langle x_i,d_j\rangle , \\ \langle Qx_i,d_j\rangle -\mu _i\langle x_i,d_j\rangle & =\langle x_i,Qd_j\rangle , \\ \langle x_i,Qd_j\rangle & =\mu _j(1-\mu _j)\langle x_i,x_j\rangle +(1-\mu _j)\langle x_i,d_j\rangle , \\ \langle d_i,d_j\rangle & =\mu _j(1-\mu _j)\langle x_i,x_j\rangle . \end{align}

Orthogonality of the generated spans follows by linearity. The action formulas

\begin{align} Px_i & =x_i, \\ Pd_i & =0, \\ Qx_i & =\mu _i x_i+d_i, \\ Qd_i & =\mu _i(1-\mu _i)x_i+(1-\mu _i)d_i \end{align}

show that both projections preserve \(B_i=\operatorname {span}\{ x_i,d_i\} \). Its dimension bound follows from its two generators.

Definition 19.9.35 Defect-block sum and complementary summand

For the compression defect blocks \((B_i)_i\), define

\begin{align} B & =\bigvee _i B_i, & K & =B^\perp . \end{align}

The index \(i\) ranges over every vector in the chosen compression eigenbasis, including repeated eigenvalues and the endpoint eigenvalues \(0\) and \(1\).

The blocks \((B_i)_i\) form an orthogonal internal direct sum in \(B\), and

\begin{align} E & =B\mathbin {\perp \! \oplus }K. \end{align}

Both \(B\) and \(K\) are invariant under \(P\) and \(Q\). Moreover, \(\operatorname {ran}P\subseteq B\), so for every \(z\in K\),

\begin{align} Pz & =0, & P(Qz) & =0, & Q(Pz) & =0. \end{align}

Consequently the restrictions \(P|_K\) and \(Q|_K\) are orthogonal projections and commute. No commutation of \(P\) and \(Q\) on the ambient space is assumed.

Proof

Pairwise orthogonality of the blocks implies independence, while their supremum is \(B\) by definition. Hence they form an internal direct sum in \(B\). The standard orthogonal-complement decomposition gives \(E=B\mathbin {\perp \! \oplus }B^\perp \).

The local action formulas show that each \(B_i\) is invariant under both projections, hence so is \(B\). Symmetry of each projection then makes \(K=B^\perp \) invariant. The compression eigenvectors form a basis of \(\operatorname {ran}P\), and each lies in its block \(B_i\); therefore \(\operatorname {ran}P\subseteq B\). Thus \(P\) vanishes on \(K\). Since \(Qz\in K\) whenever \(z\in K\), both \(P(Qz)\) and \(Q(Pz)\) vanish. The two restricted projections therefore commute.

Definition 19.9.37 Reduced projection

Let \(P\) and \(Q\) be orthogonal projections on a finite-dimensional complex inner-product space. Define \(\widetilde P\) as the orthogonal projection onto \(\operatorname {ran}(PQ)\).

The reduced projection satisfies

\begin{align} \operatorname {ran}(\widetilde P) & =\operatorname {ran}(PQ), \\ \operatorname {ran}(PQP) & =\operatorname {ran}(PQ), \\ P\widetilde P & =\widetilde P, \\ \widetilde P Q & =PQ, \\ Q\widetilde P & =QP, \\ \operatorname {ran}(\widetilde P Q) & =\operatorname {ran}(\widetilde P). \end{align}

The first two identities describe the canonical construction, while the next three establish the properties in source lines 767–769. The final range equality is the derived input to the right-factor theorem, not item (iv) of the source.

Proof

Set \(A=QP\). The standard range identity for \(A^\dagger A\) gives

\begin{align} \operatorname {ran}(PQP) & =\operatorname {ran}(A^\dagger A) =\operatorname {ran}(A^\dagger ) =\operatorname {ran}(PQ). \end{align}

By definition, \(\operatorname {ran}(\widetilde P)=\operatorname {ran}(PQ)\), which also gives \(P\widetilde P=\widetilde P\). For all \(x\) and \(z\),

\begin{align} \left\langle Qx-PQx,\, PQz\right\rangle & =0. \end{align}

Hence \(Qx-PQx\in \operatorname {ran}(PQ)^\perp \), so orthogonal projection onto this range yields \(\widetilde P Q=PQ\). Taking adjoints gives \(Q\widetilde P=QP\). Finally,

\begin{align} \operatorname {ran}(\widetilde P Q) & =\operatorname {ran}(PQ) =\operatorname {ran}(\widetilde P). \end{align}

For orthogonal projections \(P\) and \(Q\), the reduced range contains no nonzero vector annihilated by \(Q\):

\begin{align} \ker Q\cap \operatorname {ran}(\widetilde P) & =\{ 0\} . \label{eq:mpu_reduced_projection_transversality} \end{align}
Proof

If \(x=PQy\) and \(Qx=0\), then \(Px=x\) and symmetry of the two projections gives

\begin{align} \langle x,x\rangle & =\langle PQy,PQy\rangle =\langle Qy,PQy\rangle =\langle y,QPQy\rangle =0. \end{align}

Hence \(x=0\).

Theorem 19.9.40 Invertible right factor for the reduced projection

There is an invertible linear map \(Y\) such that

\begin{align} \widetilde P QY=\widetilde P. \end{align}
Proof

The equality

\begin{align} \operatorname {ran}(\widetilde P Q) & =\operatorname {ran}(\widetilde P) \end{align}

and the equal-ranges right-factor theorem give an invertible \(Y\) satisfying \(\widetilde P QY=\widetilde P\).

Definition 19.9.41 Matrix reduced projection
#

In standard Euclidean coordinates, let \(\widetilde P\) be the matrix of the reduced projection associated with two square matrices \(P\) and \(Q\).

If \(P\) and \(Q\) are orthogonal projections, then \(\widetilde P\) is an orthogonal projection and

\begin{align} P\widetilde P & =\widetilde P, & \widetilde P P & =\widetilde P, & \widetilde P Q & =PQ, & Q\widetilde P & =QP. \label{eq:mpu_reduced_projection_coordinate_identities} \end{align}

There is also an invertible matrix \(Y\) satisfying \(\widetilde P QY=\widetilde P\).

Lemma 19.9.43 Transfer map under virtual sandwiching

Let \(U=(U^{ij})\) be an MPO tensor, and let \(\widehat U^{ij}=AU^{ij}B\). Then

\begin{align} E_{\widehat U}(X) & =A E_U(BXB^\dagger )A^\dagger . \end{align}
Proof

Expand the transfer map and collect the factors \(A\), \(B\), \(B^\dagger \), and \(A^\dagger \) outside the finite sum.

Lemma 19.9.44 Positive definite support compression

Let \(\rho \succeq 0\). If

\begin{align} \operatorname {supp}(\rho )(Vx) & \neq 0 \qquad \text{for every }x\neq 0, \end{align}

then

\begin{align} V^\dagger \rho V & \succ 0. \end{align}
Proof

If the quadratic form of \(V^\dagger \rho V\) vanished at a nonzero vector \(x\), positive semidefiniteness would give \(\rho Vx=0\). The support projection would then annihilate \(Vx\), contrary to the hypothesis.

Let \(W=(W^{ij})\) be an MPO tensor. Let \(L,R\succeq 0\) satisfy

\begin{align} E_W(X) & =\operatorname {tr}(LX)R, & \operatorname {tr}(LR) & =1, \end{align}

and let \(P\) and \(Q\) be their support projections. Assume, letter by letter,

\begin{align} W^{ij}P & =W^{ij}, & QW^{ij} & =W^{ij}. \end{align}

Set \(T=\widetilde P\) and \(\widetilde W^{ij}=TW^{ij}T\). Then, for every \(N\geq 0\),

\begin{align} \operatorname {MPO}_N(\widetilde W) & =\operatorname {MPO}_N(W), \\ E_{\widetilde W}(X) & =\operatorname {tr}\! \bigl((TLT)X\bigr)\, TRT, \\ \operatorname {tr}\! \bigl((TLT)(TRT)\bigr) & =1. \end{align}

Moreover, if \(V^\dagger V=I\) and \(VV^\dagger =T\), then

\begin{align} V^\dagger LV & \succ 0, & V^\dagger RV & \succ 0. \end{align}
Proof

For every nonempty word \(w\), idempotence of \(T\) and (443) give

\begin{align} \widetilde W_w & =TW_wT. \end{align}

Cyclicity of the trace, together with \(QW_w=W_w=W_wP\), therefore gives \(\operatorname {tr}(TW_wT)=\operatorname {tr}(W_w)\). Length zero is immediate. Lemma 19.9.43, with \(A=B=T\) and \(T^\dagger =T\), gives

\begin{align} E_{\widetilde W}(X) & =\operatorname {tr}(LTXT)\, TRT =\operatorname {tr}((TLT)X)\, TRT. \end{align}

Since \(P\) and \(Q\) are the support projections of \(L\) and \(R\), respectively,

\begin{align} PL & =LP=L, & QR & =RQ=R. \end{align}

Combining these absorptions with (443), \(T^2=T\), and cyclicity of the trace gives

\begin{align} \operatorname {tr}\! \bigl((TLT)(TRT)\bigr) & =\operatorname {tr}(LTRT) =\operatorname {tr}(LRT) =\operatorname {tr}(RTL) =\operatorname {tr}(RL) =\operatorname {tr}(LR) =1. \end{align}

Let \(x\neq 0\) and set \(u=Vx\). Since \(V^\dagger V=I\), one has \(u\neq 0\); since \(VV^\dagger =T\), one has \(Tu=u\). Equation (443) gives \(Pu=u\), and (439) gives \(Qu\neq 0\). Thus

\begin{align} \operatorname {supp}(L)u & =Pu=u\neq 0, & \operatorname {supp}(R)u & =Qu\neq 0. \end{align}

Applying Lemma 19.9.44 to \(L\) and \(R\) yields

\begin{align} x^\dagger (V^\dagger LV)x & =u^\dagger Lu{\gt}0, & x^\dagger (V^\dagger RV)x & =u^\dagger Ru{\gt}0. \end{align}
Theorem 19.9.46 Invertible factor for a reduced projection

Let \(L,R\succeq 0\), let \(P\) and \(Q\) be their support projections, and let \(\widetilde P\) be the reduced projection of \(P\) and \(Q\). There are invertible support-inverse extensions of \(L\) and \(R\), together with an invertible matrix \(Y\), such that \(PQY=\widetilde P\).

Proof

Extend the inverses on the supports by the identity on their orthogonal complements, and choose an invertible right factor for the reduced projection.

Let \(W=(W^{ij})\) be an MPO tensor, let \(L\), \(R\), \(X\), \(Z\), \(Y\) be square matrices of the same size, let \(P\) be an orthogonal projection, let \(Q\) be a square matrix of the same size, and set \(\widetilde P\) equal to their reduced projection. Assume, letter by letter, \(W^{ij}P=W^{ij}=QW^{ij}\). Set \(\widehat W^{ij}=LW^{ij}R\) and \(\widetilde W^{ij}=\widetilde P W^{ij}\widetilde P\). If \(X,Z,Y\) satisfy \(XL=P\), \(RZ=Q\), and \(PQY=\widetilde P\), then

\begin{align} X\widehat W ZY & =PWQY \label{eq:mpu_reduced_hat_chain_pwqy} \\ PWQY & =PQWPQY \label{eq:mpu_reduced_hat_chain_pqwpqy} \\ PQWPQY & =\widetilde P W\widetilde P \label{eq:mpu_reduced_hat_chain_reduced} \\ \widetilde P W\widetilde P & =\widetilde W. \label{eq:mpu_reduced_hat_chain_tilde} \end{align}

If in addition \(X\), \(Z\), and \(Y\) are invertible, then

\begin{align} r[\widehat W] & =r[\widetilde W],\label{eq:mpu_reduced_hat_right_rank} \\ \ell [\widehat W] & =\ell [\widetilde W]. \label{eq:mpu_reduced_hat_left_rank} \end{align}

If \(L,R\succeq 0\) and \(P=\operatorname {supp}(L)\), \(Q=\operatorname {supp}(R)\), such invertible matrices may be chosen as \(X=L^++(\mathbb {1}-P)\), \(Z=R^++(\mathbb {1}-Q)\), together with an invertible right factor \(Y\) of the reduced projection.

Proof

Expanding each virtual sandwich and using \(XL=P\) and \(RZ=Q\) gives (459). The absorptions \(QW^{ij}=W^{ij}=W^{ij}P\) give, letter by letter,

\begin{align} PW^{ij}QY & =P(QW^{ij})QY=PQ(W^{ij}P)QY=PQW^{ij}PQY, \notag \end{align}

which is (460). Since \(\widetilde P Q=PQ\) and \(PQY=\widetilde P\),

\begin{align} PQW^{ij}PQY & =(\widetilde P Q)W^{ij}\widetilde P =\widetilde P W^{ij}\widetilde P. \end{align}

This proves (461) and (462). Thus \(\widetilde W\) is obtained from \(\widehat W\) by left multiplication by \(X\) and right multiplication by \(ZY\). Invertibility of these factors and the source ranks of a virtual sandwich (Theorem 19.9.56) give

\begin{align} r[\widehat W] & =r[X\widehat W(ZY)]=r[\widetilde W], \notag \\ \ell [\widehat W] & =\ell [X\widehat W(ZY)]=\ell [\widetilde W], \notag \end{align}

the first equality in each line being the invariance identity with invertible \(X\) and \(ZY\), and the second rewriting \(X\widehat W(ZY)\) through (459)–(461) and (462) to \(\widetilde W\), which gives (463) and (464). For positive semidefinite \(L\) and \(R\), their ambient support-inverse extensions supply \(X\) and \(Z\); the invertible right-factor identity for \(\widetilde P Q\) supplies \(Y\).

Theorem 19.9.48 Lower semicontinuity of finite matrix rank

For finite index sets \(M,N\), matrix rank defines a lower-semicontinuous map

\begin{align} \operatorname {rank}\colon \operatorname {Mat}_{M\times N}(\mathbb C) & \longrightarrow \mathbb N. \end{align}

Consequently, if \(A\colon X\to \operatorname {Mat}_{M\times N}(\mathbb C)\) is continuous, then \(x\mapsto \operatorname {rank}A(x)\) is lower semicontinuous.

Proof

For every \(k\in \mathbb N\), the sublevel set is the closed determinantal set

\begin{align} \{ A:\operatorname {rank}A\leq k\} . \end{align}

These closed sublevel sets characterize lower semicontinuity. Composition with a continuous map preserves lower semicontinuity.

Theorem 19.9.49 Fixed positive product gives local constancy

Let \(r,\ell \colon X\to \mathbb N\) be lower semicontinuous. Suppose that \(c{\gt}0\) and

\begin{align} r(x)\ell (x) & =c \end{align}

for every \(x\in X\). Then both \(r\) and \(\ell \) are locally constant.

Proof

Fix \(x\in X\). Since \(c{\gt}0\), both \(r(x)\) and \(\ell (x)\) are positive. Lower semicontinuity gives a neighborhood \(U\) of \(x\) on which

\begin{align} r(y) & \geq r(x), & \ell (y) & \geq \ell (x). \end{align}

For \(y\in U\), monotonicity of multiplication and the fixed-product identity give

\begin{align} r(x)\ell (x) & \leq r(y)\ell (x) \leq r(y)\ell (y) =r(x)\ell (x). \end{align}

Positivity permits cancellation, first of \(\ell (x)\) and then of \(r(x)\), so \(r(y)=r(x)\) and \(\ell (y)=\ell (x)\).

Theorem 19.9.50 Fixed positive product on a preconnected set

Under the hypotheses of Theorem 19.9.49, if \(S\subseteq X\) is preconnected, then for all \(x,y\in S\),

\begin{align} r(x) & =r(y), & \ell (x) & =\ell (y). \end{align}

In particular, both functions are constant when \(X\) is preconnected and when \(X=[0,1]\).

Proof

A locally constant function has open fibers. A preconnected set cannot meet two disjoint nonempty open fibers, so each of \(r\) and \(\ell \) has one value on \(S\). The unit interval is connected, hence preconnected.

Definition 19.9.51 Virtual sandwich

For an MPO tensor \(\mathcal U=(U^{ij})_{0\leq i,j{\lt}d}\) of bond dimension \(D\) and matrices \(A,B\in \mathbb C^{D\times D}\), define

\begin{align} \widehat{\mathcal U}^{ij} & =A U^{ij}B. \end{align}
Theorem 19.9.52 Continuity of the virtual sandwich

If \(x\mapsto A(x)\), \(x\mapsto \mathcal U(x)\), and \(x\mapsto B(x)\) are continuous families, then

\begin{align} x\longmapsto A(x)\mathcal U(x)B(x) \end{align}

is continuous.

Proof

Every matrix entry is a finite sum of products of entries of \(A(x)\), \(\mathcal U(x)\), and \(B(x)\), hence is continuous.

Definition 19.9.53 Doubled-virtual contraction and factor-free sandwich

Let \(q:\{ 0,\ldots ,D{-}1\} \times \{ 0,\ldots ,D{-}1\} \to \{ 0,\ldots ,D^2{-}1\} \) be the standard enumeration of a pair. For an MPO tensor \(\mathcal W\) and a doubled-virtual matrix \(E\in \mathbb C^{D^2\times D^2}\), define

\begin{align} C(\mathcal W,E)^{ij}_{ab} & =\sum _{c,e=0}^{D-1}E_{q(b,e),q(c,a)}W^{ij}_{ce}. \label{eq:mpu_factor_free_contraction} \end{align}

If \(E(\mathcal W)\) denotes the normalized physical diagonal of the physical-adjoint double layer of \(\mathcal W\), define

\begin{align} \widehat{\mathcal W}_{\mathrm{ff}} & =C\bigl(\mathcal W,E(\mathcal W)\bigr). \label{eq:mpu_factor_free_sandwich} \end{align}

This is the contraction shown in Figure IV_index4.png of [ CPGSV17 , Proposition IV.5 ] .

Let \(L,R\in \mathbb C^{D\times D}\) and let \(\operatorname {vec}\) stack matrix columns, so

\begin{align} \operatorname {vec}(L)_{q(c,a)} & =L_{ac}, \\ \operatorname {vec}(R)_{q(b,e)} & =R_{eb}. \end{align}

Then the rank-one doubled-virtual matrix satisfies

\begin{align} C\! \left(\mathcal W, \operatorname {vec}(R)\operatorname {vec}(L)^T\right)^{ij} & =L W^{ij}R. \label{eq:mpu_factor_free_rank_one} \end{align}

No adjoint or normalization factor occurs in the contraction result. The binary map \((\mathcal W,E)\mapsto C(\mathcal W,E)\), the map \(\mathcal W\mapsto E(\mathcal W)\), and consequently the map \(\mathcal W\mapsto \widehat{\mathcal W}_{\mathrm{ff}}\) are continuous. These assertions neither choose \(L\) and \(R\) continuously nor infer their positivity. They make no claim about reduced representatives or constancy of ranks.

Proof

Expanding the left-hand side of (478) and using column stacking gives

\begin{align} C\! \left(\mathcal W, \operatorname {vec}(R)\operatorname {vec}(L)^T\right)^{ij}_{ab} & =\sum _{c,e}R_{eb}L_{ac}W^{ij}_{ce} \\ & =\sum _{c,e}L_{ac}W^{ij}_{ce}R_{eb} =(L W^{ij}R)_{ab}. \end{align}

This proves (478). Each entry in (474) is a finite sum of products, so the binary contraction is continuous. The double-layer entries are finite sums of products of entries of \(\mathcal W\) and their complex conjugates, and the normalized physical diagonal is another finite sum. Hence \(\mathcal W\mapsto E(\mathcal W)\) is continuous. Substitution into (475) proves continuity of the factor-free sandwich.

The two raw source cuts satisfy the literal formulas

\begin{align} M_1(\widehat{\mathcal U}) & =(A\otimes I_d)M_1(\mathcal U)(I_d\otimes B), \\ M_2(\widehat{\mathcal U}) & =(A\otimes I_d)M_2(\mathcal U)(I_d\otimes B). \end{align}

No transpose or conjugation occurs in either identity.

Proof

For \(a=1,2\), expand an entry of the right-hand side. The identity factors collapse the two physical sums, leaving

\begin{align} \sum _{\gamma ,\delta } A_{\alpha \gamma }U^{ij}_{\gamma \delta }B_{\delta \beta } & =(A U^{ij}B)_{\alpha \beta }, \end{align}

which is the corresponding entry of \(M_a(\widehat{\mathcal U})\).

If \(A\) and \(B\) are invertible, then

\begin{align} r[\widehat{\mathcal U}] & =r[\mathcal U], & \ell [\widehat{\mathcal U}] & =\ell [\mathcal U]. \end{align}

This assertion concerns the raw source ranks only.

Proof

The determinant identities for Kronecker products show that \(A\otimes I_d\) and \(I_d\otimes B\) are invertible. Left and right multiplication by these matrices preserves rank. Apply this observation to each source-cut formula in Theorem 19.9.55.

Definition 19.9.57 Ambient support-inverse extension

Let \(A\succeq 0\) have support projection \(P\), and let \(A^+\) denote its inverse on the support. Define

\begin{align} X_A & =A^++(\mathbb {1}-P). \end{align}

Define also

\begin{align} Z_A & =A+(\mathbb {1}-P). \end{align}

This construction gives explicit witnesses for the invertible matrices required in Proposition IV.5 of [ CPGSV17 ] ; the paper requires those matrices but does not state these formulas.

The matrices \(X_A\) and \(Z_A\) are mutually inverse, and

\begin{align} X_AA=AX_A & =P. \end{align}

Consequently, \(X_A\) is a unit in the matrix algebra, with inverse \(Z_A\), and \(\det (X_A)\) is a unit in \(\mathbb C\).

In the notation of Proposition IV.5 of [ CPGSV17 ] , let \(L\) and \(R\) be the positive semidefinite outer factors with support projections \(P\) and \(Q\), respectively. The required witnesses are

\begin{align} X & =X_L=L^++(\mathbb {1}-P), \\ Z & =X_R=R^++(\mathbb {1}-Q). \end{align}

Thus

\begin{align} X_L L & =P, \\ R X_R & =Q, \end{align}

or equivalently \(XL=P\) and \(RZ=Q\). Here the paper’s matrix \(Z\) is \(X_R\), not the inverse-extension matrix \(Z_R\) from Definition 19.9.57.

Proof

The support identities give

\begin{align} A^+A=AA^+ & =P. \end{align}

Their complementary forms are

\begin{align} A^+(\mathbb {1}-P)=(\mathbb {1}-P)A^+ & =0, \end{align}
\begin{align} A(\mathbb {1}-P)=(\mathbb {1}-P)A & =0, \end{align}

and idempotence of \(P\) gives

\begin{align} (\mathbb {1}-P)^2 & =\mathbb {1}-P. \end{align}

Therefore

\begin{align} X_AZ_A & =P+(\mathbb {1}-P)=\mathbb {1}. \end{align}

Likewise,

\begin{align} Z_AX_A & =P+(\mathbb {1}-P)=\mathbb {1}, \end{align}

while multiplication by \(A\) gives \(X_AA=AX_A=P\).

Let \([0,1]\ni x\mapsto \mathcal W(x)\) be a continuous path of tensors generating MPUs, without requiring the path to be in canonical form. For each \(x\), take the canonical form and the two source-cut ranks \(r(x)\) and \(\ell (x)\) of its associated simple tensor. Then \(r(x)\) and \(\ell (x)\) are constant on \([0,1]\).

Proposition 19.9.60 Admissible constancy of the source ranks

Let \([0,1]\ni x\mapsto \mathcal W(x)\) be a continuous path of tensors generating MPUs, without requiring the path to be in canonical form. Suppose that the path carries an admissible equivalence datum with blocking length \(m\). Then the ranks \(r(x)\) and \(\ell (x)\) of the actual blocked path tensors \(\mathcal W(x)_m\) are constant on \([0,1]\).

For MPU tensors under the source’s standing canonical-form convention, take the index defined from any simple blocking. This index has the following properties.

  1. It does not change under blocking.

  2. It is additive under tensor products and composition.

  3. It is constant under continuous deformations through MPU tensors.

  4. Two MPU tensors have the same index if and only if they are equivalent.

This is the unrestricted statement of [ CPGSV17 , Theorem IV.6 ] .

Assume the following operation-specific admissibility hypotheses. Every pair of simple blockings compared for blocking independence carries reduced full-support source data. For a tensor product \(\mathcal U\otimes \mathcal V\), there is one positive length \(k\) such that \(\mathcal U_k\), \(\mathcal V_k\), and \((\mathcal U\otimes \mathcal V)_k\) are simple and carry their own data. For a composition tensor \(\mathcal W\) of \(\mathcal U\) and \(\mathcal U'\), there is one positive length \(k\) such that

\begin{align} \mathcal U_k,\quad \mathcal U’_k,\quad \mathcal W_k,\quad \mathcal U_{2k},\quad \mathcal U’_{2k},\quad \mathcal W_{4k} \end{align}

are simple and carry their own data. Continuous-deformation and equivalence claims use the path data of Definition 19.9.29. Under these hypotheses, the MPU index has the following properties.

  1. It does not change under blocking.

  2. It is additive under tensor products and composition.

  3. It is constant under continuous deformations through MPU tensors.

  4. Admissibly equivalent MPU tensors have the same index. Conversely, tensors with the same index are admissibly equivalent whenever the standard-form interpolation is supplied with an admissible path datum.

These are the conclusions of [ CPGSV17 , Theorem IV.6 ] , restricted to the explicit admissibility hypotheses above.

Corollary 19.9.63 Continuous choice of standard form

For every continuous path \(p\mapsto \mathcal U(p)\) of MPU tensors, not necessarily in canonical form, there are an integer \(k_0\leq D^4\) and continuous paths of unitaries \(u(p)\) and \(v(p)\) such that \(U^{(2k_0N)}(p)\) has standard form given by \(u(p)\) and \(v(p)\).

Corollary 19.9.64 Admissible continuous choice of standard form

Let \(k_0\) be the common blocking length in an admissible equivalence datum for a continuous path \(\mathcal U(p)\) of MPU tensors, and suppose that this supplied length satisfies \(k_0\leq D^4\). Then there are continuous paths of unitaries \(u(p)\) and \(v(p)\) such that the MPU \(U^{(2k_0N)}(p)\) has standard form given by \(u(p)\) and \(v(p)\) for every \(p\in [0,1]\). This is the reduced-source version of [ CPGSV17 , Corollary IV.7 ] .

19.9.1 Local characterization of symmetries

Proposition 19.9.1.1 Conjugation symmetry

An MPU in standard form is invariant under complex conjugation if and only if there are unitaries \(x\) and \(y\) such that either \(x=x^T\) and \(y=y^T\), or \(x=-x^T\) and \(y=-y^T\), and

\begin{align} \overline u & =(x\otimes y)u, & \overline v & =v(y^\dagger \otimes x^\dagger ). \end{align}
Proposition 19.9.1.2 Admissible conjugation symmetry

Let a tensor and its conjugate comparison tensor both be equipped with reduced full-support source data and written in admissible standard form. The generated MPU is invariant under complex conjugation if and only if there are unitaries \(x\) and \(y\) such that either \(x=x^T\) and \(y=y^T\), or \(x=-x^T\) and \(y=-y^T\), and

\begin{align} \overline u & =(x\otimes y)u, & \overline v & =v(y^\dagger \otimes x^\dagger ). \end{align}

This is the reduced-source version of the conjugation characterization in [ CPGSV17 , Proposition V.1 ] .

Lemma 19.9.1.3 Normal form of symmetric and skew-symmetric unitaries

Let \(x\) be unitary and satisfy \(x=x^T\) or \(x=-x^T\). There are a symmetric unitary \(S\) and a matrix \(\widetilde\Lambda \) such that

\begin{align} x & =S^T\widetilde\Lambda S, & S & =S^T. \end{align}

In the symmetric case, \(\widetilde\Lambda \) is real and symmetric. In the skew-symmetric case, it is real and skew-symmetric. In both cases, \(\det (\widetilde\Lambda )=1\).

Lemma 19.9.1.4 Odd-dimensional skew-symmetric determinant

Let \(I\) be a finite set of odd cardinality and let \(A\in \operatorname {Mat}_I(\mathbb C)\) satisfy \(A^T=-A\). Then

\begin{align} \det A & =0. \end{align}
Proof

Transpose invariance, skew symmetry, and the determinant of a negated matrix give

\begin{align} \det A & =\det (A^T) =\det (-A) =(-1)^{|I|}\det A =-\det A. \end{align}

Since the complex numbers have characteristic zero, \(\det A=0\).

Lemma 19.9.1.5 Parity of a nonsingular skew-symmetric matrix

Let \(I\) be finite and let \(A\in \operatorname {Mat}_I(\mathbb C)\) satisfy \(A^T=-A\) and \(\det A\ne 0\). Then \(|I|\) is even.

Proof

If \(|I|\) were odd, Lemma 19.9.1.4 would give \(\det A=0\). Hence \(|I|\) is not odd and is therefore even.

Theorem 19.9.1.6 Even dimension of a skew-symmetric unitary

Let \(U\in \operatorname {Mat}_d(\mathbb C)\) be unitary and satisfy \(U^T=-U\). Then \(d\) is even. This is the parity obstruction in [ CPGSV17 , Proposition V.2, Case II ] .

Proof

The determinant of a unitary matrix is a unit, so it is nonzero. Apply Lemma 19.9.1.5 to \(U\).

Proposition 19.9.1.7 Orthogonal and symplectic standard forms

Let an MPU in standard form be invariant under complex conjugation. Its standard-form unitaries can be chosen in exactly one of the following two cases. In Case I,

\begin{align} \overline{u'} & =u’, & \overline{v'} & =v’, \end{align}

uniquely up to local orthogonal transformations. In Case II,

\begin{align} \overline{u'} & =(\Sigma _\ell \otimes \Sigma _r)u’, & \overline{v'} & =v’(\Sigma _r\otimes \Sigma _\ell ), \end{align}

uniquely up to local symplectic transformations, where \(\Sigma _x=\mathbb {1}_{x/2}\otimes \left(\begin{smallmatrix} 0 & 1 \\ -1 & 0 \end{smallmatrix}\right)\). Case II can occur only when \(r\) and \(\ell \) are even.

Proposition 19.9.1.8 Admissible orthogonal and symplectic standard forms

Let an MPU in admissible standard form be invariant under complex conjugation, and suppose every standard-form tensor compared below carries reduced full-support source data. Its standard-form unitaries can be chosen in exactly one of the following two cases. In Case I,

\begin{align} \overline{u'} & =u’, & \overline{v'} & =v’, \end{align}

and the choice is unique up to local orthogonal transformations. In Case II,

\begin{align} \overline{u'} & =(\Sigma _\ell \otimes \Sigma _r)u’, & \overline{v'} & =v’(\Sigma _r\otimes \Sigma _\ell ), \end{align}

where

\begin{align} \Sigma _x & =\mathbb {1}_{x/2}\otimes Y, & Y & =\begin{pmatrix} 0 & 1 \\ -1 & 0 \end{pmatrix}. \end{align}

The choice is unique up to local transformations \(S_x\) satisfying \(S_x^T\Sigma _xS_x=\Sigma _x\). Case II can occur only when \(r\) and \(\ell \) are even.

Corollary 19.9.1.9 Reduction of the symplectic case to orthogonal unitaries

In Case II of Proposition 19.9.1.7, one can write

\begin{align} u’ & =(\mathbb {1}_{r\ell /4}\otimes R)\widetilde u, & v’ & =\widetilde v(\mathbb {1}_{r\ell /4}\otimes R), \end{align}

where \(\widetilde u\) and \(\widetilde v\) are orthogonal and

\begin{align} R & =\frac{1-i}{2}[\mathbb {1}_4+i(Y\otimes Y)]. \end{align}
Corollary 19.9.1.10 Admissible reduction of the symplectic case to orthogonal unitaries

In the admissible Case II of Proposition 19.9.1.8, one can write

\begin{align} u’ & =(\mathbb {1}_{r\ell /4}\otimes R)\widetilde u, & v’ & =\widetilde v(\mathbb {1}_{r\ell /4}\otimes R), \end{align}

where \(\widetilde u\) and \(\widetilde v\) are orthogonal and

\begin{align} R & =\frac{1-i}{2}[\mathbb {1}_4+i(Y\otimes Y)]. \end{align}
Proposition 19.9.1.11 Local characterization of time reversal

Let an MPU be in standard form, with one-site physical space \(\mathbb C^{d_0}\otimes \mathbb C^{d_0}\). It satisfies \(U^{(N)}=U^{(N)\dagger }\) for every even \(N\) if and only if there is a unitary \(x\) such that

\begin{align} (\mathbb {1}_1\otimes v_{23}\otimes \mathbb {1}_4)(u_{12}\otimes u_{34}) & =\pm (x_{12}\otimes x_{34}^\dagger ) (\mathbb {1}_1\otimes v_{23}^\dagger \otimes \mathbb {1}_4). \end{align}

Let an MPU be in admissible standard form. Suppose its one-site tensor \(\mathcal U\), its actual two-site block \(\mathcal U_2\), and the physical-adjoint comparison tensor \(\mathcal U_2^\sharp \) are simple and each carries its own reduced full-support source datum. Write the one-site physical space as \(\mathbb C^{d_0}\otimes \mathbb C^{d_0}\). Time-reversal symmetry then forces \(\operatorname {ind}(\mathcal U)=0\). Under these hypotheses, \(U^{(N)}=U^{(N)\dagger }\) for every even \(N\) if and only if there is a unitary \(x\) such that

\begin{align} (\mathbb {1}_1\otimes v_{23}\otimes \mathbb {1}_4)(u_{12}\otimes u_{34}) & =\pm (x_{12}\otimes x_{34}^\dagger ) (\mathbb {1}_1\otimes v_{23}^\dagger \otimes \mathbb {1}_4). \end{align}
Proof

Under time reversal, \(\mathcal U_2\) and \(\mathcal U_2^\sharp \) generate the same MPU. The admissible fundamental theorem identifies their source spaces, so their corresponding source-cut ranks agree. Physical adjunction exchanges the two cuts. Hence

\begin{align} r(\mathcal U_2) & =r(\mathcal U_2^\sharp ) =\ell (\mathcal U_2). \end{align}

Definition 19.9.17 therefore gives \(\operatorname {ind}(\mathcal U_2)=0\), and blocking independence gives \(\operatorname {ind}(\mathcal U)=0\). Apply the admissible fundamental theorem to the supplied standard forms of \(\mathcal U_2\) and \(\mathcal U_2^\sharp \) and follow the source proof. The converse follows by contracting the displayed local relation around every even periodic chain.

Proposition 19.9.1.13 Local characterization of transposition

Let an MPU be in standard form, with one-site physical space \(\mathbb C^{d_0}\otimes \mathbb C^{d_0}\). It satisfies \(U^{(N)}=U^{(N)T}\) for every even \(N\) if and only if there are a unitary \(x\) and a phase \(e^{i\phi }\) such that

\begin{align} (\mathbb {1}_1\otimes v_{23}\otimes \mathbb {1}_4)(u_{12}\otimes u_{34}) & =e^{i\phi }(x_{12}\otimes x_{34}^T) (\mathbb {1}_1\otimes v_{23}^T\otimes \mathbb {1}_4). \end{align}

Let an MPU be in admissible standard form. Suppose its one-site tensor \(\mathcal U\), its actual two-site block \(\mathcal U_2\), and the transposed comparison tensor \(\mathcal U_2^{\mathsf T}\) are simple and each carries its own reduced full-support source datum. Write the one-site physical space as \(\mathbb C^{d_0}\otimes \mathbb C^{d_0}\). Transposition symmetry then forces \(\operatorname {ind}(\mathcal U)=0\). Under these hypotheses, \(U^{(N)}=U^{(N)T}\) for every even \(N\) if and only if there are a unitary \(x\) and a phase \(e^{i\phi }\) such that

\begin{align} (\mathbb {1}_1\otimes v_{23}\otimes \mathbb {1}_4)(u_{12}\otimes u_{34}) & =e^{i\phi }(x_{12}\otimes x_{34}^T) (\mathbb {1}_1\otimes v_{23}^T\otimes \mathbb {1}_4). \end{align}
Proof

Under transposition, \(\mathcal U_2\) and \(\mathcal U_2^{\mathsf T}\) generate the same MPU. The admissible fundamental theorem identifies their source spaces, so their corresponding source-cut ranks agree. Transposition exchanges the two cuts. Hence

\begin{align} r(\mathcal U_2) & =r(\mathcal U_2^{\mathsf T}) =\ell (\mathcal U_2). \end{align}

Thus \(\operatorname {ind}(\mathcal U_2)=0\) by Definition 19.9.17, and blocking independence gives \(\operatorname {ind}(\mathcal U)=0\). Apply the admissible fundamental theorem to the supplied standard forms of \(\mathcal U_2\) and \(\mathcal U_2^{\mathsf T}\) and repeat the time-reversal argument, retaining the unrestricted phase \(e^{i\phi }\). The converse follows by contracting the local relation around every even periodic chain.

19.9.2 Equivalence under symmetries

Definition 19.9.2.1 Finite-chain operator symmetry

A finite-chain operator symmetry \(\mathcal S\) specifies a set \(A\) of applicable chain lengths and, for every physical dimension \(d\) and length \(N\), a map

\begin{align} \mathcal S_{d,N}\colon \operatorname {Mat}_{d^N}(\mathbb C) & \longrightarrow \operatorname {Mat}_{d^N}(\mathbb C). \end{align}

We suppress the subscripts on \(\mathcal S_{d,N}\) when \(d\) and \(N\) are understood. An MPO tensor \(\mathcal W\) is invariant under \(\mathcal S\) if

\begin{align} \mathcal S_{d,N}[W^{(N)}] & =W^{(N)} \end{align}

for every \(N\in A\).

Definition 19.9.2.2 Strict equivalence under a symmetry

Fix a virtual bond dimension \(D\). Two canonical-form tensors \(\mathcal U\) and \(\mathcal V\) are strictly equivalent under a symmetry \(\mathcal S\) if they have the same physical dimension and there is a continuous path \(\mathcal W(p)\) of fixed-bond MPU tensors from \(\mathcal U\) to \(\mathcal V\), not necessarily in canonical form, such that

\begin{align} \mathcal S[W^{(N)}(p)] & =W^{(N)}(p) \end{align}

for every \(p\in [0,1]\) and every applicable chain length \(N\).

Lemma 19.9.2.3 Forgetting symmetry from strict equivalence

Strict equivalence under a symmetry implies strict equivalence.

Proof

The same path lies in the larger locus obtained by forgetting invariance under \(\mathcal S\).

Definition 19.9.2.4 Admissible strict equivalence under a symmetry

Two tensors \(\mathcal U\) and \(\mathcal V\) are strictly equivalent under a symmetry \(\mathcal S\) if they have the same physical dimension and there is a continuous path \(\mathcal W(p)\) of MPU tensors from \(\mathcal U\) to \(\mathcal V\) such that

\begin{align} \mathcal S[W^{(N)}(p)] & =W^{(N)}(p) \end{align}

for every \(p\in [0,1]\) and every chain length \(N\) under consideration. The path need not be in canonical form. The equivalence is admissible if its witnessing path carries an admissible equivalence datum for the actual blocked path tensor and the fixed conjugate, physical-adjoint, transposed, and \(\mathcal S\)-image comparison paths, and if the common blocking length \(m\) of that datum satisfies \(m\leq D^4\).

Definition 19.9.2.5 Equivalence under a symmetry

Let \(\mathcal U\) and \(\mathcal V\) have physical dimensions \(d_a\) and \(d_b\). They are equivalent under a symmetry \(\mathcal S\) if there are a positive blocking length \(k\) and coprime positive ancilla dimensions \(p_a,p_b\) with \(p_ad_a=p_bd_b\) such that \((\mathcal U^{(p_a)})_k\) and \((\mathcal V^{(p_b)})_k\) are strictly equivalent under \(\mathcal S\).

Definition 19.9.2.6 Admissible equivalence under a symmetry

Let \(\mathcal U\) and \(\mathcal V\) have physical dimensions \(d_a\) and \(d_b\). They are equivalent under a symmetry \(\mathcal S\) if there are a positive blocking length \(k\) and coprime positive ancilla dimensions \(p_a,p_b\) with \(p_ad_a=p_bd_b\) such that \((\mathcal U^{(p_a)})_k\) and \((\mathcal V^{(p_b)})_k\) are strictly equivalent under \(\mathcal S\). Thus the ancillas are attached before the common blocking. The equivalence is admissible if this strict-equivalence witness carries an admissible equivalence datum for the actual blocked path and the fixed conjugate, physical-adjoint, transposed, and \(\mathcal S\)-image comparison paths, with common blocking length at most \(D^4\).

Lemma 19.9.2.7 Standard-form path criterion

Two MPUs described by simple tensors are strictly equivalent under \(\mathcal S\) if and only if their standard-form unitaries are joined by continuous families \(\widetilde u(p)\) and \(\widetilde v(p)\) whose generated MPU is invariant under \(\mathcal S\) for every \(p\in [0,1]\).

Lemma 19.9.2.8 Admissible standard-form path criterion

Fix a positive integer \(m\leq D^4\). The actual \(m\)-site blocked endpoint MPUs \(U_1^{(mN)}\) and \(U_2^{(mN)}\) are admissibly strictly equivalent under \(\mathcal S\) if and only if there are continuous families of unitaries \(\widetilde u(p)\) and \(\widetilde v(p)\) whose endpoints generate \(U_1^{(mN)}\) and \(U_2^{(mN)}\), whose generated MPU is invariant under \(\mathcal S\) for every \(p\in [0,1]\), and whose tensor path and fixed conjugate, physical-adjoint, transposed, and \(\mathcal S\)-image comparison paths carry continuously varying reduced full-support source data. These families arise from the actual blocked path \(\mathcal W(p)_m\) in an admissible equivalence datum of common length \(m\); the blocked path itself is the length-one admissible witness for the displayed endpoint families. When \(m{\gt}1\), this criterion does not assert strict equivalence of the unblocked families \(U_1^{(N)}\) and \(U_2^{(N)}\). This is the reduced-source form of [ CPGSV17 , Lemma following the symmetry-equivalence definitions ] .

Theorem 19.9.2.9 Conjugation symmetry when one source rank is odd

Suppose that either \(r\) or \(\ell \) is odd. Two conjugation-invariant MPUs \(\mathcal U_1\) and \(\mathcal U_2\) are strictly equivalent under conjugation if and only if

\begin{align} \det (u_1v_1) & =\det (u_2v_2). \end{align}
Theorem 19.9.2.10 Admissible conjugation symmetry when one source rank is odd

For \(i=1,2\), suppose the original one-site endpoint tensor \(\mathcal U_i\) and its conjugate comparison tensor are simple and each carries reduced full-support source data. Suppose that either \(r\) or \(\ell \) is odd and that the standard-form interpolation determined below carries an admissible path datum of blocking length one. Then \(\mathcal U_1\) and \(\mathcal U_2\) are admissibly strictly equivalent under conjugation symmetry if and only if

\begin{align} \det (u_1v_1) & =\det (u_2v_2). \end{align}
Corollary 19.9.2.11 Equivalence of conjugation-symmetric MPUs after blocking

If either \(r\) or \(\ell \) is odd, all conjugation-invariant MPUs with these source ranks are equivalent under conjugation symmetry.

For \(i=1,2\), suppose the endpoint tensor \(\mathcal U_i\), its conjugate comparison \(\overline{\mathcal U_i}\), the actual two-site block \(\mathcal U_{i,2}\), and its conjugate comparison \(\overline{\mathcal U_{i,2}}\) are simple and each carries reduced full-support source data. Suppose that either \(r\) or \(\ell \) is odd and that the resulting standard-form interpolation carries an admissible path datum of blocking length one. Then \(\mathcal U_1\) and \(\mathcal U_2\) are admissibly equivalent under conjugation symmetry.

Proof

If either \(r\) or \(\ell \) is odd, both endpoints lie in Case I of Proposition 19.9.1.8. The full admissible classification in Corollary 19.9.2.16 handles both possible parities of the source ranks after blocking and gives the claimed admissible equivalence.

Theorem 19.9.2.13 Strict conjugation equivalence for even source ranks

Suppose \(r\) and \(\ell \) are even. Two conjugation-invariant MPUs in the standard forms of Proposition 19.9.1.7 are strictly equivalent under conjugation if and only if they belong to the same case and

\begin{align} \det (u_1) & =\det (u_2), & \det (v_1) & =\det (v_2). \end{align}
Theorem 19.9.2.14 Admissible strict conjugation equivalence for even source ranks

Suppose that \(r\) and \(\ell \) are both even and that the standard-form interpolation determined below carries an admissible path datum of blocking length one. Choose admissible standard forms of two conjugation-invariant MPUs as in Proposition 19.9.1.8. They are admissibly strictly equivalent under conjugation symmetry if and only if they belong to the same case of Proposition 19.9.1.8 and

\begin{align} \det (u_1) & =\det (u_2), & \det (v_1) & =\det (v_2). \end{align}
Corollary 19.9.2.15 Equivalence under conjugation

Two conjugation-invariant MPUs are equivalent under conjugation symmetry if and only if they belong to the same case of Proposition 19.9.1.7.

For \(i=1,2\), suppose the original one-site endpoint tensor \(\mathcal U_i\), its conjugate comparison \(\overline{\mathcal U_i}\), the actual two-site block \(\mathcal U_{i,2}\), and its conjugate comparison \(\overline{\mathcal U_{i,2}}\) are simple and each carries reduced full-support source data. Suppose also that the resulting standard-form interpolation carries an admissible path datum of blocking length one. Then \(\mathcal U_1\) and \(\mathcal U_2\) are admissibly equivalent under conjugation symmetry if and only if they belong to the same case of Proposition 19.9.1.8.

Proof

Suppose first that either \(r\) or \(\ell \) is odd. Then both original endpoints lie in Case I. After one blocking, write \(r^{(2)}\) and \(\ell ^{(2)}\) for the blocked source ranks. Admissible blocking independence, applied to the supplied one-site and two-site endpoint data, gives

\begin{align} r^{(2)} & =dr, \end{align}

and

\begin{align} \ell ^{(2)} & =d\ell . \end{align}

For each endpoint,

\begin{align} \det (u_i^{(2)}v_i^{(2)}) & =\det (u_iv_i)^{2r\ell }=1. \end{align}

Blocking preserves Case I. If either \(r^{(2)}\) or \(\ell ^{(2)}\) is odd, the odd-rank theorem applied to the blocked endpoints gives admissible strict equivalence. If both blocked ranks are even, then \(d\) is even because one original source rank is odd. Hence \(r\ell =d^2\) is even. Consequently,

\begin{align} \det (u_i^{(2)}) & =(\det (u_i)^2\det (v_i))^{r\ell }=1, \end{align}

and

\begin{align} \det (v_i^{(2)}) & =\det (v_i)^{r\ell }=1. \end{align}

The even-rank theorem now applies to the preserved Case I endpoints. Thus the odd-source-rank branch has a single admissible equivalence class without imposing a parity condition on the blocked ranks.

If \(r\) and \(\ell \) are both even, admissible blocking independence shows that the blocked ranks remain even. Since \(r\ell \) is even, the same calculations give

\begin{align} \det (u_i^{(2)}) & =(\det (u_i)^2\det (v_i))^{r\ell }=1, \end{align}

and

\begin{align} \det (v_i^{(2)}) & =\det (v_i)^{r\ell }=1. \end{align}

Blocking preserves Case I and Case II. The even-rank theorem therefore shows that the blocked endpoints are admissibly strictly equivalent exactly when they belong to the same case. This is precisely admissible equivalence of the original endpoints.

For matrices \(A,B\in \mathbb C^{n\times n}\) and the exchange matrix \(\mathbb S\) on \(\mathbb C^n\otimes \mathbb C^n\),

\begin{align} \operatorname{tr}\! \left[\mathbb S(A\otimes B)\right] & =\operatorname{tr}(AB). \end{align}

In particular, if \(U\) is unitary, then

\begin{align} \operatorname{tr}\! \left[\mathbb S(U\otimes U^\dagger )\right] & =n. \end{align}
Proof

In product coordinates, the exchange matrix has entries \(\mathbb S_{(i,j),(k,l)}=\delta _{i,l}\delta _{j,k}\). Therefore

\begin{align} \operatorname{tr}\! \left[\mathbb S(A\otimes B)\right] & =\sum _{i,j}A_{j,i}B_{i,j} =\sum _{i,j}A_{i,j}B_{j,i} =\operatorname{tr}(AB). \end{align}

Taking \(A=U\) and \(B=U^\dagger \) gives \(\operatorname{tr}(UU^\dagger )=\operatorname{tr}(\mathbb {1})=n\).

Lemma 19.9.2.18 Trace formula for the time-reversal sign

For a time-reversal-invariant MPU in standard form, the sign \(\sigma \) is

\begin{align} \sigma & =\frac1{d^2}\operatorname{tr}\! \left[\mathbb S_{12,34} v_{23}u_{12}u_{34}v_{23}\right], \end{align}

where \(\mathbb S_{12,34}\) exchanges the pairs of sites \(12\) and \(34\).

Lemma 19.9.2.19 Admissible trace formula for the time-reversal sign

For a time-reversal-invariant MPU in admissible standard form, suppose its one-site tensor \(\mathcal U\), its actual two-site block \(\mathcal U_2\), and the physical-adjoint comparison tensor \(\mathcal U_2^\sharp \) are simple and each carries its own reduced full-support source datum. Then the sign \(\sigma \) in Proposition 19.9.1.12 is

\begin{align} \sigma & =\frac1{d^2}\operatorname{tr}\! \left[\mathbb S_{12,34}v_{23}u_{12}u_{34}v_{23}\right], \end{align}

where \(\mathbb S_{12,34}\) exchanges the pairs of sites \(12\) and \(34\).

Lemma 19.9.2.20 Gauge invariance of the time-reversal sign

Under \(u'=(x\otimes y)u\) and \(v'=v(y^\dagger \otimes x^\dagger )\), the trace formula in Lemma 19.9.2.18 has the same value for \((u',v')\) as for \((u,v)\). Thus \(\sigma \) is independent of the standard-form gauge.

Lemma 19.9.2.21 Admissible gauge invariance of the time-reversal sign

Under the hypotheses of Lemma 19.9.2.19, let \(x\) and \(y\) be unitaries and change the standard-form gauge by

\begin{align} u’ & =(x\otimes y)u, & v’ & =v(y^\dagger \otimes x^\dagger ). \end{align}

Then the trace formula has the same value for \((u',v')\) as for \((u,v)\). Consequently the sign \(\sigma \) is independent of the standard-form gauge.

Proof

Write \(x_i\) and \(y_i\) for the unitaries acting on site \(i\). Then \(u'_{12}=x_1y_2u_{12}\), \(u'_{34}=x_3y_4u_{34}\), and \(v'_{23}=v_{23}y_2^\dagger x_3^\dagger \). Substitution into Lemma 19.9.2.19, cancellation on the middle legs, and \(\mathbb S_{12,34}x_1y_4=x_3y_2\mathbb S_{12,34}\) give

\begin{align} \operatorname{tr}\! \left[\mathbb S_{12,34}v’_{23}u’_{12}u’_{34}v’_{23}\right] & =\operatorname{tr}\! \left[x_3y_2\mathbb S_{12,34} v_{23}u_{12}u_{34}v_{23}y_2^\dagger x_3^\dagger \right] \\ & =\operatorname{tr}\! \left[\mathbb S_{12,34}v_{23}u_{12}u_{34}v_{23}\right]. \end{align}

The last equality follows from cyclicity of the trace and unitarity of \(x\) and \(y\).

Proposition 19.9.2.22 Well-definedness of the time-reversal sign

For a time-reversal-invariant MPU in standard form, the sign \(\sigma =\pm 1\) is independent of the gauge and is constant under continuous deformations that preserve time-reversal symmetry.

Proposition 19.9.2.23 Admissible well-definedness of the time-reversal sign

Suppose a time-reversal-preserving path carries an admissible equivalence datum of blocking length one. At every path point, require the one-site tensor, its actual two-site block, and the physical-adjoint comparison of that two-site block to be simple and to carry continuously varying reduced full-support source data. Then the sign \(\sigma =\pm 1\) is independent of the admissible standard-form gauge and is constant along the path.

Proposition 19.9.2.24 Blocking parity of the time-reversal sign

Let \(\mathcal U\) be a time-reversal-invariant MPU in standard form, and let \(\sigma ^{(k)}\) be the sign obtained from a standard form of its \(k\)-site block. Then

\begin{align} \sigma ^{(k)} & =\begin{cases} \sigma ^{(1)}, & k\text{ odd}, \\ \sigma ^{(2)}, & k\text{ even}. \end{cases}\end{align}

For every blocking length \(k\) under consideration, suppose the endpoint tensor \(\mathcal U_k\), its actual two-site block \(\mathcal U_{2k}\), and the physical-adjoint comparison tensor \(\mathcal U_{2k}^\sharp \) are simple and each carries its own reduced full-support source datum. Let \(\sigma ^{(k)}\) be the sign of the resulting standard form. Then

\begin{align} \sigma ^{(k)} & =\begin{cases} \sigma ^{(1)}, & k\text{ odd}, \\ \sigma ^{(2)}, & k\text{ even}. \end{cases}\end{align}
Proof

Gauge invariance permits the standard blocked representatives used in the source proof. Choose

\begin{align} u^{(2)} & =v_{23}u_{12}u_{34}, & v^{(2)} & =v_{45}, \end{align}

and put \(\zeta =\sigma ^{(2)}/\sigma ^{(1)}\). The one- and two-site local time-reversal relations give the unwinding identity

\begin{align} \zeta \big[(\mathbb {1}\otimes v\otimes \mathbb {1})(u\otimes x)\big]^\dagger & =(\mathbb {1}\otimes v\otimes \mathbb {1})(x^\dagger \otimes u). \end{align}

Applying this identity successively through a block of length \(k\) gives

\begin{align} \sigma ^{(k)} & =\sigma ^{(1)}\zeta ^{k-1}. \end{align}

Since each sign is \(\pm 1\), this equals \(\sigma ^{(1)}\) for odd \(k\) and \(\sigma ^{(2)}\) for even \(k\).

Corollary 19.9.2.26 Necessary time-reversal invariants

Strictly equivalent time-reversal-invariant MPUs have the same \(\sigma ^{(1)}\) and \(\sigma ^{(2)}\). Equivalent time-reversal-invariant MPUs have the same \(\sigma ^{(2)}\).

Corollary 19.9.2.27 Admissible time-reversal invariants of supplied witnesses

Let two time-reversal-invariant MPUs be compared by an admissible symmetry path \(p\mapsto \mathcal W(p)\) whose path datum has blocking length one. Suppose that the five path families \(\mathcal W(p)\), \(\mathcal W(p)_2\), \(\mathcal W(p)_4\), \(\mathcal W(p)_2^\sharp \), and \(\mathcal W(p)_4^\sharp \) are simple and carry continuously varying reduced full-support source data. Then admissible strict equivalence requires equality of both \(\sigma ^{(1)}\) and \(\sigma ^{(2)}\).

For admissible equivalence after adjoining ancillas and applying the source blocking, let \(p\mapsto \mathcal W_{\mathrm{anc}}(p)\) be the actual ancilla-enlarged blocked witness path, not a reduced or equivalent representative, and regard it as a length-one admissible path. Require continuous reduced full-support source data along \(\mathcal W_{\mathrm{anc}}(p)\), its actual two- and four-site blocks, and the physical-adjoint comparisons of those two blocks. The conclusion is only

\begin{align} \sigma ^{(2)}(\mathcal W_{\mathrm{anc}}(0)) & =\sigma ^{(2)}(\mathcal W_{\mathrm{anc}}(1)). \end{align}

No equality between \(\sigma ^{(2)}\) of the original two MPUs is asserted. Such a conclusion requires preservation under adjoining identity ancillas, which is not established here.

Proposition 19.9.2.28 Conjugation class and the MPS sign

For a conjugation-invariant MPU, the MPS sign \(G\overline G=\pm \mathbb {1}\) agrees with Case I and Case II, respectively, of Proposition 19.9.1.7.

Proposition 19.9.2.29 Admissible conjugation class and the MPS sign

Let \(\mathcal U\) be conjugation invariant. Suppose \(\mathcal U\) and its conjugate comparison tensor are simple and carry their own reduced full-support source data. Let \(p\mapsto \mathcal W(p)\) be the actual standard-form interpolation from \(\mathcal U\) to its Case I or Case II representative. Suppose this interpolation carries an admissible conjugation-symmetry path datum of blocking length one: the path \(\mathcal W(p)\) and its fixed conjugate, physical-adjoint, transposed, and conjugation-image comparison paths are simple and carry continuously varying source data. Then the sign \(G\overline G=\pm \mathbb {1}\) in the MPS classification agrees with Case I and Case II, respectively, of Proposition 19.9.1.8.

Proposition 19.9.2.30 Time-reversal class and the MPS sign

For a time-reversal-invariant MPU in standard form, the MPS sign \(G\overline G=\pm \mathbb {1}\) equals the sign \(\sigma ^{(2)}\) obtained after blocking two sites.

Proposition 19.9.2.31 Admissible time-reversal class and the MPS sign

Let an MPU in admissible standard form be invariant under time reversal. Suppose \(\mathcal U\), \(\mathcal U_2\), and \(\mathcal U_4\), together with the physical-adjoint comparison tensors \(\mathcal U_2^\sharp \) and \(\mathcal U_4^\sharp \), are simple and each carries its own reduced full-support source datum. Then the MPS sign \(G\overline G=\pm \mathbb {1}\) equals the sign \(\sigma ^{(2)}\).

19.9.3 Symmetry examples

Proposition 19.9.3.1 All phases under conjugation

Let \(u_+=v_+=\mathbb {1}_{r\ell }\) and \(u_-=v_-=\operatorname {diag}(-1,1,\ldots ,1)\). If at least one of \(r\) and \(\ell \) is odd, \((u_+,v_+)\) and \((u_+,v_-)\) realize the two classes of Theorem 19.9.2.9. If both are even, the four pairs \((u_\pm ,v_\pm )\) realize the four Case I classes. Multiplying by the matrix \(R\) of Corollary 19.9.1.9 realizes the four Case II classes.

Assume the standard-form tensors displayed below carry reduced full-support source data and that the comparison paths used to classify them carry admissible path data of blocking length one. Let

\begin{align} u_+ & =v_+=\mathbb {1}_{r\ell }, & u_- & =v_-=\operatorname {diag}(-1,1,\ldots ,1). \end{align}

If at least one of \(r\) and \(\ell \) is odd, the pairs \((u_+,v_+)\) and \((u_+,v_-)\) represent the two classes of Theorem 19.9.2.10. If \(r\) and \(\ell \) are both even, the four pairs \((u_\pm ,v_\pm )\) represent the four Case I classes of Theorem 19.9.2.14. With

\begin{align} u’_\pm & =(\mathbb {1}_{r\ell /4}\otimes R)u_\pm , & v’_\pm & =v_\pm (\mathbb {1}_{r\ell /4}\otimes R), \end{align}

the four pairs \((u'_\pm ,v'_\pm )\) represent the four Case II classes.

The CZX MPU is defined by

\begin{align} u & =CZ(\sigma _x\otimes \sigma _x), & v & =CZ, \end{align}

where \(CZ=\operatorname {diag}(1,1,1,-1)\). Let

\begin{align} H & =\frac1{\sqrt2}\begin{pmatrix} 1 & 1 \\ 1 & -1 \end{pmatrix}\end{align}

be the Hadamard matrix. The corresponding tensor is

\begin{align} \mathcal U_{\ell r}^{ud} & =(\sigma _x)_{ud}\delta _{u,\ell }H_{\ell ,r}. \end{align}

It has \(\sigma ^{(1)}=\sigma ^{(2)}=-1\).

Suppose the CZX tensors \(\mathcal U\), \(\mathcal U_2\), and \(\mathcal U_4\), together with the physical-adjoint comparison tensors \(\mathcal U_2^\sharp \) and \(\mathcal U_4^\sharp \), are simple and each carries its own reduced full-support source datum. The CZX MPU is defined by

\begin{align} u & =CZ(\sigma _x\otimes \sigma _x), & v & =CZ, \\ CZ & =\operatorname {diag}(1,1,1,-1). \end{align}

Its tensor is

\begin{align} \mathcal U_{\ell r}^{ud} & =(\sigma _x)_{ud} \delta _{u,\ell }H_{\ell ,r}, & H & =\frac1{\sqrt2}\begin{pmatrix} 1 & 1 \\ 1 & -1 \end{pmatrix}. \end{align}

It has \(\sigma ^{(1)}=\sigma ^{(2)}=-1\).

Proposition 19.9.3.5 The other three time-reversal classes

The standard forms

\begin{align} u & =v=\mathbb {1}, \\ u & =i\mathbb {1}, & v & =\mathbb {1}, \\ u & =i\, CZ(\sigma _x\otimes \sigma _x), & v & =CZ \end{align}

realize, respectively,

\begin{align} (\sigma ^{(1)},\sigma ^{(2)}) & =(1,1),\quad (-1,1),\quad (1,-1). \end{align}
Proposition 19.9.3.6 Admissible remaining time-reversal classes

For each displayed example, suppose the tensors \(\mathcal U\), \(\mathcal U_2\), and \(\mathcal U_4\), together with the physical-adjoint comparison tensors \(\mathcal U_2^\sharp \) and \(\mathcal U_4^\sharp \), are simple and each carries its own reduced full-support source datum. The standard forms

\begin{align} u & =v=\mathbb {1}, \\ u & =i\mathbb {1}, & v & =\mathbb {1}, \\ u & =i\, CZ(\sigma _x\otimes \sigma _x), & v & =CZ \end{align}

realize, respectively,

\begin{align} (\sigma ^{(1)},\sigma ^{(2)}) & =(1,1),\quad (-1,1),\quad (1,-1). \end{align}
Definition 19.9.3.7 Right- and left-shift tensors

The right shift on \((\mathbb C^d)^{\otimes N}\) is

\begin{align} T^{(N)}\lvert n_1,\ldots ,n_N\rangle & =\lvert n_N,n_1,\ldots ,n_{N-1}\rangle . \end{align}

It has bond dimension \(d\) and local matrices \(A^{ij}=\lvert i\rangle \langle j\rvert \). The physical adjoint of \(A\) defines the left-shift tensor.

Proof

Closing the virtual indices of \(A^{ij}=\lvert i\rangle \langle j\rvert \) leaves one nonzero virtual configuration. If \(\tau _n=\sigma _{n+1}\) for every \(n\) modulo \(N\), then

\begin{align} \langle \sigma \vert T^{(N)}\vert \tau \rangle & =1. \end{align}

Otherwise the same matrix entry is zero. Hence the output configuration is \((\tau _{N-1},\tau _0,\ldots ,\tau _{N-2})\), which fixes the permutation-matrix convention and gives the right shift. Its permutation matrix is unitary. Physical adjunction gives the conjugate transpose, hence the left shift and its unitarity.

Let \(A\) be the right-shift tensor and let \(z_{(\alpha ,i)}=\delta _{\alpha i}\) be the vectorized identity. Its raw source cuts are

\begin{align} M_1(A) & =\mathbb {1}, \\ M_2(A) & =zz^{\mathsf T}. \end{align}

For the left-shift tensor \(A^\sharp \) they are

\begin{align} M_1(A^\sharp ) & =zz^{\mathsf T}, \\ M_2(A^\sharp ) & =\mathbb {1}. \end{align}

In every physical dimension, including \(d=0\),

\begin{align} r(A) & =d^2, \\ \ell (A^\sharp ) & =d^2. \end{align}

If \(d{\gt}0\), the remaining two ranks are

\begin{align} \ell (A) & =1, \\ r(A^\sharp ) & =1. \end{align}

These are the raw source-cut quantities from [ CPGSV17 , Section III ] ; no choice of source factors is involved.

Proof

The entry formula \(A^{ij}_{\alpha \beta } =\delta _{i\alpha }\delta _{j\beta }\) gives \(M_1(A)=\mathbb {1}\) and \(M_2(A)=zz^{\mathsf T}\). Hence the first rank is \(d^2\). For \(d{\gt}0\), the vector \(z\) has a nonzero diagonal coordinate, so its outer product has rank one. Physical adjunction exchanges the two cuts and their ranks, which gives all four left-shift formulas without a second rank calculation.

On two \(d\)-dimensional spins per site, define

\begin{align} U_1^{(N)} & =(\mathbb {1}\otimes \mathbb {1})^{\otimes N}, \\ U_2^{(N)} & =T^{(N)\dagger }\otimes T^{(N)}, \\ U_3^{(N)} & =T^{(N)}\otimes T^{(N)\dagger }. \end{align}
Proof

The bond-one identity tensor generates the identity at every length. The finite-chain operator of an independent tensor product is the Kronecker product of the two finite-chain operators, after the sitewise identification \((\mathbb C^d\otimes \mathbb C^d)^{\otimes N} \cong (\mathbb C^d)^{\otimes N}\otimes (\mathbb C^d)^{\otimes N}\). Applying this identity to the right and left shifts gives the three displayed formulas. The identity tensor is unitary, and tensor products preserve unitarity, so all three tensors are matrix product unitaries.

Assume \(d{\gt}0\). Equip \(U_1\) with the tensor product of the two bond-one identity source factorizations. Equip \(U_2\) with the tensor product of the left-shift and right-shift source factorizations, in that order, and equip \(U_3\) with the tensor product in the reverse order. In all three cases the virtual weight is the identity. These are explicit supplied source-factor witnesses; no identification with an arbitrary compact singular-value decomposition is asserted.

Assume \(d{\gt}0\). In the source coordinates of the supplied factorizations,

\begin{align} u_1 & =\mathbb {1}\otimes \mathbb {1}, & v_1 & =\mathbb {1}\otimes \mathbb {1}, \\ u_3 & =\mathbb S, & v_3 & =\mathbb S. \end{align}

These are the formulas in [ CPGSV17 , lines 2009–2016 ] . The \(U_3\) equalities use the source-coordinate presentation of that equation; they are distinct from the blocked four-spin presentation below.

Proof

Write the source and physical indices as \(((a,b),(c,e))\) and \(((i,j),(k,l))\), respectively, and set \(s=d^{-1/2}\). The identity factors give

\begin{align} [u_1]_{((a,b),(c,e)),((i,j),(k,l))} & =[v_1]_{((i,j),(k,l)),((a,b),(c,e))} \\ & =(\delta _{a,i}\delta _{c,k})(\delta _{b,j}\delta _{e,l}) \\ & =\delta _{a,i}\delta _{b,j}\delta _{c,k}\delta _{e,l}. \end{align}

For \(U_3\), the right-shift and left-shift factors instead give

\begin{align} [u_3]_{((a,b),(c,e)),((i,j),(k,l))} & =(s\delta _{c,i}\delta _{a,k}) (d\, s\delta _{e,j}\delta _{b,l}) \\ & =\delta _{a,k}\delta _{b,l}\delta _{c,i}\delta _{e,j}, \\ {}[v_3]_{((i,j),(k,l)),((a,b),(c,e))} & =(d\, s\delta _{a,k}\delta _{c,i}) (s\delta _{b,l}\delta _{e,j}) \\ & =\delta _{a,k}\delta _{b,l}\delta _{c,i}\delta _{e,j}, \end{align}

where \(d\, s^2=1\). The first display is the identity matrix, while the second interchanges the two composite indices and is therefore \(\mathbb S\) in both orientations.

Assume \(d{\gt}0\). The two open-leg factorizations of the blocked tensor have, as their respective central gates,

\begin{align} u_2^{(2)} & =\mathbb {1}\otimes \mathbb S\otimes \mathbb {1}, \\ v_2^{(2)} & =(\mathbb S\otimes \mathbb S) (\mathbb {1}\otimes \mathbb S\otimes \mathbb {1}). \end{align}

Here the two physical sites are decoded into the four-spin order of the paper, and the factors \(X_1,Y_2\) and \(X_2,Y_1\) remain as the two open boundaries of the respective factorizations. This is [ CPGSV17 , lines 2018–2026 ] .

Proof

Write \(B_2=\mathcal U_{2,2}\) for the actual two-site block and decode \(I=(I_1,I_2)\) and \(J=(J_1,J_2)\). If \(\pi \) is the four-spin physical identification and \(\iota _u,\iota _v\) are the two source-rank identifications, the asserted open-leg equalities are

\begin{align} (B_2)_{I,J;\alpha ,\gamma } & =\sum _{r,l}(X_1)_{(\alpha ,J_1),r} [\mathbb {1}\otimes \mathbb S\otimes \mathbb {1}]_{\iota _u^{-1}(l,r),\pi ^{-1}(I_1,I_2)} (Y_2)_{l,(J_2,\gamma )}, \\ (B_2)_{I,J;\alpha ,\gamma } & =\sum _{l,r}(X_2)_{(\alpha ,I_1),l} [(\mathbb S\otimes \mathbb S) (\mathbb {1}\otimes \mathbb S\otimes \mathbb {1})]_{\pi ^{-1}(J_2,J_1),\iota _v^{-1}(r,l)} (Y_1)_{r,(I_2,\gamma )}. \end{align}

Write \(s=d^{-1/2}\), so that \(d\, s^2=1\). In the product-rank coordinates, the entries of the two central gates are

\begin{align} [u_2^{(2)}]_{((a,b),(c,e)),((i,j),(k,l))} & =(d\, s\delta _{a,i}\delta _{b,k}) (s\delta _{c,j}\delta _{e,l}) \\ & =\delta _{a,i}\delta _{b,k}\delta _{c,j}\delta _{e,l}, \\ {}[v_2^{(2)}]_{((i,j),(k,l)),((a,b),(c,e))} & =(s\delta _{e,k}\delta _{c,i}) (d\, s\delta _{b,l}\delta _{a,j}) \\ & =\delta _{a,j}\delta _{b,l}\delta _{c,i}\delta _{e,k}. \end{align}

These are respectively the entries of \(\mathbb {1}\otimes \mathbb S\otimes \mathbb {1}\) and \((\mathbb S\otimes \mathbb S)(\mathbb {1}\otimes \mathbb S\otimes \mathbb {1})\); substituting them into the exact open two-site factorizations gives the asserted equalities.

Assume \(d{\gt}0\). The two open-leg factorizations of the blocked tensor have, as their respective central gates,

\begin{align} u_3^{(2)} & =(\mathbb {1}\otimes \mathbb S\otimes \mathbb {1}) (\mathbb S\otimes \mathbb S), \\ v_3^{(2)} & =\mathbb {1}\otimes \mathbb S\otimes \mathbb {1}. \end{align}

Again the four-spin order is the one displayed in the paper, and the source factors at the two open boundaries are retained explicitly. This is [ CPGSV17 , lines 2028–2034 ] .

Proof

With \(B_3=\mathcal U_{3,2}\) for the actual two-site block and the corresponding identifications \(\pi ,\iota _u,\iota _v\), the exact equalities are

\begin{align} (B_3)_{I,J;\alpha ,\gamma } & =\sum _{r,l}(X_1)_{(\alpha ,J_1),r} [(\mathbb {1}\otimes \mathbb S\otimes \mathbb {1}) (\mathbb S\otimes \mathbb S)]_{\iota _u^{-1}(l,r),\pi ^{-1}(I_1,I_2)} (Y_2)_{l,(J_2,\gamma )}, \\ (B_3)_{I,J;\alpha ,\gamma } & =\sum _{l,r}(X_2)_{(\alpha ,I_1),l} [\mathbb {1}\otimes \mathbb S\otimes \mathbb {1}]_{\pi ^{-1}(J_2,J_1),\iota _v^{-1}(r,l)} (Y_1)_{r,(I_2,\gamma )}. \end{align}

Again write \(s=d^{-1/2}\), so that \(d\, s^2=1\). Reversing the two supplied factors gives the entry identities

\begin{align} [u_3^{(2)}]_{((a,b),(c,e)),((i,j),(k,l))} & =(s\delta _{c,i}\delta _{e,k}) (d\, s\delta _{a,j}\delta _{b,l}) \\ & =\delta _{a,j}\delta _{b,l}\delta _{c,i}\delta _{e,k}, \\ {}[v_3^{(2)}]_{((i,j),(k,l)),((a,b),(c,e))} & =(d\, s\delta _{b,k}\delta _{a,i}) (s\delta _{e,l}\delta _{c,j}) \\ & =\delta _{a,i}\delta _{b,k}\delta _{c,j}\delta _{e,l}. \end{align}

The first is the entry of \((\mathbb {1}\otimes \mathbb S\otimes \mathbb {1})(\mathbb S\otimes \mathbb S)\), and the second is the entry of \(\mathbb {1}\otimes \mathbb S\otimes \mathbb {1}\). Substitution into the two exact open two-site factorizations proves the claim.

Proposition 19.9.3.16 Standard forms of the three shift MPUs

Assume \(d{\gt}0\). For the fixed standard-form source factors, the corresponding unitaries satisfy

\begin{align} u_1 & =\mathbb {1}\otimes \mathbb {1}, & v_1 & =\mathbb {1}\otimes \mathbb {1}, \\ u_3 & =\mathbb S, & v_3 & =\mathbb S, \\ u_2^{(2)} & =\mathbb {1}\otimes \mathbb S\otimes \mathbb {1}, & v_2^{(2)} & =(\mathbb S\otimes \mathbb S) (\mathbb {1}\otimes \mathbb S\otimes \mathbb {1}), \\ u_3^{(2)} & =(\mathbb {1}\otimes \mathbb S\otimes \mathbb {1}) (\mathbb S\otimes \mathbb S), & v_3^{(2)} & =\mathbb {1}\otimes \mathbb S\otimes \mathbb {1}. \end{align}

The preceding theorems establish these matrices for explicit supplied source-factor witnesses. It remains to identify that supplied choice with the fixed standard-form choice.

Definition 19.9.3.17 Transformation with swaps

Define

\begin{align} \widetilde U_1^{(N)} & =\mathbb S^{\otimes N}, \\ \widetilde U_2^{(N)} & =\mathbb S^{\otimes N} (T^{(N)\dagger }\otimes T^{(N)}), \\ \widetilde U_3^{(N)} & =\mathbb S^{\otimes N} (T^{(N)}\otimes T^{(N)\dagger }) =\mathbb S^{\otimes N}\widetilde U_2^{(N)}\mathbb S^{\otimes N}. \end{align}

The first two standard forms are

\begin{align} \widetilde u_1 & =\mathbb S, & \widetilde v_1 & =\mathbb {1}\otimes \mathbb {1}, \\ \widetilde u_2 & =\mathbb {1}\otimes \mathbb {1}, & \widetilde v_2 & =\mathbb S. \end{align}
Lemma 19.9.3.18 Transformation with swaps and symmetry equivalence

Two MPUs \(U_i^{(N)}\) and \(U_j^{(N)}\) are in the same phase under time reversal combined with the local swap, transposition combined with the local swap, or conjugation if and only if the corresponding \(\widetilde U_i^{(N)}\) and \(\widetilde U_j^{(N)}\) are in the same phase under time reversal, transposition, or conjugation, respectively.

Lemma 19.9.3.19 Admissible transformation with swaps and symmetry equivalence

Fix one of the three symmetry comparisons and let \(p\mapsto \mathcal W(p)\) be the actual source-constructed interpolation used in that comparison. Suppose it carries an admissible symmetry-path datum of blocking length one. Thus the path \(\mathcal W(p)\) and the four fixed comparison paths \(\overline{\mathcal W(p)}\), \(\mathcal W(p)^\sharp \), \(\mathcal W(p)^{\mathsf T}\), and \(\mathcal S[\mathcal W(p)]\) are simple and carry continuously varying reduced full-support source data. Suppose also that every one- and two-site endpoint tensor used in the local comparison carries its own datum. Then \(U_i^{(N)}\) and \(U_j^{(N)}\) are in the same admissible phase under time reversal combined with the local swap, transposition combined with the local swap, or conjugation if and only if the corresponding \(\widetilde U_i^{(N)}\) and \(\widetilde U_j^{(N)}\) are in the same admissible phase under time reversal, transposition, or conjugation, respectively.

Proposition 19.9.3.20 Conjugation phases of \(U_2\) and \(U_3\)

After blocking two sites, \(U_2^{(N)}\) and \(U_3^{(N)}\) are both in the trivial phase under conjugation.

Proposition 19.9.3.21 Admissible conjugation phases of \(U_2\) and \(U_3\)

Suppose the two-site blocked tensors carry reduced full-support source data and the standard-form comparison paths carry admissible path data of blocking length one. Then the actual two-site blocked tensors of \(U_2\) and \(U_3\) are both in the trivial admissible strict phase under conjugation. Consequently \(U_2\) and \(U_3\) lie in the same admissible equivalence class.

Proposition 19.9.3.22 Time-reversal and transposition phases of \(U_2\) and \(U_3\)

The MPUs \(\widetilde U_2^{(N)}\) and \(\widetilde U_3^{(N)}\) are in the same phase under both time reversal and transposition.

Proposition 19.9.3.23 Admissible time-reversal and transposition phases of \(U_2\) and \(U_3\)

For each of time reversal and transposition, let \(p\mapsto \mathcal W_{\mathcal S}(p)\) be the actual source-constructed interpolation from \(\widetilde U_2\) to \(\widetilde U_3\). Suppose it carries an admissible symmetry-path datum of blocking length one. Thus the path \(\mathcal W_{\mathcal S}(p)\) and its fixed conjugate, physical-adjoint, transposed, and \(\mathcal S\)-image comparison paths are simple and carry continuously varying reduced full-support source data. Suppose also that the one- and two-site endpoint tensors used in the local comparison carry their own data. Then \(\widetilde U_2^{(N)}\) and \(\widetilde U_3^{(N)}\) are in the same admissible phase under both time reversal and transposition.

Proposition 19.9.3.24 Conjugation phases of \(U_1\) and \(U_2\)

The MPUs \(U_1^{(N)}\) and \(U_2^{(N)}\) lie in different strict conjugation phases exactly when \(d=4k+2\) for a nonnegative integer \(k\). Otherwise they lie in the same strict phase. All these phases are in Case I, so the two MPUs are equivalent after blocking and adjoining ancillas.

Proposition 19.9.3.25 Admissible conjugation phases of \(U_1\) and \(U_2\)

For each endpoint tensor \(\mathcal A\) among \(U_1\), \(U_2\), \(\widetilde U_1\), and \(\widetilde U_2\), suppose that \(\mathcal A\), its conjugate comparison \(\overline{\mathcal A}\), its actual two-site block \(\mathcal A_2\), and the conjugate comparison \(\overline{\mathcal A_2}\) are simple and carry their own reduced full-support source data. For every strict comparison used below, let \(p\mapsto \mathcal W(p)\) be the actual source-constructed interpolation between the corresponding endpoint tensors. Suppose it carries an admissible conjugation-symmetry path datum of blocking length one: the path and its fixed conjugate, physical-adjoint, transposed, and conjugation-image comparison paths are simple and carry continuously varying source data. Require a length-one admissible path datum also for the actual blocked interpolation used in the final equivalence claim. Under these hypotheses, \(U_1^{(N)}\) and \(U_2^{(N)}\) are in different admissible strict phases under conjugation exactly when \(k\) is a nonnegative integer and

\begin{align} d & =4k+2. \end{align}

Otherwise they are in the same admissible strict phase. All these standard forms lie in Case I, so \(U_1^{(N)}\) and \(U_2^{(N)}\) are admissibly equivalent after blocking and adding ancillas.

Proposition 19.9.3.26 Ancilla equivalence of \(U_1\) and \(U_2\)

After adding one ancilla of dimension \(d\) per site, \(\widetilde U_1^{(N)}\) and \(\widetilde U_2^{(N)}\) are strictly equivalent under both time reversal and transposition.

Proposition 19.9.3.27 Admissible ancilla equivalence of \(U_1\) and \(U_2\)

After adding one ancilla of dimension \(d\) per site and performing the source blocking, denote the actual enlarged blocked endpoints by \(\widehat{\mathcal U}_1\) and \(\widehat{\mathcal U}_2\). For either time reversal or transposition, let \(p\mapsto \mathcal W_{\mathcal S}(p)\) be the source-constructed interpolation from \(\widehat{\mathcal U}_1\) to \(\widehat{\mathcal U}_2\). Suppose this interpolation carries an admissible symmetry-path datum of blocking length one. Thus the path \(\mathcal W_{\mathcal S}(p)\) and its fixed conjugate, physical-adjoint, transposed, and \(\mathcal S\)-image comparison paths are simple and carry continuously varying reduced full-support source data. Suppose also that the one- and two-site enlarged endpoint tensors used in the local comparison carry their own data. Then \(\widehat{\mathcal U}_1\) and \(\widehat{\mathcal U}_2\) are admissibly strictly equivalent under both time reversal and transposition. No equivalence of the original ancilla-free tensors is asserted, because the equal ancilla dimensions used here need not satisfy the coprimality condition in Definition 19.9.2.6.

19.10 Finite-region observable algebras

Definition 19.10.1 Finite-region local algebra
#

Fix a nonnegative integer \(d\). For a finite region \(\Lambda \subset \mathbb Z\), the configuration space and local observable algebra are

\begin{align} \operatorname {Config}_d(\Lambda ) & =\{ 0,\ldots ,d-1\} ^{\Lambda }, \\ \mathcal A_\Lambda & =M_{\operatorname {Config}_d(\Lambda )}(\mathbb C) \cong M_d^{\otimes \Lambda }. \end{align}

The algebra \(\mathcal A_\Lambda \) carries its operator norm. This is the finite-region construction used in the QCA appendix of [ CPGSV17 ] ; no completion is taken here.

Definition 19.10.2 Canonical inclusion of finite-region observables
#

If \(\Lambda \subseteq \Gamma \) are finite regions, define

\begin{align} \iota _{\Lambda ,\Gamma } & \colon \mathcal A_\Lambda \longrightarrow \mathcal A_\Gamma , \\ \iota _{\Lambda ,\Gamma }(A) & = A\otimes \mathbf1_{\Gamma \setminus \Lambda }. \end{align}

Here product configurations are identified with configurations on \(\Gamma \). This is a unital star-algebra homomorphism.

Lemma 19.10.3 Injectivity of the finite-region inclusion

For \(d{\gt}0\), the map \(\iota _{\Lambda ,\Gamma }\) is injective.

Proof

Since \(d{\gt}0\), choose a configuration \(c\) on \(\Gamma \setminus \Lambda \). For configurations \(x,y\) on \(\Lambda \),

\begin{align} \iota _{\Lambda ,\Gamma }(A)((x,c),(y,c)) & =A(x,y)\cdot \mathbf1_{\Gamma \setminus \Lambda }(c,c) =A(x,y). \end{align}

Hence \(\iota _{\Lambda ,\Gamma }(A)=\iota _{\Lambda ,\Gamma }(B)\) implies \(A=B\).

Lemma 19.10.4 Operator-norm preservation

If \(d{\gt}0\), then for every \(A\in \mathcal A_\Lambda \),

\begin{align} \lVert \iota _{\Lambda ,\Gamma }(A)\rVert & =\lVert A\rVert . \end{align}
Proof

An injective star-algebra homomorphism between complex \(C^*\)-algebras is isometric.

Lemma 19.10.5 Isometry of the finite-region inclusion

For \(d{\gt}0\), the map \(\iota _{\Lambda ,\Gamma }\) is an operator-norm isometry.

Proof

Use injectivity and the general theorem that an injective star-algebra homomorphism between complex \(C^*\)-algebras is an isometry.

Lemma 19.10.6 Identity inclusion
#

For every finite region \(\Lambda \), \(\iota _{\Lambda ,\Lambda }\) is the identity on \(\mathcal A_\Lambda \).

Proof

When the two regions agree, their complement is empty. Hence

\begin{align} \iota _{\Lambda ,\Lambda }(A) & =A\otimes \mathbf1_{\varnothing } =A\otimes 1 =A. \end{align}
Lemma 19.10.7 Composition of finite-region inclusions
#

If \(\Lambda \subseteq \Gamma \subseteq \Delta \), then

\begin{align} \iota _{\Gamma ,\Delta }\circ \iota _{\Lambda ,\Gamma } & =\iota _{\Lambda ,\Delta }. \end{align}
Proof

The complement decomposes as the disjoint union

\begin{align} \Delta \setminus \Lambda & =(\Gamma \setminus \Lambda )\sqcup (\Delta \setminus \Gamma ), \\ \mathbf1_{\Delta \setminus \Lambda } & =\mathbf1_{\Gamma \setminus \Lambda } \otimes \mathbf1_{\Delta \setminus \Gamma }. \end{align}

Therefore

\begin{align} \iota _{\Gamma ,\Delta }(\iota _{\Lambda ,\Gamma }(A)) & =(A\otimes \mathbf1_{\Gamma \setminus \Lambda }) \otimes \mathbf1_{\Delta \setminus \Gamma } \\ & =A\otimes \mathbf1_{\Delta \setminus \Lambda } \\ & =\iota _{\Lambda ,\Delta }(A). \end{align}
Theorem 19.10.8 Disjoint finite-region observables commute

If \(\Lambda \cap \Gamma =\varnothing \), \(A\in \mathcal A_\Lambda \), and \(B\in \mathcal A_\Gamma \), then their images in the algebra of the union satisfy

\begin{align} \iota _{\Lambda ,\Lambda \cup \Gamma }(A) \iota _{\Gamma ,\Lambda \cup \Gamma }(B) & =\iota _{\Gamma ,\Lambda \cup \Gamma }(B) \iota _{\Lambda ,\Lambda \cup \Gamma }(A). \end{align}

This tensor-factor identity is local algebra infrastructure for the quasi-local algebra used in [ CPGSV17 , lines 2292–2300 ] ; the appendix does not state it as a separate theorem.

Proof

For configurations \(x,y\) on \(\Lambda \cup \Gamma \), split each intermediate configuration \(z\) into \((z|_\Lambda ,z|_\Gamma )\). The matrix entry of the left-hand side is

\begin{align} \sum _{z_\Lambda ,z_\Gamma } A(x|_\Lambda ,z_\Lambda ) \delta _{x|_\Gamma ,z_\Gamma } B(z_\Gamma ,y|_\Gamma ) \delta _{z_\Lambda ,y|_\Lambda } & =A(x|_\Lambda ,y|_\Lambda ) B(x|_\Gamma ,y|_\Gamma ). \end{align}

Reversing the two factors gives the same scalar product.

Theorem 19.10.9 Finite relative commutant

Let \(\Lambda \subseteq \Gamma \) be finite regions and \(A\in \mathcal A_\Gamma \). Then \(A\) acts trivially outside \(\Lambda \) if and only if it commutes with the full observable algebra on the complement:

\begin{align} A\in \iota _{\Lambda ,\Gamma }(\mathcal A_\Lambda ) & \iff \forall B\in \mathcal A_{\Gamma \setminus \Lambda },\quad A\iota _{\Gamma \setminus \Lambda ,\Gamma }(B) =\iota _{\Gamma \setminus \Lambda ,\Gamma }(B)A. \end{align}

This is a finite-dimensional tensor-factor identity. It is not a statement of Schumacher–Werner or the QCA appendix of [ CPGSV17 ] .

Proof

Splitting configurations gives \(\mathcal A_\Gamma \cong \mathcal A_\Lambda \otimes \mathcal A_{\Gamma \setminus \Lambda }\). Under this identification, the two inclusions are \(C\mapsto C\otimes \mathbf1\) and \(B\mapsto \mathbf1\otimes B\). Write an arbitrary matrix in blocks indexed by configurations on \(\Lambda \). If it commutes with every \(\mathbf1\otimes B\), each block commutes with the full matrix algebra on \(\Gamma \setminus \Lambda \) and is therefore scalar. Thus the matrix is \(C\otimes \mathbf1\) for some \(C\in \mathcal A_\Lambda \). The converse follows from

\begin{align} (C\otimes \mathbf1)(\mathbf1\otimes B) & =C\otimes B =(\mathbf1\otimes B)(C\otimes \mathbf1). \end{align}

19.10.1 Bipartite support algebras

Let \(I\) and \(J\) be finite sets. For \(X\in M_{I\times J}(\mathbb C)\), the existing left coefficient slice and right operator block are the matrices

\begin{align} X^{\mathrm L}_{rs}(i,j) & =X((i,r),(j,s)), & X^{\mathrm R}_{ij}(r,s) & =X((i,r),(j,s)). \end{align}

For a star-subalgebra \(\mathcal D\subseteq M_K(\mathbb C)\), write \(\mathcal D'=\{ Y\in M_K(\mathbb C)\mid [Y,Z]=0\text{ for every }Z\in \mathcal D\} \) for its commutant. The construction below is the full complex matrix-algebra specialization of the coefficient construction in [ SW04 , Section 4.3 ] and of the leastness and commutant statements in [ GNVW12 , Lemma 7 ] .

Local fix (star-closed input): The printed statement of Lemma 7 says only that \(\mathcal A\) is a subalgebra. Its coefficient-generated support and leastness clauses remain valid for arbitrary algebra input, but its commutant characterization fails unless \(\mathcal A\) is closed under the involution. For instance, the algebra \(\operatorname {span}\{ \mathbf1,E_{12}\} \subset M_2(\mathbb C)\) gives a counterexample to the printed commutant clause if involution closure is omitted. Schumacher and Werner first define the ordinary algebra generated by the coefficient support and then derive its adjoint closure from adjoint closure of the observable-algebra input [ SW04 , lines 1161–1171 ] . Accordingly, the statements below use a star-subalgebra. This is a necessary correction to the printed hypothesis. In the QCA application the input is the image of a local observable algebra under a star-automorphism [ GNVW12 , Section 7.2 ] . The MPU application likewise takes the image of a local star-algebra under a star-automorphism [ CPGSV17 , lines 2300–2306 ] . This correction, including the counterexample without involution closure and the exact application ranges, is recorded in [ con26w ] .

Scope restriction (full matrix factors): After this correction, Lemma 7 permits arbitrary finite-dimensional \(C^*\)-algebras \(\mathcal B_1\) and \(\mathcal B_2\), whereas the statements below take \(\mathcal B_1=M_I(\mathbb C)\) and \(\mathcal B_2=M_J(\mathbb C)\). Full matrix factors suffice for the MPU spin chain because every one-site algebra is isomorphic to \(M_d(\mathbb C)\) [ CPGSV17 , lines 2320–2334 ] . The general-factor result is not claimed here; see [ con26k ] .

No full-matrix structure theorem, dimension formula, Margolus gate, depth-two circuit, MPU standard form, or index comparison is asserted here.

Theorem 19.10.1.1 Adjoint of a right coefficient block
#

For every \(X\in M_{I\times J}(\mathbb C)\) and \(i,j\in I\),

\begin{align} (X^\dagger )^{\mathrm R}_{ij} & =\bigl(X^{\mathrm R}_{ji}\bigr)^\dagger . \end{align}
Proof

Compare the \((r,s)\) entries. Both sides equal \(\overline{X((j,s),(i,r))}\).

Definition 19.10.1.2 Left coefficient set
#

Let \(\mathcal A\subseteq M_{I\times J}(\mathbb C)\) be a star-subalgebra. Its set of left coefficients is

\begin{align} \operatorname {Coeff}_{\mathrm L}(\mathcal A) & =\bigl\{ X^{\mathrm L}_{rs}\ \bigm | X\in \mathcal A,\ r,s\in J\bigr\} \subseteq M_I(\mathbb C). \end{align}
Definition 19.10.1.3 Right coefficient set
#

Let \(\mathcal A\subseteq M_{I\times J}(\mathbb C)\) be a star-subalgebra. Its set of right coefficients is

\begin{align} \operatorname {Coeff}_{\mathrm R}(\mathcal A) & =\bigl\{ X^{\mathrm R}_{ij}\ \bigm | X\in \mathcal A,\ i,j\in I\bigr\} \subseteq M_J(\mathbb C). \end{align}
Definition 19.10.1.4 Left support algebra
#

The left support algebra of \(\mathcal A\) is the star-subalgebra

\begin{align} \operatorname {Spp}_{\mathrm L}(\mathcal A) & =\operatorname {alg}^* \bigl(\operatorname {Coeff}_{\mathrm L}(\mathcal A)\bigr) \subseteq M_I(\mathbb C) \end{align}

generated by all left coefficients.

Definition 19.10.1.5 Right support algebra
#

The right support algebra of \(\mathcal A\) is the star-subalgebra

\begin{align} \operatorname {Spp}_{\mathrm R}(\mathcal A) & =\operatorname {alg}^* \bigl(\operatorname {Coeff}_{\mathrm R}(\mathcal A)\bigr) \subseteq M_J(\mathbb C) \end{align}

generated by all right coefficients.

Theorem 19.10.1.6 Left coefficient generators belong to the support algebra

If \(X\in \mathcal A\) and \(r,s\in J\), then

\begin{align} X^{\mathrm L}_{rs} & \in \operatorname {Spp}_{\mathrm L}(\mathcal A). \end{align}
Proof

The matrix \(X^{\mathrm L}_{rs}\) is one of the generators of the defining star-algebra closure.

Theorem 19.10.1.7 Right coefficient generators belong to the support algebra

If \(X\in \mathcal A\) and \(i,j\in I\), then

\begin{align} X^{\mathrm R}_{ij} & \in \operatorname {Spp}_{\mathrm R}(\mathcal A). \end{align}
Proof

The matrix \(X^{\mathrm R}_{ij}\) is one of the generators of the defining star-algebra closure.

Theorem 19.10.1.8 Reconstruction from left coefficients

For every \(X\in M_{I\times J}(\mathbb C)\),

\begin{align} X & =\sum _{r,s\in J}X^{\mathrm L}_{rs}\otimes E_{rs}. \end{align}
Proof

At the entry \(((i,r),(j,s))\), only the matrix unit \(E_{rs}\) contributes, and its coefficient is \(X^{\mathrm L}_{rs}(i,j)=X((i,r),(j,s))\).

Theorem 19.10.1.9 Reconstruction from right coefficients

For every \(X\in M_{I\times J}(\mathbb C)\),

\begin{align} X & =\sum _{i,j\in I}E_{ij}\otimes X^{\mathrm R}_{ij}. \end{align}
Proof

At the entry \(((i,r),(j,s))\), only the matrix unit \(E_{ij}\) contributes, and its coefficient is \(X^{\mathrm R}_{ij}(r,s)=X((i,r),(j,s))\).

Theorem 19.10.1.10 Minimality of the left support algebra

For every star-subalgebra \(\mathcal B\subseteq M_I(\mathbb C)\),

\begin{align} \operatorname {Spp}_{\mathrm L}(\mathcal A)\subseteq \mathcal B \quad \Longleftrightarrow \quad & X^{\mathrm L}_{rs}\in \mathcal B \quad \text{for every }X\in \mathcal A\text{ and }r,s\in J. \end{align}
Proof

Use the universal property of the star-subalgebra generated by \(\operatorname {Coeff}_{\mathrm L}(\mathcal A)\).

Theorem 19.10.1.11 Minimality of the right support algebra

For every star-subalgebra \(\mathcal C\subseteq M_J(\mathbb C)\),

\begin{align} \operatorname {Spp}_{\mathrm R}(\mathcal A)\subseteq \mathcal C \quad \Longleftrightarrow \quad & X^{\mathrm R}_{ij}\in \mathcal C \quad \text{for every }X\in \mathcal A\text{ and }i,j\in I. \end{align}
Proof

Use the universal property of the star-subalgebra generated by \(\operatorname {Coeff}_{\mathrm R}(\mathcal A)\).

Theorem 19.10.1.12 Tensor-submodule criterion for left support

Let \(I=\{ 0,\ldots ,p-1\} \) and \(J=\{ 0,\ldots ,q-1\} \). For every star-subalgebra \(\mathcal B\subseteq M_p(\mathbb C)\),

\begin{align} \operatorname {Spp}_{\mathrm L}(\mathcal A)\subseteq \mathcal B \quad \Longleftrightarrow \quad \mathcal A & \subseteq \mathcal B\otimes M_q(\mathbb C), \end{align}

where the right-hand side is the tensor-product submodule consisting of matrices whose left coefficient slices lie in \(\mathcal B\).

Proof

Membership in the tensor-product submodule is equivalent to membership of every left coefficient in \(\mathcal B\). Apply Theorem 19.10.1.10.

Definition 19.10.1.13 Kronecker-product submodule
#

For submodules \(S\subseteq M_I(\mathbb C)\) and \(T\subseteq M_J(\mathbb C)\), define

\begin{align} S\boxtimes T & =\operatorname {span}_{\mathbb C} \{ A\otimes B\mid A\in S,\ B\in T\} \subseteq M_{I\times J}(\mathbb C). \end{align}

This is the tensor-product subspace used below, rather than a conjunction of two one-sided containment conditions.

Theorem 19.10.1.14 Coefficient criterion for the Kronecker-product submodule

For \(X\in M_{I\times J}(\mathbb C)\),

\begin{align} X\in S\boxtimes T \quad \Longleftrightarrow \quad & X^{\mathrm L}_{rs}\in S \quad \text{for every }r,s\in J, \\ & \text{and}\quad X^{\mathrm R}_{ij}\in T \quad \text{for every }i,j\in I. \end{align}
Proof

For \(A\in S\) and \(B\in T\), direct evaluation gives

\begin{align} (A\otimes B)^{\mathrm L}_{rs} & =B_{rs}A, & (A\otimes B)^{\mathrm R}_{ij} & =A_{ij}B. \end{align}

Thus the forward implication extends by linearity. Conversely, expand the right blocks in a basis of \(T\). Applying the dual coordinate functionals to those blocks produces matrices in \(S\) from the left-slice hypothesis, and the resulting finite sum reconstructs \(X\) as a sum of Kronecker products from \(S\) and \(T\).

Theorem 19.10.1.15 The one-sided tensor submodule as a Kronecker span

Let \(I=\{ 0,\ldots ,p-1\} \) and \(J=\{ 0,\ldots ,q-1\} \). For every submodule \(S\subseteq M_p(\mathbb C)\),

\begin{align} S\boxtimes M_q(\mathbb C) & =\{ X\in M_{p\times q}(\mathbb C) \mid X^{\mathrm L}_{rs}\in S\text{ for all }r,s\} . \end{align}

Thus the two-sided Kronecker span extends the existing left-oriented tensor-submodule construction.

Proof

In Theorem 19.10.1.14, take the right submodule to be the full matrix space. The right-block condition is then automatic, leaving precisely the left-slice criterion.

For every star-subalgebra \(\mathcal C\subseteq M_J(\mathbb C)\),

\begin{align} \operatorname {Spp}_{\mathrm R}(\mathcal A)\subseteq \mathcal C \quad \Longleftrightarrow \quad \mathcal A & \subseteq M_I(\mathbb C)\boxtimes \mathcal C. \end{align}
Proof

The coefficient criterion reduces the right-hand containment to membership of every right block in \(\mathcal C\). This is equivalent to the left-hand inclusion by Theorem 19.10.1.11.

Every bipartite star-subalgebra satisfies the literal two-sided containment

\begin{align} \mathcal A & \subseteq \operatorname {Spp}_{\mathrm L}(\mathcal A) \boxtimes \operatorname {Spp}_{\mathrm R}(\mathcal A). \end{align}

This is the full complex matrix-algebra specialization of the support-algebra containment in Schumacher–Werner Section 4.3 and of Gross–Nesme–Vogts–Werner equation (22). The latter follows after applying Lemma 7 to both tensor factors. The displayed assertion is an inclusion in an actual tensor-product span, not merely the conjunction of the two one-sided inclusions. The MPU appendix applies this construction to \(\mathcal R_{2x}\) in [ CPGSV17 , lines 2320–2334 ] .

Proof

If \(X\in \mathcal A\), every left coefficient of \(X\) lies in the left support algebra and every right coefficient lies in the right support algebra. Theorem 19.10.1.14 places \(X\) in their Kronecker-product submodule.

Theorem 19.10.1.18 Left Kronecker commutation is coefficientwise
#

For \(B\in M_I(\mathbb C)\) and \(X\in M_{I\times J}(\mathbb C)\),

\begin{align} [B\otimes \mathbf1_J,X]=0 \quad \Longleftrightarrow \quad [B,X^{\mathrm L}_{rs}] & =0 \quad \text{for every }r,s\in J. \end{align}
Proof

For \(i,j\in I\) and \(r,s\in J\), the two product entries are

\begin{align} [(B\otimes \mathbf1_J)X]_{(i,r),(j,s)} & =\sum _{k\in I}B_{ik}X_{(k,r),(j,s)} =(BX^{\mathrm L}_{rs})_{ij}, \\ {}[X(B\otimes \mathbf1_J)]_{(i,r),(j,s)} & =\sum _{k\in I}X_{(i,r),(k,s)}B_{kj} =(X^{\mathrm L}_{rs}B)_{ij}. \end{align}

Equality of the two products is therefore equivalent to \([B,X^{\mathrm L}_{rs}]=0\) for every \(r,s\in J\).

Theorem 19.10.1.19 Right Kronecker commutation is coefficientwise
#

For \(C\in M_J(\mathbb C)\) and \(X\in M_{I\times J}(\mathbb C)\),

\begin{align} [\mathbf1_I\otimes C,X]=0 \quad \Longleftrightarrow \quad [C,X^{\mathrm R}_{ij}] & =0 \quad \text{for every }i,j\in I. \end{align}
Proof

For \(i,j\in I\) and \(r,s\in J\), the two product entries are

\begin{align} [(\mathbf1_I\otimes C)X]_{(i,r),(j,s)} & =\sum _{t\in J}C_{rt}X_{(i,t),(j,s)} =(CX^{\mathrm R}_{ij})_{rs}, \\ {}[X(\mathbf1_I\otimes C)]_{(i,r),(j,s)} & =\sum _{t\in J}X_{(i,r),(j,t)}C_{ts} =(X^{\mathrm R}_{ij}C)_{rs}. \end{align}

Equality of the two products is therefore equivalent to \([C,X^{\mathrm R}_{ij}]=0\) for every \(i,j\in I\).

For \(B\in M_I(\mathbb C)\),

\begin{align} B\otimes \mathbf1_J\in \mathcal A’ \quad \Longleftrightarrow \quad B & \in \operatorname {Spp}_{\mathrm L}(\mathcal A)’. \end{align}

Equivalently,

\begin{align} \operatorname {Spp}_{\mathrm L}(\mathcal A)’ & =\{ B\in M_I(\mathbb C)\mid B\otimes \mathbf1_J\in \mathcal A’\} . \end{align}
Proof

By Theorem 19.10.1.18, commutation of \(B\otimes \mathbf1_J\) with every \(X\in \mathcal A\) is equivalent to commutation of \(B\) with every left coefficient. The coefficient set is closed under adjoints because \(\mathcal A\) is a star-subalgebra. The universal property of the generated star-subalgebra then extends this commutation relation to all of \(\operatorname {Spp}_{\mathrm L}(\mathcal A)\); the converse follows from generator membership.

For \(C\in M_J(\mathbb C)\),

\begin{align} \mathbf1_I\otimes C\in \mathcal A’ \quad \Longleftrightarrow \quad C & \in \operatorname {Spp}_{\mathrm R}(\mathcal A)’. \end{align}

Equivalently,

\begin{align} \operatorname {Spp}_{\mathrm R}(\mathcal A)’ & =\{ C\in M_J(\mathbb C)\mid \mathbf1_I\otimes C\in \mathcal A’\} . \end{align}
Proof

By Theorem 19.10.1.19, commutation of \(\mathbf1_I\otimes C\) with every \(X\in \mathcal A\) is equivalent to commutation of \(C\) with every right coefficient. Theorem 19.10.1.1 shows that the right coefficient set is closed under adjoints. The universal property of the generated star-subalgebra extends commutation to all of \(\operatorname {Spp}_{\mathrm R}(\mathcal A)\); the converse follows from generator membership.

Let \(I\), \(J\), and \(K\) be finite sets, and let

\begin{align} \mathcal A_1 & \subseteq M_{I\times J}(\mathbb C), & \mathcal A_2 & \subseteq M_{J\times K}(\mathbb C) \end{align}

be star-subalgebras. Suppose that, for every \(X\in \mathcal A_1\) and \(Y\in \mathcal A_2\),

\begin{align} (X\otimes \mathbf1_K)(\mathbf1_I\otimes Y) & =(\mathbf1_I\otimes Y)(X\otimes \mathbf1_K). \label{eq:qca_overlapping_support_lifts} \end{align}

Then the two support algebras on the common factor commute:

\begin{align} RL & =LR & & \text{for every } R\in \operatorname {Spp}_{\mathrm R}(\mathcal A_1),\quad L\in \operatorname {Spp}_{\mathrm L}(\mathcal A_2). \label{eq:qca_overlapping_support_conclusion} \end{align}

This is the full complex matrix-algebra specialization of Schumacher–Werner Lemma sppcomm [ SW04 , lines 1174–1194 ] and GNVW Lemma sppcomm [ GNVW12 , Section 7.1, lines 1221–1246 ] .

Local fixes. For the generated \(C^*\)-algebras, the input must be closed under the involution; see [ con26w ] . In both printed coefficient proofs, the basis element \(e'_\nu \) belongs to the third factor, rather than the second factor as printed; see [ con26o ] .

Scope restriction (full matrix factors). The cited lemmas permit arbitrary finite-dimensional \(C^*\)-algebra factors, whereas the statement above treats full complex matrix algebras. This restriction suffices for the MPU application in [ CPGSV17 , lines 2300–2306 and 2313–2334 ] ; the general-factor statement remains open as recorded in [ con26k ] .

Proof

For \(X\in \mathcal A_1\), \(Y\in \mathcal A_2\), \(i,i'\in I\), and \(k,k'\in K\), applying Theorem  [ LTC26 , “Middle-factor coefficients of commuting overlapping operators” ] to (664) gives

\begin{align} X^{\mathrm R}_{ii'}Y^{\mathrm L}_{kk'} & =Y^{\mathrm L}_{kk'}X^{\mathrm R}_{ii'}. \label{eq:qca_overlapping_support_coefficients} \end{align}

Applying the same identity to \(Y^\dagger \in \mathcal A_2\), with \(k\) and \(k'\) exchanged, also gives commutation with \((Y^{\mathrm L}_{kk'})^\dagger \). Hence every generator of \(\operatorname {Spp}_{\mathrm R}(\mathcal A_1)\) lies in the star-subalgebra commuting with all left coefficients of \(\mathcal A_2\). Passing to the generated star-subalgebra yields

\begin{align} LR & =RL, & L^\dagger R & =RL^\dagger \label{eq:qca_overlapping_support_first_closure} \end{align}

for every \(R\in \operatorname {Spp}_{\mathrm R}(\mathcal A_1)\) and every left coefficient \(L\) of \(\mathcal A_2\).

Since \(\operatorname {Spp}_{\mathrm R}(\mathcal A_1)\) is closed under adjoints, applying the first equality in (667) to \(R^\dagger \) shows that each left coefficient commutes with both \(R\) and \(R^\dagger \). Passing to the star-subalgebra generated by the left coefficients gives (665).

For a finite region \(\Lambda \subset \mathbb Z\) and \(a\in \mathbb Z\), define

\begin{align} \Lambda +a & =\{ i+a\mid i\in \Lambda \} . \end{align}

Thus

\begin{align} j\in \Lambda +a & \iff j-a\in \Lambda , \\ \Lambda +0 & =\Lambda , \\ (\Lambda +a)+b & =\Lambda +(a+b). \end{align}

The map \(\Lambda \mapsto \Lambda +a\) is injective. If \(\Lambda \subseteq \Gamma \), then \(\Lambda +a\subseteq \Gamma +a\).

For finite regions \(\Lambda ,\mathcal N\subset \mathbb Z\), define their sumset by

\begin{align} \Lambda +\mathcal N & =\{ i+a\mid i\in \Lambda ,\ a\in \mathcal N\} . \end{align}

Hence \(j\in \Lambda +\mathcal N\) exactly when \(j=i+a\) for some \(i\in \Lambda \) and \(a\in \mathcal N\). The sumset is monotone in both arguments, either empty argument gives the empty region, and the singleton \(\{ 0\} \) is an identity. A singleton displacement recovers translation,

\begin{align} \Lambda +\{ a\} & =\Lambda +a. \end{align}

Translation commutes with either argument, and sumsets are associative:

\begin{align} (\Lambda +a)+\mathcal N & =(\Lambda +\mathcal N)+a, \\ \Lambda +(\mathcal N+a) & =(\Lambda +\mathcal N)+a, \\ (\Lambda +\mathcal N)+\mathcal M & =\Lambda +(\mathcal N+\mathcal M). \end{align}

These identities support composition of propagation bounds. This is the finite region \(\Lambda +\mathcal N\) in the propagation condition of the QCA appendix of [ CPGSV17 ] .

Translation gives bijections

\begin{align} t_{a,\Lambda } & \colon \Lambda \longrightarrow \Lambda +a, & t_{a,\Lambda }(i) & =i+a, \\ T_{a,\Lambda } & \colon \operatorname {Config}_d(\Lambda ) \longrightarrow \operatorname {Config}_d(\Lambda +a), & (T_{a,\Lambda }x)(j) & =x(j-a). \end{align}

Their inverses translate sites and configurations by \(-a\).

Definition 19.10.1.26 Finite-region translation of local observables

Relabelling configurations defines a star-algebra equivalence

\begin{align} \tau _{a,\Lambda } & \colon \mathcal A_\Lambda \longrightarrow \mathcal A_{\Lambda +a}, \\ \tau _{a,\Lambda }(A)(x,y) & =A(T_{a,\Lambda }^{-1}x,T_{a,\Lambda }^{-1}y). \end{align}

This is the finite-region translation associated with the lattice translation \(i\mapsto i+a\).

Lemma 19.10.1.27 Norm preservation of finite-region translation

Let \(d\) be a nonnegative integer, let \(\Lambda \subset \mathbb Z\) be finite, and let \(a\in \mathbb Z\). For every \(A\in \mathcal A_\Lambda \),

\begin{align} \lVert \tau _{a,\Lambda }(A)\rVert & =\lVert A\rVert . \end{align}

Consequently, \(\tau _{a,\Lambda }\) is an isometry from \(\mathcal A_\Lambda \) onto \(\mathcal A_{\Lambda +a}\).

Proof

Translation bijectively relabels the configuration basis. The resulting star-algebra equivalence between the two finite matrix algebras preserves the operator norm. Applying this identity to \(A-B\) and using linearity gives

\begin{align} \lVert \tau _{a,\Lambda }(A)-\tau _{a,\Lambda }(B)\rVert & =\lVert \tau _{a,\Lambda }(A-B)\rVert =\lVert A-B\rVert , \end{align}

so distances are preserved.

Translation by zero acts trivially on sites, configurations, and finite-region observables. Successive translations by \(a\) and \(b\) agree with translation by \(a+b\). Pointwise, for \(i\in \Lambda \), \(x,y\in \operatorname {Config}_d(\Lambda )\), and \(A\in \mathcal A_\Lambda \),

\begin{align} t_{0,\Lambda }(i) & =i, \\ t_{b,\Lambda +a}(t_{a,\Lambda }(i)) & =t_{a+b,\Lambda }(i), \\ (T_{a,\Lambda }x)(t_{a,\Lambda }(i)) & =x(i), \\ (T_{b,\Lambda +a}T_{a,\Lambda }x) (t_{b,\Lambda +a}(t_{a,\Lambda }(i))) & =(T_{a+b,\Lambda }x)(t_{a+b,\Lambda }(i)), \\ (\tau _{a,\Lambda }(A))(T_{a,\Lambda }x,T_{a,\Lambda }y) & =A(x,y), \\ (\tau _{b,\Lambda +a}\tau _{a,\Lambda }(A)) (T_{b,\Lambda +a}T_{a,\Lambda }x, T_{b,\Lambda +a}T_{a,\Lambda }y) & =(\tau _{a+b,\Lambda }(A)) (T_{a+b,\Lambda }x,T_{a+b,\Lambda }y). \end{align}

These identities express the additive action without identifying the distinct configuration spaces before applying their canonical translation bijections.

Proof

The site identities follow from

\begin{align} (i+a)+b & =i+(a+b). \end{align}

Relabelling a configuration is precomposition by the inverse site translation, so evaluation at the corresponding translated site gives

\begin{align} (T_{a,\Lambda }x)(t_{a,\Lambda }(i)) & =x(i). \end{align}

Applying this identity successively proves the configuration action laws. Finally, the local observable translation relabels both matrix indices by \(T_{a,\Lambda }^{-1}\); evaluating on translated configurations therefore gives \(A(x,y)\), and applying this entrywise identity successively proves the local observable action laws.

If \(\Lambda \subseteq \Gamma \), then for every \(A\in \mathcal A_\Lambda \),

\begin{align} \tau _{a,\Gamma }(\iota _{\Lambda ,\Gamma }(A)) & =\iota _{\Lambda +a,\Gamma +a}(\tau _{a,\Lambda }(A)). \end{align}
Proof

The lattice translation identifies \(\Gamma \setminus \Lambda \) with \((\Gamma +a)\setminus (\Lambda +a)\). Hence relabelling configurations carries the identity on the first complement to the identity on the second complement, and

\begin{align} \tau _{a,\Gamma }(A\otimes \mathbf1_{\Gamma \setminus \Lambda }) & =\tau _{a,\Lambda }(A) \otimes \mathbf1_{(\Gamma +a)\setminus (\Lambda +a)}. \end{align}

The finite-region local algebras form a directed system under the maps \(\iota _{\Lambda ,\Gamma }\). Its algebraic direct limit is the algebraic local algebra

\begin{align} \mathcal A_{\mathrm{loc}} & =\varinjlim _{\Lambda \Subset \mathbb Z}\mathcal A_\Lambda . \end{align}

For every finite region \(\Lambda \), write \(\iota _\Lambda \colon \mathcal A_\Lambda \to \mathcal A_{\mathrm{loc}}\) for the canonical inclusion. Whenever \(\Lambda \subseteq \Gamma \), these inclusions satisfy

\begin{align} \iota _\Gamma \circ \iota _{\Lambda ,\Gamma } & =\iota _\Lambda . \end{align}

This is the union \(\bigcup _{\Lambda \Subset \mathbb Z}\mathcal A_\Lambda \) in the Appendix of [ CPGSV17 ] . No norm completion is taken.

For \(a\in \mathbb Z\) and \(N\in \mathbb N\), let \(I_{a,N}=[a,a+N)\cap \mathbb Z\). The increasing bijection \(\{ 0,\ldots ,N-1\} \simeq I_{a,N}\) identifies the periodic configuration basis with \(\operatorname {Config}_d(I_{a,N})\). In this basis, write \(U_{a}^{(N)}\) for the matrix obtained by simultaneously reindexing the rows and columns of \(U^{(N)}\).

If \(\mathcal U\) is an MPU and \(N{\gt}1\), then \(U_a^{(N)}\) is unitary and defines the star-algebra automorphism

\begin{align} \operatorname {Ad}_{U_a^{(N)}}(B) & =U_a^{(N)}B\, U_a^{(N)\dagger }. \end{align}

The factor order is the one in [ CPGSV17 , lines 2300–2306 ] .

Proof

Simultaneous reindexing preserves both unitarity equations for \(U^{(N)}\). Conjugation by the resulting unitary preserves multiplication, the unit, complex scalars, and the adjoint.

Definition 19.10.1.33 Finite-chain term on a local observable

Let \(\Lambda \subseteq I_{a,N}\) and \(A\in \mathcal A_\Lambda \). Set \(A_L=\iota _{\Lambda ,I_{a,N}}(A)\). For \(N{\gt}1\), the finite-chain term in [ CPGSV17 , lines 2300–2306 ] is the star-algebra map

\begin{align} \mathcal A_\Lambda & \longrightarrow \mathcal A_{\mathrm{loc}}, \\ A & \longmapsto \iota _{I_{a,N}}\! \left(U_a^{(N)}A_LU_a^{(N)\dagger }\right). \end{align}

No stabilization in \(N\) is asserted.

Assume \(d{\gt}0\). If a local observable \(A\in \mathcal A_{\mathrm{loc}}\) is represented by \(A_\Lambda \in \mathcal A_\Lambda \) on a finite region \(\Lambda \), define its operator norm by

\begin{align} \lVert A\rVert & =\lVert A_\Lambda \rVert . \end{align}

This does not depend on the finite region or representative. It is the norm whose topology is completed in the Appendix of [ CPGSV17 ] to obtain the quasi-local algebra.

Proof

If \(A_\Lambda \) and \(A_\Gamma \) represent the same local observable, then the defining relation for the inductive limit gives a finite region \(\Delta \supseteq \Lambda \cup \Gamma \) such that

\begin{align} \iota _{\Lambda ,\Delta }(A_\Lambda ) & =\iota _{\Gamma ,\Delta }(A_\Gamma ). \end{align}

Both inclusions preserve the operator norm, so \(\lVert A_\Lambda \rVert =\lVert A_\Gamma \rVert \).

For \(d{\gt}0\), the local algebra with the operator norm is a normed complex star algebra. No completeness assertion is made.

Proof

Choose finite-region representatives of the local observables and enlarge their regions to a common finite region. The norm axioms, the product inequality, and compatibility with complex scalar multiplication then follow from the corresponding properties of the finite-region matrix algebra.

Lemma 19.10.1.36 C-star identity for local observables
#

For \(d{\gt}0\) and every local observable \(A\),

\begin{align} \lVert A^*A\rVert & =\lVert A\rVert ^2. \end{align}
Proof

Choose a finite region \(\Lambda \) and a representative \(A_\Lambda \in \mathcal A_\Lambda \). Then

\begin{align} \lVert A^*A\rVert & =\lVert A_\Lambda ^*A_\Lambda \rVert =\lVert A_\Lambda \rVert ^2 =\lVert A\rVert ^2. \end{align}
Definition 19.10.1.37 C-star axiom structure on local observables

The normed star algebra of local observables satisfies the C-star axiom. This structure does not include completeness.

Proof

The C-star identity gives the required norm inequality in the reverse direction; submultiplicativity gives the other direction.

Lemma 19.10.1.38 Local observables embed isometrically

For \(d{\gt}0\) and every finite region \(\Lambda \), the canonical map

\begin{align} \iota _\Lambda \colon \mathcal A_\Lambda & \longrightarrow \mathcal A_{\mathrm{loc}} \end{align}

is an operator-norm isometry.

Proof

The norm of \(\iota _\Lambda (A)\) is defined to be the operator norm of its representative \(A\in \mathcal A_\Lambda \). Hence \(\lVert \iota _\Lambda (A)\rVert =\lVert A\rVert \); applying this equality to differences proves preservation of distance.

Let \(B\) be a complex star algebra. Suppose that for every finite region \(\Lambda \) there is a star-algebra homomorphism \(g_\Lambda \colon \mathcal A_\Lambda \to B\). Whenever \(\Lambda \subseteq \Gamma \), suppose that

\begin{align} g_\Gamma \circ \iota _{\Lambda ,\Gamma } & =g_\Lambda . \end{align}

Then there is a unique star-algebra homomorphism \(g\colon \mathcal A_{\mathrm{loc}}\to B\) satisfying \(g\circ \iota _\Lambda =g_\Lambda \) for every finite region \(\Lambda \).

Proof

Define \(g\) on a local observable represented by \(A\in \mathcal A_\Lambda \) as \(g_\Lambda (A)\). Compatibility makes this independent of the representing region. Since each \(g_\Lambda \) preserves addition, multiplication, the unit, scalar multiplication, and the star operation, so does \(g\). Every element of the direct limit has a finite-region representative, which also proves uniqueness.

Definition 19.10.1.40 Support of a local observable
#

A local observable \(A\in \mathcal A_{\mathrm{loc}}\) is supported in the finite region \(\Lambda \) if it lies in the image of \(\iota _\Lambda \colon \mathcal A_\Lambda \to \mathcal A_{\mathrm{loc}}\).

Every local observable is supported in some finite region. If \(A\) is supported in \(\Lambda \) and \(\Lambda \subseteq \Gamma \), then \(A\) is supported in \(\Gamma \). If \(A\) and \(B\) are supported in \(\Lambda \) and \(\Gamma \), respectively, then \(A+B\) and \(AB\) are supported in \(\Lambda \cup \Gamma \). The adjoint \(A^*\) is supported in \(\Lambda \), and both \(0\) and \(1\) are supported in every finite region. For positive on-site dimension, each map \(\iota _\Lambda \) is injective.

Proof

Enlarge representatives by tensoring with the identity on the added sites. Two representatives can therefore be moved to the finite union of their regions, where sums and products are formed. The canonical inclusions are unital star-algebra homomorphisms, so they preserve the adjoint, zero, and the unit. For \(d{\gt}0\), every transition map between finite-region algebras is injective; the canonical map from each member of such a directed system to its direct limit is therefore injective. Finally, every element of an algebraic direct limit has a representative in one member of the directed system.

Theorem 19.10.1.42 Disjointly supported local observables commute

If \(X,Y\in \mathcal A_{\mathrm{loc}}\) are supported in finite regions \(\Lambda \) and \(\Gamma \) with \(\Lambda \cap \Gamma =\varnothing \), then

\begin{align} XY & =YX. \end{align}

In particular, for \(A\in \mathcal A_\Lambda \) and \(B\in \mathcal A_\Gamma \),

\begin{align} \iota _\Lambda (A)\iota _\Gamma (B) & =\iota _\Gamma (B)\iota _\Lambda (A). \end{align}

This is the algebraic-local form of the tensor-factor identity underlying the construction in [ CPGSV17 , lines 2292–2300 ] , not a named theorem of the appendix.

Proof

Choose representatives \(X=\iota _\Lambda (A)\) and \(Y=\iota _\Gamma (B)\). Embed both into \(\mathcal A_{\Lambda \cup \Gamma }\). The preceding finite-region identity gives

\begin{align} XY & =\iota _{\Lambda \cup \Gamma } \bigl(\iota _{\Lambda ,\Lambda \cup \Gamma }(A) \iota _{\Gamma ,\Lambda \cup \Gamma }(B)\bigr) \\ & =\iota _{\Lambda \cup \Gamma } \bigl(\iota _{\Gamma ,\Lambda \cup \Gamma }(B) \iota _{\Lambda ,\Lambda \cup \Gamma }(A)\bigr) =YX. \end{align}

Assume \(d{\gt}0\). A local observable \(X\in \mathcal A_{\mathrm{loc}}\) is supported in a finite region \(\Lambda \) if and only if it commutes with every local observable having a finite support disjoint from \(\Lambda \):

\begin{align} X\in \iota _\Lambda (\mathcal A_\Lambda ) & \iff \forall \Gamma ,\ \forall Y,\quad \bigl(\Gamma \cap \Lambda =\varnothing \ \land Y\in \iota _\Gamma (\mathcal A_\Gamma )\bigr)\Longrightarrow XY=YX. \end{align}

This is an algebraic-local statement. It does not assert the corresponding characterization in the norm completion and is not stated by Schumacher–Werner. Their Theorem 6 gives the generalized Margolus structure, and Corollary 7 says that the inverse of a nearest-neighbor QCA exists and is a nearest-neighbor QCA.

Proof

The forward implication is commutation of disjointly supported observables. Conversely, choose a finite support \(\Delta \) of \(X\) and represent \(X\) by \(A\in \mathcal A_{\Lambda \cup \Delta }\). The hypothesis says that \(A\) commutes with every observable on \((\Lambda \cup \Delta )\setminus \Lambda \). Since \(d{\gt}0\), the canonical map from \(\mathcal A_{\Lambda \cup \Delta }\) into \(\mathcal A_{\mathrm{loc}}\) is injective, so this commutation already holds in the finite-region algebra. The finite relative-commutant theorem gives \(A=\iota _{\Lambda ,\Lambda \cup \Delta }(A_\Lambda )\) for some \(A_\Lambda \in \mathcal A_\Lambda \). Hence \(X=\iota _\Lambda (A_\Lambda )\).

For \(a\in \mathbb Z\), finite-region translations induce a star-algebra homomorphism

\begin{align} \tau _a\colon \mathcal A_{\mathrm{loc}} & \longrightarrow \mathcal A_{\mathrm{loc}} \end{align}

determined on every finite-region representative by

\begin{align} \tau _a(\iota _\Lambda (A)) & =\iota _{\Lambda +a}(\tau _{a,\Lambda }(A)). \end{align}
Proof

Naturality gives, for \(\Lambda \subseteq \Gamma \),

\begin{align} \iota _{\Gamma +a}\! \left(\tau _{a,\Gamma }(\iota _{\Lambda ,\Gamma }(A))\right) & =\iota _{\Gamma +a}\! \left(\iota _{\Lambda +a,\Gamma +a}(\tau _{a,\Lambda }(A))\right) \\ & =\iota _{\Lambda +a}(\tau _{a,\Lambda }(A)). \end{align}

Hence the finite-region maps form a compatible family and induce the stated homomorphism on the algebraic direct limit.

The homomorphisms satisfy

\begin{align} \tau _0 & =\operatorname {id}, \\ \tau _b\circ \tau _a & =\tau _{a+b}. \end{align}
Proof

On a representative \(A\in \mathcal A_\Lambda \), the two sides of the composition law are represented on the equal finite regions \((\Lambda +a)+b\) and \(\Lambda +(a+b)\). The finite-region action identity therefore gives

\begin{align} \tau _b(\tau _a(\iota _\Lambda (A))) & =\iota _{\Lambda +(a+b)} (\tau _{a+b,\Lambda }(A)) =\tau _{a+b}(\iota _\Lambda (A)). \end{align}

The zero law is proved in the same way. Equality on all finite-region representatives determines equality on the algebraic direct limit.

For every \(a\in \mathbb Z\), the homomorphism \(\tau _a\) is a star-algebra automorphism of \(\mathcal A_{\mathrm{loc}}\), and on finite-region representatives

\begin{align} \tau _a(\iota _\Lambda (A)) & =\iota _{\Lambda +a}(\tau _{a,\Lambda }(A)). \end{align}
Proof

The homomorphism action laws give

\begin{align} \tau _{-a}\circ \tau _a & =\tau _0=\operatorname {id}, \\ \tau _a\circ \tau _{-a} & =\tau _0=\operatorname {id}. \end{align}

Thus \(\tau _a\) is bijective with inverse \(\tau _{-a}\). The formula on finite-region representatives is the defining formula for the homomorphism.

The automorphisms form an additive action of \(\mathbb Z\):

\begin{align} \tau _0 & =\operatorname {id}, \\ \tau _b\circ \tau _a & =\tau _{a+b}, \\ \tau _a^{-1} & =\tau _{-a}. \end{align}
Proof

The first two identities are the homomorphism action laws expressed as identities of star-algebra automorphisms. Taking \(b=-a\) in the composition law gives

\begin{align} \tau _{-a}\circ \tau _a & =\tau _0=\operatorname {id}, \\ \tau _a\circ \tau _{-a} & =\tau _0=\operatorname {id}, \end{align}

which is the inverse identity.

A local observable \(A\) is supported in a finite region \(\Lambda \) if and only if \(\tau _a(A)\) is supported in \(\Lambda +a\).

Proof

If \(A=\iota _\Lambda (A_\Lambda )\), then

\begin{align} \tau _a(A) & =\iota _{\Lambda +a}(\tau _{a,\Lambda }(A_\Lambda )), \end{align}

which proves the forward implication. Apply the same argument to \(\tau _a(A)\) with displacement \(-a\), and use \(\tau _{-a}\tau _a=\operatorname {id}\) and \((\Lambda +a)-a=\Lambda \), for the reverse implication.

If \(d{\gt}0\), then every translation preserves the operator norm and distance:

\begin{align} \lVert \tau _a(A)\rVert & =\lVert A\rVert , \\ \lVert \tau _a(A)-\tau _a(B)\rVert & =\lVert A-B\rVert . \end{align}
Proof

Choose \(A_\Lambda \in \mathcal A_\Lambda \) with \(A=\iota _\Lambda (A_\Lambda )\). Finite-region translation is a star-algebra equivalence of matrix algebras, hence preserves the operator norm. Therefore

\begin{align} \lVert \tau _a(A)\rVert & =\lVert \tau _{a,\Lambda }(A_\Lambda )\rVert =\lVert A_\Lambda \rVert =\lVert A\rVert . \end{align}

Applying this equality to \(A-B\) gives distance preservation.

The norm completion of a complex normed star algebra satisfying the C-star identity carries the continuously extended involution and a C-star algebra structure. The canonical map into the completion is a star-algebra homomorphism.

Proof

Extend the isometric involution continuously. The algebraic identities and the C-star identity hold on the dense canonical image and hence on the completion.

Definition 19.10.1.51 Quasi-local observable algebra

For \(d{\gt}0\), the quasi-local observable algebra is the operator-norm completion

\begin{align} \mathcal A & =\overline{\mathcal A_{\mathrm{loc}}}^{\lVert \cdot \rVert }. \end{align}

The involution on \(\mathcal A\) is the unique continuous extension of the involution on \(\mathcal A_{\mathrm{loc}}\). It makes \(\mathcal A\) a complex C-star algebra, and the canonical star-algebra homomorphism

\begin{align} j\colon \mathcal A_{\mathrm{loc}} & \longrightarrow \mathcal A \end{align}

is the completion map.

Proof

The involution is isometric on \(\mathcal A_{\mathrm{loc}}\), hence extends continuously to the completion. The identities \((A+B)^*=A^*+B^*\), \((AB)^*=B^*A^*\), and \((A^*)^*=A\) hold on the dense subalgebra \(j(\mathcal A_{\mathrm{loc}})\) and therefore on \(\mathcal A\). The conjugate-linear identity \((cA)^*=\overline cA^*\) extends by the same argument. Density also extends

\begin{align} \lVert A^*A\rVert & =\lVert A\rVert ^2, \end{align}

so the completed algebra satisfies the C-star identity.

Definition 19.10.1.52 Finite-region quasi-local observable

For every finite region \(\Lambda \), define

\begin{align} j_\Lambda \colon \mathcal A_\Lambda & \longrightarrow \mathcal A, & j_\Lambda & =j\circ \iota _\Lambda . \end{align}

For every finite region \(\Lambda \), the map \(j_\Lambda \) is an injective isometry. If \(\Lambda \subseteq \Gamma \), then

\begin{align} j_\Gamma \circ \iota _{\Lambda ,\Gamma } & =j_\Lambda . \end{align}
Proof

The algebraic inclusion \(\iota _\Lambda \) and the completion map \(j\) both preserve the norm. Their composite is therefore an isometry and hence injective. Compatibility follows from \(\iota _\Gamma \circ \iota _{\Lambda ,\Gamma }=\iota _\Lambda \).

Theorem 19.10.1.54 Density of the canonical completion map

The canonical star-algebra homomorphism from a normed star algebra into its completion has dense range.

Proof

This is the defining density property of the completion.

Algebraic local observables are dense in the quasi-local algebra. Equivalently,

\begin{align} \overline{j(\mathcal A_{\mathrm{loc}})} & =\overline{\bigcup _{\Lambda \Subset \mathbb Z} j_\Lambda (\mathcal A_\Lambda )} =\mathcal A. \end{align}
Proof

Density of \(j(\mathcal A_{\mathrm{loc}})\) is the defining density property of the completion. Every algebraic local observable has a representative in some finite region, so

\begin{align} j(\mathcal A_{\mathrm{loc}}) & =\bigcup _{\Lambda \Subset \mathbb Z}j_\Lambda (\mathcal A_\Lambda ). \end{align}

Taking closures proves the second equality. This theorem concerns the union over all finite regions and makes no claim about the image of a fixed region.

The appendix introduces the quasi-local algebra and then invokes lattice translation in the definition of a QCA and in its covariance condition [ CPGSV17 , lines 2292–2306 ] . The following completion result and translation laws make this background structure explicit; they are not stated as separate results in the appendix.

Definition 19.10.1.56 Extension of star-algebra equivalences to completions

Let \(A\) and \(B\) be complex normed star algebras, and let \(f\colon A\to B\) be a star-algebra equivalence. If \(f\) and \(f^{-1}\) are continuous, then \(f\) extends to a star-algebra equivalence

\begin{align} \widehat f\colon \widehat A & \longrightarrow \widehat B \end{align}

between their norm completions. Its underlying map is the continuous extension of \(f\). Writing \(j_A\) and \(j_B\) for the canonical maps into the completions, the extension satisfies

\begin{align} \widehat f(j_A(x)) & =j_B(f(x)). \end{align}
Proof

Apply the completion functor to \(f\) and \(f^{-1}\). On the dense canonical images their composites satisfy

\begin{align} \widehat{f^{-1}}\bigl(\widehat f(j_A(x))\bigr) & =j_A(x), & \widehat f\bigl(\widehat{f^{-1}}(j_B(y))\bigr) & =j_B(y). \end{align}

Continuity gives the two inverse identities on the completions. Multiplication, addition, scalar multiplication, and the star operation commute with the extension because the corresponding identities hold on the canonical images and both sides are continuous.

For \(d{\gt}0\) and every \(a\in \mathbb Z\), algebraic-local translation extends to a star-algebra automorphism

\begin{align} \tau _a\colon \mathcal A & \longrightarrow \mathcal A \end{align}

of the quasi-local algebra. On the dense algebraic-local subalgebra and on finite-region observables it satisfies

\begin{align} \tau _a(j(A)) & =j(\tau _a(A)), \\ \tau _a(j_\Lambda (A_\Lambda )) & =j_{\Lambda +a}(\tau _{a,\Lambda }(A_\Lambda )). \end{align}
Proof

Algebraic-local translation and its inverse translation by \(-a\) are operator-norm isometries, hence continuous. Extend this equivalence to the norm completion. The extension identity gives

\begin{align} \tau _a(j(A)) & =j(\tau _a(A)). \end{align}

Substituting \(A=\iota _\Lambda (A_\Lambda )\) and using the finite-region translation formula gives the second identity.

For \(d{\gt}0\), the quasi-local automorphisms form an additive action of \(\mathbb Z\):

\begin{align} \tau _0 & =\operatorname {id}, \\ \tau _b\circ \tau _a & =\tau _{a+b}, \\ \tau _a^{-1} & =\tau _{-a}. \end{align}
Proof

On the canonical image of the algebraic-local algebra,

\begin{align} \tau _b\bigl(\tau _a(j(A))\bigr) & =j\bigl(\tau _b(\tau _a(A))\bigr) =j(\tau _{a+b}(A)), \\ \tau _0(j(A)) & =j(A). \end{align}

Continuity extends these identities to the completion. Taking \((a,b)=(a,-a)\) and \((-a,a)\) identifies the inverse of \(\tau _a\) with \(\tau _{-a}\).

Definition 19.10.1.59 Translation covariance

For \(d{\gt}0\), a star-algebra automorphism \(\omega \) of the quasi-local algebra \(\mathcal A\) is translation covariant if

\begin{align} \omega \circ \tau _a=\tau _a\circ \omega \end{align}

for every \(a\in \mathbb Z\). This is the covariance condition in the definition of a one-dimensional QCA [ CPGSV17 , line 2298 ] .

Theorem 19.10.1.60 Pointwise characterization of translation covariance

For \(d{\gt}0\), a star-algebra automorphism \(\omega \) of \(\mathcal A\) is translation covariant if and only if

\begin{align} \omega \bigl(\tau _a(A)\bigr)=\tau _a\bigl(\omega (A)\bigr) \end{align}

for every \(a\in \mathbb Z\) and every \(A\in \mathcal A\).

Proof

Two automorphisms are equal if and only if they agree on every observable. Apply this observation to the two composites in the definition.

Theorem 19.10.1.61 Translation covariance from unit translation

For \(d{\gt}0\), a star-algebra automorphism \(\omega \) of \(\mathcal A\) commutes with every integer translation if and only if it commutes with unit translation:

\begin{align} \omega \circ \tau _1=\tau _1\circ \omega . \end{align}
Proof

The forward implication follows by taking the translation step to be \(1\). Conversely, use two-sided integer induction. The zero case follows from \(\tau _0=\operatorname {id}\). For the successor step, the additive action law gives \(\tau _{i+1}=\tau _1\circ \tau _i\), and

\begin{align} \omega \circ \tau _{i+1} & =\omega \circ \tau _1\circ \tau _i =\tau _1\circ \omega \circ \tau _i \\ & =\tau _1\circ \tau _i\circ \omega =\tau _{i+1}\circ \omega . \end{align}

For the predecessor step, \(\tau _{-1}=\tau _1^{-1}\) transfers commutation to \(\tau _{-1}\), while \(\tau _{i-1}=\tau _{-1}\circ \tau _i\), so

\begin{align} \omega \circ \tau _{i-1} & =\omega \circ \tau _{-1}\circ \tau _i =\tau _{-1}\circ \omega \circ \tau _i \\ & =\tau _{-1}\circ \tau _i\circ \omega =\tau _{i-1}\circ \omega . \end{align}
Theorem 19.10.1.62 Identity and composition of translation-covariant automorphisms

For \(d{\gt}0\), the identity automorphism of \(\mathcal A\) is translation covariant. If \(\omega \) and \(\eta \) are translation covariant, then so is the composite

\begin{align} A\longmapsto \eta \bigl(\omega (A)\bigr), \end{align}

which first applies \(\omega \) and then \(\eta \).

Proof

The identity commutes pointwise with every translation. For the composite, translation covariance gives

\begin{align} \eta \bigl(\omega (\tau _a(A))\bigr) & =\eta \bigl(\tau _a(\omega (A))\bigr) =\tau _a\bigl(\eta (\omega (A))\bigr). \end{align}

For \(d{\gt}0\), every quasi-local translation preserves the operator norm and distance:

\begin{align} \lVert \tau _a(A)\rVert & =\lVert A\rVert , \\ \lVert \tau _a(A)-\tau _a(B)\rVert & =\lVert A-B\rVert . \end{align}
Proof

A star-algebra equivalence between complex C-star algebras preserves the C-star norm. Since \(\tau _a\) is linear,

\begin{align} \lVert \tau _a(A)-\tau _a(B)\rVert & =\lVert \tau _a(A-B)\rVert =\lVert A-B\rVert . \end{align}

For \(d{\gt}0\), a quasi-local observable \(x\in \mathcal A\) is supported in a finite region \(\Lambda \subset \mathbb Z\) if it belongs to the image of the canonical embedding \(j_\Lambda \colon \mathcal A_\Lambda \to \mathcal A\). Equivalently,

\begin{align} x\text{ is supported in }\Lambda & \iff x\in j_\Lambda (\mathcal A_\Lambda ) \\ & \iff \text{there exists }A\in \mathcal A_\Lambda \text{ such that }j_\Lambda (A)=x. \end{align}

In particular, \(j_\Lambda (A)\) is supported in \(\Lambda \) for every \(A\in \mathcal A_\Lambda \).

Proof

These assertions follow directly from membership in the range of \(j_\Lambda \).

Let \(d{\gt}0\). If \(x\in \mathcal A\) is supported in \(\Lambda \) and \(\Lambda \subseteq \Gamma \), then \(x\) is supported in \(\Gamma \).

Proof

Write \(x=j_\Lambda (A)\). Enlarge \(A\) to \(\mathcal A_\Gamma \) by tensoring with the identity on \(\Gamma \setminus \Lambda \). Compatibility of the canonical embeddings identifies the image of this enlargement with \(x\).

Theorem 19.10.1.66 Disjointly supported quasi-local observables commute

Let \(d{\gt}0\). If \(x,y\in \mathcal A\) are supported in finite regions \(\Lambda \) and \(\Gamma \) with \(\Lambda \cap \Gamma =\varnothing \), then

\begin{align} xy & =yx. \end{align}

Equivalently, for \(A\in \mathcal A_\Lambda \) and \(B\in \mathcal A_\Gamma \),

\begin{align} j_\Lambda (A)j_\Gamma (B) & =j_\Gamma (B)j_\Lambda (A). \end{align}

This is finite-support infrastructure for the quasi-local algebra in [ CPGSV17 , lines 2292–2300 ] ; it is not stated there as a separate theorem.

Proof

The completion map is a star-algebra homomorphism. Applying it to the algebraic-local identity gives

\begin{align} j(XY)=j(YX), \end{align}

and finite-region support provides representatives \(x=j_\Lambda (A)\) and \(y=j_\Gamma (B)\).

The following finite-dimensional averaging fact is standard and is not a claim from [ CPGSV17 ] . It will be used to characterize exact local support by commutation with observables on the complementary tensor factor.

Definition 19.10.1.67 Complementary-factor Weyl twirl

Let \(S\) and \(C\) be finite sets, let \(n=|C|{\gt}0\), and choose an equivalence \(e\colon C\simeq \mathbb Z/n\mathbb Z\) and a primitive \(n\)-th root of unity \(\zeta \). If \(W_{a,b}\) is the corresponding reindexed Weyl operator on \(\mathbb C^C\), define

\begin{align} \mathcal E_C(M) & = \frac{1}{n^2}\sum _{a,b\in \mathbb Z/n\mathbb Z} (\mathbf1_S\otimes W_{a,b})M (\mathbf1_S\otimes W_{a,b})^\dagger . \end{align}

For every \(M\in \operatorname {Mat}_{S\times C}(\mathbb C)\),

\begin{align} \mathcal E_C(M) & = \operatorname {tr}_C(M)\otimes \frac{\mathbf1_C}{n} \\ & = \left(\frac{1}{n}\operatorname {tr}_C(M)\right)\otimes \mathbf1_C. \end{align}
Proof

Split \(M\) into \(C\times C\) blocks indexed by \(S\). The Weyl average sends each block \(B\) to \((\operatorname {tr}B/n)\mathbf1_C\), so its block trace is the corresponding entry of \(\operatorname {tr}_C(M)\).

Lemma 19.10.1.69 Norm estimate for the complementary-factor twirl

Let \(S\) be a nonempty finite set. The average is contractive in the operator norm. Moreover, if every complementary Weyl conjugation fixes \(X\), then

\begin{align} \lVert \mathcal E_C(M)-X\rVert & \le \lVert M-X\rVert . \end{align}
Proof

Unitary conjugation preserves the operator norm, and the coefficients of the finite average sum to one. Apply this contraction to \(M-X\).

Let \(d{\gt}0\), let \(\Lambda \subset \mathbb Z\) be finite, and let \(x\in \mathcal A\). Then

\begin{align} x\in \mathcal A_\Lambda \quad \Longleftrightarrow \quad [x,y]=0 \text{ for every }y\in \mathcal A_\Gamma \text{ and every finite }\Gamma \cap \Lambda =\varnothing . \end{align}

This is standard infrastructure for the UHF algebra used in TNLean. It is not a theorem stated in [ CPGSV17 ] or by Schumacher–Werner, quant-ph/0405174.

Proof

The forward implication follows from commutation of disjoint tensor factors. Conversely, approximate \(x\) by \(A\in \mathcal A_\Delta \), where \(\Lambda \subseteq \Delta \). Split \(\Delta =\Lambda \sqcup (\Delta \setminus \Lambda )\) and average \(A\) over Weyl unitaries on the second factor. Every such conjugation fixes \(x\) by the commutation hypothesis, so unitary invariance and convexity give

\begin{align} \lVert \mathcal E_{\Delta \setminus \Lambda }(A)-x\rVert & \le \lVert A-x\rVert . \end{align}

The Weyl identity writes the average as \(B\otimes \mathbf1\) for some \(B\in \mathcal A_\Lambda \). Thus elements of \(\mathcal A_\Lambda \) approximate \(x\) arbitrarily well. The embedding of \(\mathcal A_\Lambda \) is isometric, hence has closed range, and therefore \(x\in \mathcal A_\Lambda \).

The following theorem gives the scalar center of the homogeneous UHF algebra. GNVW invoke this conclusion for the quasi-local algebra at [ GNVW12 , lines 1276–1282 ] , specifically line 1279, but do not prove it there. The homogeneous construction in [ CPGSV17 , lines 2292–2298 ] likewise defines the norm completion without stating this conclusion separately. Schumacher–Werner give the corresponding homogeneous norm-completed construction in [ SW04 , lines 263–285 ] , again without a separate scalar-center theorem.

Let \(d{\gt}0\) and let \(z\in \mathcal A\). If

\begin{align} \forall \Gamma \subset \mathbb Z\text{ finite},\quad \forall A\in \mathcal A_\Gamma ,\qquad z j_\Gamma (A) & =j_\Gamma (A)z, \end{align}

then there is a unique \(c\in \mathbb C\) such that

\begin{align} z & =c\mathbf1. \end{align}

Consequently,

\begin{align} \{ z\in \mathcal A\mid za=az\text{ for every }a\in \mathcal A\} & =\mathbb C\mathbf1. \end{align}

This statement concerns the fixed on-site dimension \(d\); it does not assert the site-dependent extension used in GNVW.

Proof

Apply Theorem 19.10.1.70 with \(\Lambda =\varnothing \). The commutation hypothesis gives \(z\in j_\varnothing (\mathcal A_\varnothing )\), so there is an \(A\in \mathcal A_\varnothing \) with \(z=j_\varnothing (A)\). Let \(e\colon \mathbb C\xrightarrow {\ \sim \ }\mathcal A_\varnothing \) be the canonical isomorphism and write \(A=e(c)\). If also \(z=j_\varnothing (e(c'))\), then injectivity of \(j_\varnothing \) and of \(e\) give

\begin{align} j_\varnothing (e(c))=j_\varnothing (e(c’)) & \Longrightarrow e(c)=e(c’) \Longrightarrow c=c’. \end{align}

Thus \(c\) is the unique scalar for which \(z=c\mathbf1\).

Write \(Z(\mathcal A)\) for the center. Since every \(j_\Gamma (B)\) belongs to \(\mathcal A\), the two inclusions are

\begin{align} z\in Z(\mathcal A) & \Longrightarrow \bigl(\forall \Gamma \subset \mathbb Z\text{ finite}, \forall B\in \mathcal A_\Gamma , z j_\Gamma (B)=j_\Gamma (B)z\bigr) \Longrightarrow z\in \mathbb C\mathbf1, \\ z=c\mathbf1\quad (c\in \mathbb C) & \Longrightarrow \bigl(\forall a\in \mathcal A, (c\mathbf1)a=a(c\mathbf1)\bigr) \Longrightarrow z\in Z(\mathcal A). \end{align}

Hence \(Z(\mathcal A)=\mathbb C\mathbf1\).

The appendix requires that “there exists a finite subset \(\mathcal N\subset \mathbb Z\)” such that, for every finite \(\Lambda \subset \mathbb Z\),

\begin{align} \omega (\mathcal A_\Lambda ) & \subset \mathcal A_{\Lambda +\mathcal N}. \end{align}

This is the propagation condition stated at line 2298 of [ CPGSV17 ] . The definitions below isolate this forward inclusion. They do not impose translation covariance or a propagation bound for \(\omega ^{-1}\).

Definition 19.10.1.72 Propagation within a fixed neighborhood

Let \(d{\gt}0\), let \(\omega \colon \mathcal A\to \mathcal A\) be a star-algebra automorphism, and let \(\mathcal N\subset \mathbb Z\) be finite. The automorphism \(\omega \) propagates within \(\mathcal N\) if, for every finite region \(\Lambda \) and every \(x\in \mathcal A\),

\begin{align} x\text{ supported in }\Lambda & \implies \omega (x)\text{ supported in }\Lambda +\mathcal N. \end{align}
Definition 19.10.1.73 Finite forward propagation
#

Let \(d{\gt}0\) and let \(\omega \colon \mathcal A\to \mathcal A\) be a star-algebra automorphism. The automorphism \(\omega \) has finite forward propagation if there exists a finite set \(\mathcal N\subset \mathbb Z\) such that \(\omega \) propagates within \(\mathcal N\).

Let \(d{\gt}0\), let \(\omega \colon \mathcal A\to \mathcal A\) be a star-algebra automorphism, and let \(\mathcal N\subset \mathbb Z\) be finite. The following conditions are equivalent:

\begin{align} & \omega \text{ propagates within }\mathcal N; \tag {a} \\ & \forall \Lambda \Subset \mathbb Z,\ \forall A\in \mathcal A_\Lambda , \quad \omega (j_\Lambda (A))\text{ is supported in } \Lambda +\mathcal N; \tag {b} \\ & \forall \Lambda \Subset \mathbb Z,\quad \omega \bigl(j_\Lambda (\mathcal A_\Lambda )\bigr) \subseteq j_{\Lambda +\mathcal N} (\mathcal A_{\Lambda +\mathcal N}); \tag {c} \\ & \forall \Lambda \Subset \mathbb Z,\ \forall A\in \mathcal A_\Lambda , \ \exists B\in \mathcal A_{\Lambda +\mathcal N},\quad \omega (j_\Lambda (A))=j_{\Lambda +\mathcal N}(B). \tag {d} \end{align}

Thus condition (c) is precisely the range-inclusion formula from line 2298 of [ CPGSV17 ] , expressed inside the quasi-local algebra, and condition (d) is its finite-region witness form.

Proof

Every observable supported in \(\Lambda \) has the form \(j_\Lambda (A)\), so (a) and (b) are equivalent. Replacing support by membership in the range of the canonical embedding gives (c). Finally, membership in \(j_{\Lambda +\mathcal N}(\mathcal A_{\Lambda +\mathcal N})\) is equivalent to the existence of a representative \(B\) in that finite-region algebra, which gives (d).

Let \(d{\gt}0\) and let \(\omega \colon \mathcal A\to \mathcal A\) be a star-algebra automorphism. If \(\mathcal N\subseteq \mathcal M\) and \(\omega \) propagates within \(\mathcal N\), then \(\omega \) propagates within \(\mathcal M\). More explicitly, for every finite region \(\Lambda \),

\begin{align} \Lambda +\mathcal N & \subseteq \Lambda +\mathcal M, \\ x\text{ supported in }\Lambda & \implies \omega (x)\text{ supported in }\Lambda +\mathcal N \implies \omega (x)\text{ supported in }\Lambda +\mathcal M. \end{align}

Consequently, if \(\omega \) has finite forward propagation, then for every prescribed finite set \(\mathcal N\subset \mathbb Z\) there is a finite set \(\mathcal M\subset \mathbb Z\) such that

\begin{align} \mathcal N & \subseteq \mathcal M, \\ & \forall \Lambda \Subset \mathbb Z,\quad \omega (\mathcal A_\Lambda )\subseteq \mathcal A_{\Lambda +\mathcal M}. \end{align}

This is the neighborhood-enlargement consequence of the propagation condition in Appendix line 2298 of [ CPGSV17 ] .

Proof

Monotonicity of the sumset in its second argument gives \(\Lambda +\mathcal N\subseteq \Lambda +\mathcal M\). Monotonicity of quasi-local support then enlarges the support of \(\omega (x)\) from \(\Lambda +\mathcal N\) to \(\Lambda +\mathcal M\). For the final assertion, choose a finite propagation neighborhood \(\mathcal K\) and take \(\mathcal M=\mathcal K\cup \mathcal N\).

Let \(d{\gt}0\), let \(\omega ,\eta \colon \mathcal A\to \mathcal A\) be star-algebra automorphisms, and let \(\mathcal N,\mathcal M\subset \mathbb Z\) be finite. Suppose that \(\omega \) propagates within \(\mathcal N\) and \(\eta \) propagates within \(\mathcal M\). The composite \(\eta \circ \omega \) first applies \(\omega \) and then \(\eta \), so

\begin{align} (\eta \circ \omega )(x) & =\eta (\omega (x)), \\ x\text{ supported in }\Lambda & \implies \omega (x)\text{ supported in }\Lambda +\mathcal N \\ & \implies \eta (\omega (x))\text{ supported in } (\Lambda +\mathcal N)+\mathcal M =\Lambda +(\mathcal N+\mathcal M). \end{align}

Hence \(\eta \circ \omega \) propagates within \(\mathcal N+\mathcal M\). In particular, if \(\omega \) and \(\eta \) have finite forward propagation, then \(\eta \circ \omega \) has finite forward propagation. This is the composition closure of the finite-neighborhood condition in Appendix line 2298 of [ CPGSV17 ] .

Proof

Apply the two propagation bounds successively and use associativity of finite region sumsets. For the existential assertion, choose one finite propagation neighborhood for each automorphism and use their sumset.

Every finite set \(\mathcal N\subset \mathbb Z\) is contained in \([-R,R]\cap \mathbb Z\) for some \(R\in \mathbb N\). This includes \(\mathcal N=\varnothing \), for which one may take \(R=0\). If \(\omega \) propagates within a fixed set \(\mathcal N\), then \(R\) may be chosen so that

\begin{align} \mathcal N & \subseteq [-R,R]\cap \mathbb Z, \\ \omega & \text{ propagates within }[-R,R]\cap \mathbb Z. \end{align}

Consequently, finite forward propagation admits a symmetric-interval witness:

\begin{align} \exists R\in \mathbb N,\qquad \omega & \text{ propagates within }[-R,R]\cap \mathbb Z. \end{align}
Proof

Let \(R\) be the maximum of \(|n|\) over \(n\in \mathcal N\), with value \(0\) when \(\mathcal N\) is empty. Then \(-R\leq n\leq R\) for every \(n\in \mathcal N\), which gives \(\mathcal N\subseteq [-R,R]\cap \mathbb Z\). Enlargement of propagation neighborhoods gives the fixed-neighborhood assertion. Finally, choose a finite propagation neighborhood and apply this enlargement to its symmetric interval.

Definition 19.10.1.78 Two-sided finite-propagation predicates

Let \(d{\gt}0\), and let \(\omega \) be a star-algebra automorphism of \(\mathcal A\). A finite set \(\mathcal N\subset \mathbb Z\) is a two-sided propagation neighborhood when

\begin{align} \omega (\mathcal A_\Lambda ) & \subseteq \mathcal A_{\Lambda +\mathcal N} \end{align}

and

\begin{align} \omega ^{-1}(\mathcal A_\Lambda ) & \subseteq \mathcal A_{\Lambda +\mathcal N} \end{align}

for every finite \(\Lambda \subset \mathbb Z\). The automorphism has two-sided finite propagation when \(\omega \) and \(\omega ^{-1}\) each have a finite forward propagation neighborhood; the two neighborhoods need not initially agree.

These are derived convenience predicates. The definition in [ CPGSV17 , line 2298 ] and Schumacher–Werner, quant-ph/0405174, Definition 1, both define locality by the forward inclusion for \(\omega \) alone. In particular, these predicates do not assert that forward locality implies locality of the inverse.

For a finite set \(\mathcal N\subset \mathbb Z\), write

\begin{align} -\mathcal N & =\{ -a\mid a\in \mathcal N\} . \end{align}

Reflection fixes every symmetric interval:

\begin{align} -\bigl([-R,R]\cap \mathbb Z\bigr) & =[-R,R]\cap \mathbb Z. \end{align}

If \(\mathcal N\subseteq \mathcal M\), then every common two-sided bound by \(\mathcal N\) is also a common bound by \(\mathcal M\). More generally, if \(\omega \) propagates within \(\mathcal N\) and \(\omega ^{-1}\) propagates within \(-\mathcal N\), then both propagate within

\begin{align} \mathcal N\cup (-\mathcal N). \end{align}

Finally, any two-sided finite-propagation pair admits a common finite neighborhood and therefore a common symmetric-interval neighborhood. Thus there is an \(R\in \mathbb N\) such that

\begin{align} \omega (\mathcal A_\Lambda ) & \subseteq \mathcal A_{\Lambda +([-R,R]\cap \mathbb Z)} \end{align}

and

\begin{align} \omega ^{-1}(\mathcal A_\Lambda ) & \subseteq \mathcal A_{\Lambda +([-R,R]\cap \mathbb Z)} \end{align}

for every finite \(\Lambda \subset \mathbb Z\).

Proof

Negation reverses the two endpoint inequalities, so it preserves \([-R,R]\cap \mathbb Z\). Enlarge each supplied neighborhood using monotonicity. For reflected bounds, both inclusions into \(\mathcal N\cup (-\mathcal N)\) are immediate. For arbitrary finite forward and inverse neighborhoods, first take their union and then enlarge that common neighborhood to a symmetric interval.

Lemma 19.10.1.80 Disjointness for reflected neighborhoods

Let \(\Gamma ,\Lambda ,\mathcal N\subset \mathbb Z\) be finite. Then

\begin{align} \Gamma \cap (\Lambda +(-\mathcal N))=\varnothing \quad \Longleftrightarrow \quad (\Gamma +\mathcal N)\cap \Lambda =\varnothing . \end{align}
Proof

If \(g+n=\lambda \) for \(g\in \Gamma \), \(n\in \mathcal N\), and \(\lambda \in \Lambda \), then \(g=\lambda +(-n)\in \Lambda +(-\mathcal N)\), proving the forward implication. Conversely, if \(g\in \Gamma \) and \(g=\lambda +(-n)\) with \(\lambda \in \Lambda \) and \(n\in \mathcal N\), then \(\lambda =g+n\in \Gamma +\mathcal N\).

Let \(d{\gt}0\), let \(\omega \) be a star-algebra automorphism of \(\mathcal A\), and let \(\mathcal N\subset \mathbb Z\) be finite. If \(\omega \) propagates within \(\mathcal N\), then \(\omega ^{-1}\) propagates within \(-\mathcal N\). Therefore finite forward propagation of \(\omega \) implies finite forward propagation of \(\omega ^{-1}\). Moreover, if \(\omega \) is translation covariant, then so is \(\omega ^{-1}\).

This is a consequence of the forward-only locality definition, not part of that definition. Schumacher–Werner, quant-ph/0405174, derive inverse locality through Lemma 5, Theorem 6, and Corollary 7 using the generalized Margolus structure. The proof here instead uses TNLean’s exact outside-commutant infrastructure.

Proof

Let \(x\in \mathcal A_\Lambda \) and let \(y\in \mathcal A_\Gamma \), where \(\Gamma \cap (\Lambda +(-\mathcal N))=\varnothing \). Forward locality gives \(\omega (y)\in \mathcal A_{\Gamma +\mathcal N}\). The regions \(\Gamma +\mathcal N\) and \(\Lambda \) are disjoint, so \(x\) commutes with \(\omega (y)\). Applying \(\omega ^{-1}\) shows that \(\omega ^{-1}(x)\) commutes with \(y\). The exact outside-commutant characterization now gives

\begin{align} \omega ^{-1}(x) & \in \mathcal A_{\Lambda +(-\mathcal N)}. \end{align}

Choosing the reflected finite neighborhood proves the existential statement. For each lattice translation \(\tau _a\), multiplying the covariance identity on the left and right by \(\omega ^{-1}\) gives

\begin{align} \omega \circ \tau _a=\tau _a\circ \omega & \implies \omega ^{-1}\circ \tau _a=\tau _a\circ \omega ^{-1}. \end{align}

Let \(d{\gt}0\) and let \(\omega \) be a star-algebra automorphism of \(\mathcal A\). Then

\begin{align} \omega \text{ has finite forward propagation} & \iff \omega \text{ has two-sided finite propagation}. \end{align}

More precisely, a forward bound by a finite set \(\mathcal N\) gives an inverse bound by \(-\mathcal N\), so both maps propagate within the enlarged common neighborhood

\begin{align} \mathcal N\cup (-\mathcal N). \end{align}

This statement remains valid when \(\mathcal N=\varnothing \). For an arbitrary finite forward-propagation witness, there is therefore a common finite neighborhood for \(\omega \) and \(\omega ^{-1}\), and it may be enlarged to a symmetric interval \([-R,R]\cap \mathbb Z\).

This equivalence is a derived convenience theorem. It does not change the forward-only locality definition at line 2298 of [ CPGSV17 ] or in Schumacher–Werner, Definition 1.

Proof

Derived inverse locality sends a forward bound by \(\mathcal N\) to an inverse bound by \(-\mathcal N\). Enlarging both bounds to \(\mathcal N\cup (-\mathcal N)\) gives the fixed-neighborhood statement. At the existential level, forward finite propagation gives finite propagation of the inverse, hence the two-sided predicate. The converse is its first conjunct. Finally, apply the common-neighborhood and symmetric-interval enlargement results to the resulting two-sided pair.

Definition 19.10.1.83 One-dimensional quantum cellular automaton

Let \(d{\gt}0\) and let \(\omega \colon \mathcal A\to \mathcal A\) be a star-algebra automorphism. The automorphism \(\omega \) is a one-dimensional quantum cellular automaton if it is translation covariant and has finite forward propagation. Thus there exists a finite set \(\mathcal N\subset \mathbb Z\) such that, for every finite \(\Lambda \subset \mathbb Z\),

\begin{align} \omega (\mathcal A_\Lambda ) & \subseteq \mathcal A_{\Lambda +\mathcal N}. \end{align}

This is the definition at line 2298 of [ CPGSV17 ] . It imposes only the displayed forward propagation condition and makes no locality claim for \(\omega ^{-1}\).

For \(d{\gt}0\), a star-algebra automorphism \(\omega \) of \(\mathcal A\) is a quantum cellular automaton if and only if it is translation covariant and has finite forward propagation. These two properties therefore construct an automaton, and every quantum cellular automaton satisfies each of them.

Proof

This is the conjunction defining a quantum cellular automaton, together with its introduction and elimination rules.

For \(d{\gt}0\), the identity automorphism of \(\mathcal A\) is a quantum cellular automaton.

Proof

The identity is translation covariant. For finite forward propagation, choose the neighborhood \(\{ 0\} \). An observable supported in \(\Lambda \) remains unchanged, and \(\Lambda +\{ 0\} =\Lambda \).

Theorem 19.10.1.86 Composition of quantum cellular automata

Let \(d{\gt}0\). If \(\omega \) and \(\eta \) are quantum cellular automata on \(\mathcal A\), then \(\eta \circ \omega \) is a quantum cellular automaton. This composite first applies \(\omega \) and then \(\eta \). No propagation condition for either inverse is required.

Proof

Both defining properties are closed under composition in the stated order. Translation covariance of \(\omega \) and \(\eta \) gives translation covariance of \(\eta \circ \omega \). If \(\mathcal N\) and \(\mathcal M\) are forward propagation neighborhoods for \(\omega \) and \(\eta \), respectively, then \(\mathcal N+\mathcal M\) is a forward propagation neighborhood for \(\eta \circ \omega \). Hence the composite is a quantum cellular automaton.

Theorem 19.10.1.87 Inverse quantum cellular automaton

Let \(d{\gt}0\). If \(\omega \) is a quantum cellular automaton on \(\mathcal A\), then \(\omega ^{-1}\) is also a quantum cellular automaton.

Proof

Translation covariance and finite forward propagation are both preserved by inversion. These are the two defining properties of a quantum cellular automaton.

Definition 19.10.1.88 Expansion of a blocked region

Let \(L{\gt}0\) be a block length and let \(\Lambda \subset \mathbb Z\) be finite. The expansion of \(\Lambda \) is

\begin{align} \widehat\Lambda _L & =\{ Lx+r:x\in \Lambda ,\ 0\leq r{\lt}L\} . \end{align}

Every \(z\in \widehat\Lambda _L\) has unique block coordinates \((x,r)\) with \(x\in \Lambda \) and \(0\leq r{\lt}L\). Increasing \(r\) gives the consecutive-site order within the block. This is the finite-region site grouping used in [ CPGSV17 , lines 2308 and 2313–2320 ] ; it does not assert the QCA-to-MPU converse stated at line 2308.

An integer \(z\) belongs to \(\widehat\Lambda _L\) exactly when its quotient by \(L\) belongs to \(\Lambda \). The inverse block-coordinate map sends \((x,r)\) to \(Lx+r\), and \(\Lambda \subseteq \Gamma \) implies \(\widehat\Lambda _L\subseteq \widehat\Gamma _L\).

Proof

Apply Euclidean division by the positive integer \(L\). Its quotient and residue are unique, the residue lies in \(\{ 0,\ldots ,L-1\} \), and taking the inverse coordinates gives \(Lx+r\). Monotonicity follows by retaining the same quotient and residue.

Definition 19.10.1.90 Blocking equivalence for finite configurations
#

For on-site dimension \(d\), configurations of \(d^L\)-level blocked spins on \(\Lambda \) are equivalent to configurations of \(d\)-level spins on \(\widehat\Lambda _L\). The blocked label at \(x\) is decoded into its ordered length-\(L\) word, and its \(r\)th letter is placed at \(Lx+r\).

Evaluation at \(Lx+r\) reads the \(r\)th letter of the blocked spin at \(x\). Conversely, encoding collects these letters in increasing residue order. At the configuration level, decoding commutes with restriction from \(\Gamma \) to \(\Lambda \subseteq \Gamma \).

Proof

The two evaluation formulas are the inverse laws for the ordered word encoding. Restriction preserves the quotient block coordinate and residue, so both sides have the same value at every expanded site.

Definition 19.10.1.92 Finite local-algebra blocking equivalence

Reindexing matrix rows and columns by the configuration blocking equivalence gives a star-algebra equivalence

\begin{align} \mathcal A^{(d^L)}_\Lambda & \simeq \mathcal A^{(d)}_{\widehat\Lambda _L}. \end{align}

This is the finite-region equivalence underlying the algebraic-local blocking construction below. No quasi-local completion, compatibility with translations, propagation statement, circuit representation, or transport of quantum cellular automata is asserted here.

Matrix entries are transported by applying the inverse configuration equivalence to both indices. The blocking equivalence preserves the operator norm and is an isometry.

Proof

The entry formula is matrix reindexing. A star-algebra equivalence between finite matrix algebras preserves the C-star norm and hence distances.

Definition 19.10.1.94 Finite block hull
#

For \(L{\gt}0\) and a finite original region \(\Delta \subset \mathbb Z\), define its block hull by

\begin{align} \operatorname {Hull}_L(\Delta ) & =\{ \lfloor z/L\rfloor :z\in \Delta \} . \end{align}

This is the finite set of length-\(L\) blocks that meet \(\Delta \).

Let \(L{\gt}0\). Every finite original region is contained in the expansion of its block hull,

\begin{align} \Delta \subseteq \widehat{\operatorname {Hull}_L(\Delta )}_L. \end{align}

If \(\Delta \subseteq \Theta \), then \(\operatorname {Hull}_L(\Delta )\subseteq \operatorname {Hull}_L(\Theta )\). Expansion and block hull are exact in the blocked direction:

\begin{align} \operatorname {Hull}_L(\widehat\Lambda _L) & =\Lambda . \end{align}
Proof

The quotient coordinate of each \(z\in \Delta \) belongs to the image defining the hull, which proves containment after expansion. Monotonicity follows by retaining the same witness \(z\) in the larger region. Finally, every expanded site has quotient in \(\Lambda \), while for each \(x\in \Lambda \) the site \(Lx\) belongs to \(\widehat\Lambda _L\) and has quotient \(x\). Hence \(\operatorname {Hull}_L(\widehat\Lambda _L)=\Lambda \).

Let \(L{\gt}0\). For finite blocked regions \(\Lambda \subseteq \Gamma \), the square formed by finite blocking and the two canonical local inclusions commutes:

\begin{align} \mathcal B_{L,\Gamma }\, \iota _{\Lambda ,\Gamma } & =\iota _{\widehat\Lambda _L,\widehat\Gamma _L}\, \mathcal B_{L,\Lambda }. \end{align}
Proof

For configurations \(x,y\) on \(\widehat\Gamma _L\), let \(\widetilde x\) and \(\widetilde y\) be their block encodings on \(\Gamma \). The two matrix entries in the commuting square are

\begin{align} \bigl[\mathcal B_{L,\Gamma } (\iota _{\Lambda ,\Gamma }(A))\bigr]_{x,y} & =A_{\widetilde x|_\Lambda ,\widetilde y|_\Lambda } \mathbf1_{\{ \widetilde x|_{\Gamma \setminus \Lambda } =\widetilde y|_{\Gamma \setminus \Lambda }\} }, \\ \bigl[\iota _{\widehat\Lambda _L,\widehat\Gamma _L} (\mathcal B_{L,\Lambda }(A))\bigr]_{x,y} & =A_{\widetilde x|_\Lambda ,\widetilde y|_\Lambda } \mathbf1_{\{ x|_{\widehat\Gamma _L\setminus \widehat\Lambda _L} =y|_{\widehat\Gamma _L\setminus \widehat\Lambda _L}\} }. \end{align}

Restriction commutes with block encoding, and decoding identifies \(\Gamma \setminus \Lambda \) with the original sites in \(\widehat\Gamma _L\setminus \widehat\Lambda _L\). Hence the two indicator conditions are equivalent and the entries agree.

Let \(L{\gt}0\). The compatible finite blocking maps induce a star-algebra homomorphism

\begin{align} \mathcal B_L:\mathcal A_{\mathrm{loc}}^{(d^L)} \longrightarrow \mathcal A_{\mathrm{loc}}^{(d)}. \end{align}

In the reverse direction there is a star-algebra homomorphism

\begin{align} \mathcal U_L:\mathcal A_{\mathrm{loc}}^{(d)} \longrightarrow \mathcal A_{\mathrm{loc}}^{(d^L)}. \end{align}

A representative on an arbitrary original region \(\Delta \) is first included into \(\widehat{\operatorname {Hull}_L(\Delta )}_L\) and then finite blocking is inverted. Thus the reverse construction does not require \(\Delta \) itself to be an expanded region.

Proof

For \(\Lambda \subseteq \Gamma \), the finite commuting square gives

\begin{align} \iota _{\widehat\Gamma _L} \mathcal B_{L,\Gamma }(\iota _{\Lambda ,\Gamma }(A)) & =\iota _{\widehat\Lambda _L}\mathcal B_{L,\Lambda }(A), \end{align}

so the forward finite-region maps form a compatible family. For the reverse family, block-hull monotonicity sends \(\Delta \subseteq \Theta \) to \(\operatorname {Hull}_L(\Delta )\subseteq \operatorname {Hull}_L(\Theta )\). Transitivity of the canonical inclusions and the inverse commuting square then give

\begin{align} \iota _{\operatorname {Hull}_L(\Theta )} \mathcal U_{L,\Theta }(\iota _{\Delta ,\Theta }(A)) & =\iota _{\operatorname {Hull}_L(\Delta )} \mathcal U_{L,\Delta }(A). \end{align}

The universal property of the algebraic direct limit therefore induces both star-algebra homomorphisms.

Let \(L{\gt}0\). The forward map applies finite blocking to a representative on \(\Lambda \) and regards the result as supported on \(\widehat\Lambda _L\). The reverse map enlarges a representative on \(\Delta \) to the expansion of its block hull, applies the inverse finite blocking equivalence, and regards the result as supported on \(\operatorname {Hull}_L(\Delta )\). Their composites have the orientations

\begin{align} \mathcal B_L\mathcal U_L & =\operatorname {id}_{\mathcal A_{\mathrm{loc}}^{(d)}}, \\ \mathcal U_L\mathcal B_L & =\operatorname {id}_{\mathcal A_{\mathrm{loc}}^{(d^L)}}. \end{align}

The first identity holds for every on-site dimension \(d\); the second also assumes \(d{\gt}0\).

Proof

These are the evaluation formulas for the two compatible-family lifts. Forward compatibility is the commuting square. Reverse compatibility follows by monotonicity of block hulls, composition of local inclusions, and the same square applied in the inverse direction. Forward blocking after the reverse map reduces to the canonical inclusion of \(\Delta \) into its expanded block hull. The other composite is the identity by injectivity of the forward map.

Definition 19.10.1.99 Algebraic-local site-blocking equivalence

For \(d{\gt}0\) and \(L{\gt}0\), forward blocking gives a star-algebra equivalence

\begin{align} \mathcal A_{\mathrm{loc}}^{(d^L)} & \simeq \mathcal A_{\mathrm{loc}}^{(d)}. \end{align}

This formalizes only the algebraic-local change of site grouping used at lines 2308 and 2313–2320 of [ CPGSV17 ] . It does not assert the QCA-to-MPU converse at line 2308, extend blocking to the quasi-local completion, intertwine translations, provide a propagation estimate or a circuit representation, or transport a quantum cellular automaton.

Proof

The identity \(\mathcal B_L\mathcal U_L=\operatorname {id}\) makes \(\mathcal B_L\) surjective. For injectivity, represent two blocked observables on finite regions \(\Lambda \) and \(\Gamma \), enlarge both to \(\Omega =\Lambda \cup \Gamma \), and suppose their images under \(\mathcal B_L\) agree. Injectivity of the canonical map from the finite algebra on \(\widehat\Omega _L\), followed by injectivity of \(\mathcal B_{L,\Omega }\) and the finite commuting square, gives equality of the two representatives in the algebra on \(\Omega \). Thus \(\mathcal B_L\) is bijective, with inverse \(\mathcal U_L\), and defines the stated star-algebra equivalence.

Let \(d{\gt}0\) and \(L{\gt}0\). The equivalence agrees with the forward lift, and its inverse agrees with the block-hull unblocking map. For \(A\in \mathcal A^{(d^L)}_\Lambda \) and \(C\in \mathcal A^{(d)}_\Delta \),

\begin{align} \mathcal B_L\bigl(\iota _\Lambda (A)\bigr) & =\iota _{\widehat\Lambda _L} \bigl(\mathcal B_{L,\Lambda }(A)\bigr), \\ \mathcal B_L^{-1}\bigl(\iota _\Delta (C)\bigr) & =\iota _{\operatorname {Hull}_L(\Delta )} \left(\mathcal B^{-1}_{L,\operatorname {Hull}_L(\Delta )} \bigl(\iota _{\Delta , \widehat{\operatorname {Hull}_L(\Delta )}_L}(C)\bigr)\right). \end{align}

In particular, if \(C\in \mathcal A^{(d)}_{\widehat\Lambda _L}\), then

\begin{align} \mathcal B_L^{-1}\bigl(\iota _{\widehat\Lambda _L}(C)\bigr) & =\iota _\Lambda \bigl(\mathcal B_{L,\Lambda }^{-1}(C)\bigr). \end{align}

The forward and inverse maps are operator-norm isometries. In particular, each preserves the norm.

Proof

The first two formulas are the evaluation identities for the forward and reverse maps. For an expanded region, the block-hull identity removes the enlargement and leaves the inverse finite blocking equivalence. On a finite representative, forward norm preservation reduces to norm preservation by finite blocking and by the canonical local inclusion; distance preservation follows by applying the norm identity to differences. The inverse isometry follows from the forward isometry and the right-inverse identity, rather than repeating the norm calculation.

For \(d{\gt}0\) and \(L{\gt}0\), site blocking extends uniquely by continuity to a star-algebra equivalence

\begin{align} \overline{\mathcal B}_L: \mathcal A^{(d^L)} & \simeq \mathcal A^{(d)}. \end{align}

This is only the completion of the change of site grouping used at [ CPGSV17 , lines 2308 and 2313–2320 ] . This slice does not define scaled translations or carry-aware neighborhoods, prove propagation estimates, construct blocked automorphisms, transport the property of being a quantum cellular automaton, realize a circuit, or reconstruct an MPU from a quantum cellular automaton.

Proof

The algebraic blocking equivalence and its inverse are isometries, hence continuous. Extending both maps to the operator-norm completions gives mutually inverse star-algebra homomorphisms.

Let \(d{\gt}0\) and \(L{\gt}0\). Write \(j^{(d)}\) for the canonical map from the algebraic-local algebra into its completion. The underlying map of \(\overline{\mathcal B}_L\) is the continuous extension of \(\mathcal B_L\), and on the two dense algebraic-local subalgebras,

\begin{align} \overline{\mathcal B}_L\bigl(j^{(d^L)}(A)\bigr) & =j^{(d)}\bigl(\mathcal B_L(A)\bigr), \\ \overline{\mathcal B}_L^{-1}\bigl(j^{(d)}(C)\bigr) & =j^{(d^L)}\bigl(\mathcal B_L^{-1}(C)\bigr). \end{align}

For \(A\in \mathcal A^{(d^L)}_\Lambda \) and \(C\in \mathcal A^{(d)}_\Delta \), these identities become

\begin{align} \overline{\mathcal B}_L\bigl(j^{(d^L)}_\Lambda (A)\bigr) & =j^{(d)}_{\widehat\Lambda _L} \bigl(\mathcal B_{L,\Lambda }(A)\bigr), \\ \overline{\mathcal B}_L^{-1}\bigl(j^{(d)}_\Delta (C)\bigr) & =j^{(d^L)}_{\operatorname {Hull}_L(\Delta )} \left(\mathcal B^{-1}_{L,\operatorname {Hull}_L(\Delta )} \bigl(\iota _{\Delta , \widehat{\operatorname {Hull}_L(\Delta )}_L}(C)\bigr)\right). \end{align}

If \(C\in \mathcal A^{(d)}_{\widehat\Lambda _L}\), the inverse formula reduces exactly to

\begin{align} \overline{\mathcal B}_L^{-1} \bigl(j^{(d)}_{\widehat\Lambda _L}(C)\bigr) & =j^{(d^L)}_\Lambda \bigl(\mathcal B_{L,\Lambda }^{-1}(C)\bigr). \end{align}

Finally, for all quasi-local observables \(X\) and \(Y\) on the blocked chain,

\begin{align} \lVert \overline{\mathcal B}_L(X)\rVert & =\lVert X\rVert , \\ \lVert \overline{\mathcal B}_L(X)- \overline{\mathcal B}_L(Y)\rVert & =\lVert X-Y\rVert . \end{align}
Proof

The first identity is the defining completion map, and the two dense formulas are its canonical-image formulas for the algebraic equivalence and its inverse. Composing them with the finite-region canonical maps gives the two finite formulas. The exact expanded-region formula uses \(\operatorname {Hull}_L(\widehat\Lambda _L)=\Lambda \). A star-algebra equivalence of C-star algebras preserves the norm; applying this to \(X-Y\) gives distance preservation.

Let \(d{\gt}0\) and \(L{\gt}0\). If a blocked-chain observable \(X\) is supported in \(\Lambda \), then \(\overline{\mathcal B}_L(X)\) is supported in \(\widehat\Lambda _L\). If an original-chain observable \(Y\) is supported in \(\Delta \), then \(\overline{\mathcal B}_L^{-1}(Y)\) is supported in \(\operatorname {Hull}_L(\Delta )\). In particular, support transport is exact on expanded regions:

\begin{align} \overline{\mathcal B}_L(X) \text{ is supported in }\widehat\Lambda _L \quad \Longleftrightarrow \quad X\text{ is supported in }\Lambda . \end{align}
Proof

Choose a finite-region representative and apply the corresponding forward or inverse finite evaluation formula. For the reverse implication in the exact statement, apply inverse blocking and simplify the inverse composite and the block hull of the expanded region.

Lemma 19.10.1.104 Expansion of a translated blocked region

Let \(L{\gt}0\), let \(a\in \mathbb Z\), and let \(\Lambda \subset \mathbb Z\) be a finite blocked region. Then

\begin{align} \widehat{\Lambda +a}_L & =\widehat\Lambda _L+aL. \end{align}

This is the translation rule for the site grouping used in [ CPGSV17 , lines 2308 and 2313–2320 ] .

Proof

Write each original site uniquely as \(xL+r\), with \(0\leq r{\lt}L\). Translation of the blocked coordinate by \(a\) sends this site to \((x+a)L+r=xL+r+aL\). The argument is unchanged when \(x\) or \(a\) is negative.

Definition 19.10.1.105 Carry-aware blocked neighborhood

Let \(L{\gt}0\) and let \(\mathcal N\subset \mathbb Z\) be a finite original-lattice neighborhood. Write \(\widehat{\{ 0\} }_L=\{ 0,\ldots ,L-1\} \) for the sites in the reference block. Define the induced blocked-lattice neighborhood by

\begin{align} \mathcal C_L(\mathcal N) & =\operatorname {Hull}_L \bigl(\widehat{\{ 0\} }_L+\mathcal N\bigr). \end{align}

The full reference block is included because the block reached after a displacement can depend on the starting residue. This is the finite-region geometry implicit in the grouping of sites in [ CPGSV17 , lines 2308 and 2313–2320 ] .

Let \(L{\gt}0\), let \(\mathcal N\subset \mathbb Z\) be finite, and let \(b\in \mathbb Z\). Then

\begin{align} b\in \mathcal C_L(\mathcal N) \quad \Longleftrightarrow \quad & \exists r\in \{ 0,\ldots ,L-1\} ,\ \exists n\in \mathcal N, \quad \left\lfloor \frac{r+n}{L}\right\rfloor =b, \\ \quad \Longleftrightarrow \quad & \exists r,s\in \{ 0,\ldots ,L-1\} ,\ \exists n\in \mathcal N, \quad r+n=bL+s. \end{align}

The quotient is the Euclidean quotient by the positive integer \(L\). Consequently these formulas include negative displacements and both positive and negative block carries. In general, \(\operatorname {Hull}_L(\mathcal N)\) alone omits some carries caused by the starting residue. This makes explicit the boundary geometry implicit in [ CPGSV17 , lines 2313–2320 ] .

Proof

An element of \(\widehat{\{ 0\} }_L+\mathcal N\) has the form \(r+n\) with \(0\leq r{\lt}L\) and \(n\in \mathcal N\). Taking its block coordinate gives the first equivalence. Euclidean division writes \(r+n=bL+s\) with the unique remainder \(0\leq s{\lt}L\), which gives the second equivalence without any sign restriction on \(n\).

Let \(L{\gt}0\). For every finite blocked region \(\Lambda \) and every finite original-lattice neighborhood \(\mathcal N\),

\begin{align} \operatorname {Hull}_L \bigl(\widehat\Lambda _L+\mathcal N\bigr) & =\Lambda +\mathcal C_L(\mathcal N). \end{align}

This is the exact finite-region carry identity associated with the site grouping in [ CPGSV17 , lines 2313–2320 ] .

Proof

Write an expanded site as \(xL+r\), where \(x\in \Lambda \) and \(0\leq r{\lt}L\). For \(n\in \mathcal N\), Euclidean division gives

\begin{align} \left\lfloor \frac{xL+r+n}{L}\right\rfloor & =x+\left\lfloor \frac{r+n}{L}\right\rfloor . \end{align}

The second summand ranges over exactly \(\mathcal C_L(\mathcal N)\) by Lemma 19.10.1.106. This proves both inclusions, including for negative \(x\) and \(n\).

Let \(L{\gt}0\). For every finite blocked region \(\Lambda \) and every finite original-lattice neighborhood \(\mathcal N\),

\begin{align} \widehat\Lambda _L+\mathcal N & \subseteq \widehat{\Lambda +\mathcal C_L(\mathcal N)}_L. \end{align}
Proof

Every finite original region is contained in the expansion of its block hull. Apply this containment to \(\widehat\Lambda _L+\mathcal N\) and use Lemma 19.10.1.107 to identify that hull.

Let \(d,L{\gt}0\). If \(\mathcal B_L\) denotes the algebraic-local blocking equivalence, then for every \(a\in \mathbb Z\),

\begin{align} \mathcal B_L\, \tau ^{(d^L)}_a & =\tau ^{(d)}_{aL}\, \mathcal B_L. \end{align}

This is the translation rule induced by the blocking of sites used in [ CPGSV17 , lines 2308 and 2313–2320 ] .

Proof

On a representative supported in a finite blocked region \(\Lambda \), both sides apply the same matrix reindexing. Their supports agree because

\begin{align} \widehat{\Lambda +a}_L & =\widehat\Lambda _L+aL. \end{align}

Equality on all finite-region representatives proves the identity on the algebraic direct limit.

Let \(d,L{\gt}0\). If \(\overline{\mathcal B}_L\) denotes the quasi-local blocking equivalence, then for every \(a\in \mathbb Z\),

\begin{align} \overline{\mathcal B}_L\, \overline\tau ^{(d^L)}_a & =\overline\tau ^{(d)}_{aL}\, \overline{\mathcal B}_L. \end{align}

This is the completion of the translation rule induced by the site grouping in [ CPGSV17 , lines 2308 and 2313–2320 ] .

Proof

The two sides are continuous maps on the quasi-local algebra. On the dense algebraic-local subalgebra their values agree by Lemma 19.10.1.109. They therefore agree on the completion.

Line 2308 of [ CPGSV17 ] invokes a converse representation theorem: after blocking finitely many sites, the restriction of every one-dimensional QCA to the local algebra is represented by a depth-two circuit of nearest-neighbor unitaries, and hence by an MPU in standard form. The results below concern the preceding change of grouping. A given QCA remains a QCA under any prescribed finite blocking. This one-way transport alone does not produce the circuit or MPU representation asserted by the converse theorem.

Definition 19.10.1.111 Blocked automorphism
#

Let \(d,L{\gt}0\), let \(\overline{\mathcal B}_L\colon \mathcal A^{(d^L)}\simeq \mathcal A^{(d)}\) be the quasi-local blocking equivalence, and let \(\omega \) be a star-algebra automorphism of \(\mathcal A^{(d)}\). The induced automorphism of the blocked chain is

\begin{align} \omega ^{[L]} & =\overline{\mathcal B}_L^{-1}\circ \omega \circ \overline{\mathcal B}_L. \end{align}
Lemma 19.10.1.112 Evaluation of the blocked automorphism

For every \(X\in \mathcal A^{(d^L)}\),

\begin{align} \omega ^{[L]}(X) & =\overline{\mathcal B}_L^{-1} \bigl(\omega (\overline{\mathcal B}_L(X))\bigr). \end{align}
Proof

This is the evaluation formula for the defining conjugation.

Lemma 19.10.1.113 Inverse of the blocked automorphism

The inverse of the blocked automorphism is the blocked inverse:

\begin{align} (\omega ^{[L]})^{-1} & =(\omega ^{-1})^{[L]}. \end{align}
Proof

Evaluate both sides on a blocked-chain observable. The blocking equivalence and its inverse cancel on the two sides of \(\omega ^{-1}\).

Let \(\mathcal N\subset \mathbb Z\) be finite. If \(\omega \) propagates within \(\mathcal N\), then \(\omega ^{[L]}\) propagates within the carry-aware blocked neighborhood \(\mathcal C_L(\mathcal N)\). Explicitly, for every finite blocked region \(\Lambda \) and every \(X\in \mathcal A^{(d^L)}\) supported in \(\Lambda \), the observable \(\omega ^{[L]}(X)\) is supported in \(\Lambda +\mathcal C_L(\mathcal N)\).

Proof

Forward blocking sends support in \(\Lambda \) to support in \(\widehat\Lambda _L\). Apply the propagation bound for \(\omega \), then apply inverse blocking. The resulting support is

\begin{align} \operatorname {Hull}_L \bigl(\widehat\Lambda _L+\mathcal N\bigr) & =\Lambda +\mathcal C_L(\mathcal N) \end{align}

by the exact carry-aware block-hull identity.

Lemma 19.10.1.115 Finite propagation after site blocking

If \(\omega \) has finite forward propagation, then \(\omega ^{[L]}\) has finite forward propagation.

Proof

Choose a finite propagation neighborhood \(\mathcal N\) for \(\omega \). Lemma 19.10.1.114 gives the finite neighborhood \(\mathcal C_L(\mathcal N)\) for \(\omega ^{[L]}\).

The blocked automorphism is translation covariant if and only if the original automorphism commutes with translation by one whole block:

\begin{align} \omega ^{[L]}\text{ is translation covariant} \quad \Longleftrightarrow \quad \omega \circ \overline\tau ^{(d)}_L & =\overline\tau ^{(d)}_L\circ \omega . \end{align}

Thus blocked unit-translation covariance records commutation of \(\omega \) with translation by \(L\), not necessarily commutation with original unit translation. For \(L\neq 1\), the latter condition can be strictly stronger.

Proof

By Theorem 19.10.1.61, translation covariance on the blocked chain is equivalent to commutation with blocked unit translation. Under \(\overline{\mathcal B}_L\), blocked unit translation corresponds to original translation by \(L\). Conjugating the commutation identity by the blocking equivalence proves both implications.

Lemma 19.10.1.117 Translation covariance after site blocking

If \(\omega \) is translation covariant, then \(\omega ^{[L]}\) is translation covariant.

Proof

Translation covariance of \(\omega \) gives commutation with \(\overline\tau ^{(d)}_L\). Apply Theorem 19.10.1.116.

Let \(d,L{\gt}0\). If \(\omega \) is a one-dimensional QCA on \(\mathcal A^{(d)}\), then

\begin{align} \omega ^{[L]} & =\overline{\mathcal B}_L^{-1}\circ \omega \circ \overline{\mathcal B}_L \end{align}

is a one-dimensional QCA on \(\mathcal A^{(d^L)}\).

Proof

The two defining properties of a one-dimensional QCA are finite forward propagation and translation covariance. They are preserved by Lemma 19.10.1.115 and Lemma 19.10.1.117, respectively.

Line 2308 of [ CPGSV17 ] invokes the external Schumacher–Werner structure theorem to obtain support algebras, a depth-two nearest-neighbor circuit, and an MPU in standard form. The next statements establish only the elementary normalization of a finite propagation neighborhood to \([-1,1]\cap \mathbb Z\) by regrouping sites. They do not prove any part of that external structure theorem beyond this change of grouping.

Lemma 19.10.1.119 Nearest-neighbor carries from a strict displacement bound

Let \(L{\gt}0\) and let \(\mathcal N\subset \mathbb Z\) be finite. If \(|n|{\lt}L\) for every \(n\in \mathcal N\), then

\begin{align} \mathcal C_L(\mathcal N) & \subseteq [-1,1]\cap \mathbb Z. \end{align}
Proof

If \(\mathcal N=\varnothing \), membership in \(\mathcal C_L(\mathcal N)\) is impossible. In general, a point \(b\in \mathcal C_L(\mathcal N)\) supplies residues \(0\leq r,s{\lt}L\) and \(n\in \mathcal N\) such that

\begin{align} r+n & =bL+s. \end{align}

Since \(-L{\lt}n{\lt}L\), the equality is incompatible with \(b\leq -2\) and with \(b\geq 2\). Thus \(-1\leq b\leq 1\), including for negative displacements.

Lemma 19.10.1.120 Supremum criterion for nearest-neighbor carries

Let \(L{\gt}0\) and let \(\mathcal N\subset \mathbb Z\) be finite. If

\begin{align} \sup _{n\in \mathcal N}|n| & {\lt}L, \end{align}

then

\begin{align} \mathcal C_L(\mathcal N) & \subseteq [-1,1]\cap \mathbb Z. \end{align}
Proof

Every \(|n|\) is bounded above by the finite supremum, so the strict supremum bound supplies the pointwise hypothesis. When \(\mathcal N=\varnothing \), the supremum is \(0\) and the conclusion is vacuous.

Lemma 19.10.1.121 Canonical nearest-neighbor block length

For every finite \(\mathcal N\subset \mathbb Z\), define

\begin{align} L_{\mathcal N} & =1+\sup _{n\in \mathcal N}|n|. \end{align}

Then \(L_{\mathcal N}{\gt}0\) and

\begin{align} \mathcal C_{L_{\mathcal N}}(\mathcal N) & \subseteq [-1,1]\cap \mathbb Z. \end{align}

With the finite-supremum convention, \(L_{\varnothing }=1\).

Proof

The finite supremum is strictly smaller than its successor, which is positive even when \(\mathcal N\) is empty.

Let \(d,L{\gt}0\). Suppose that \(\omega \) propagates within the finite neighborhood \(\mathcal N\) and that \(|n|{\lt}L\) for every \(n\in \mathcal N\). For every finite blocked region \(\Lambda \) and every observable \(X\in \mathcal A^{(d^L)}\) supported in \(\Lambda \), the observable \(\omega ^{[L]}(X)\) is supported in

\begin{align} \Lambda +([-1,1]\cap \mathbb Z). \end{align}
Proof

Blocking first gives propagation within \(\mathcal C_L(\mathcal N)\). The carry bound places this neighborhood inside \([-1,1]\cap \mathbb Z\), and enlargement of propagation neighborhoods gives the claim.

Let \(d{\gt}0\). If \(\omega \) has finite forward propagation, then there exists a positive integer \(L\) such that, for every finite blocked region \(\Lambda \) and every observable \(X\in \mathcal A^{(d^L)}\) supported in \(\Lambda \), the observable \(\omega ^{[L]}(X)\) is supported in

\begin{align} \Lambda +([-1,1]\cap \mathbb Z). \end{align}
Proof

Choose a finite propagation neighborhood \(\mathcal N\) and take \(L=1+\sup _{n\in \mathcal N}|n|\). This \(L\) is positive, including when \(\mathcal N\) is empty. Blocking gives propagation within \(\mathcal C_L(\mathcal N)\); the canonical carry inclusion and propagation monotonicity then give the nearest-neighbor bound.

Let \(d{\gt}0\) and let \(\omega \) be a one-dimensional QCA on \(\mathcal A^{(d)}\). There exists a positive integer \(L\) such that \(\omega ^{[L]}\) is a QCA on \(\mathcal A^{(d^L)}\) and, for every finite blocked region \(\Lambda \) and every observable \(X\in \mathcal A^{(d^L)}\) supported in \(\Lambda \), the observable \(\omega ^{[L]}(X)\) is supported in

\begin{align} \Lambda +([-1,1]\cap \mathbb Z). \end{align}
Proof

Apply the preceding theorem to the finite propagation of \(\omega \), using the canonical positive block length. At the same length, site blocking preserves both finite propagation and translation covariance, hence the QCA property.

Lines 2292–2298 of [ CPGSV17 ] specify the finite local algebras, their canonical inclusions, translation covariance, and the locality inclusion

\begin{align} \omega (\mathcal A_\Lambda ) & \subseteq \mathcal A_{\Lambda +\mathcal N}. \end{align}

The following statements record the finite-dimensional maps determined by this inclusion. They are consequences and coordinate formulations of the cited locality relation, not separately named theorems of the paper. The left/right order below agrees with the adjacent-pair convention in [ GNVW12 ] .

Let \(d{\gt}0\), let \(\omega \) be a star-algebra automorphism of \(\mathcal A^{(d)}\), and suppose that \(\omega \) propagates within the finite neighborhood \(\mathcal N\). For every finite region \(\Lambda \), define the finite-region restriction

\begin{align} \omega _{\Lambda ,\mathcal N}\colon \mathcal A_\Lambda & \longrightarrow \mathcal A_{\Lambda +\mathcal N} \end{align}

to be the unique star-algebra homomorphism satisfying

\begin{align} \iota _{\Lambda +\mathcal N} \bigl(\omega _{\Lambda ,\mathcal N}(A)\bigr) & =\omega \bigl(\iota _\Lambda (A)\bigr) \qquad (A\in \mathcal A_\Lambda ). \end{align}

The finite-region restriction is injective, and for every \(A\in \mathcal A_\Lambda \) one has

\begin{align} \iota _{\Lambda +\mathcal N} \bigl(\omega _{\Lambda ,\mathcal N}(A)\bigr) & =\omega \bigl(\iota _\Lambda (A)\bigr). \end{align}
Proof

The defining equality follows from the locality inclusion and the injectivity of \(\iota _{\Lambda +\mathcal N}\). If two local observables have the same image under \(\omega _{\Lambda ,\mathcal N}\), their canonical images become equal after applying \(\omega \). Injectivity of \(\omega \) and of \(\iota _\Lambda \) then gives equality of the two local observables.

Theorem 19.10.1.127 Enlargement of finite-region restrictions

If \(\Lambda \subseteq \Gamma \), then for every \(A\in \mathcal A_\Lambda \),

\begin{align} \omega _{\Gamma ,\mathcal N} \bigl(\iota _{\Lambda ,\Gamma }(A)\bigr) & =\iota _{\Lambda +\mathcal N,\Gamma +\mathcal N} \bigl(\omega _{\Lambda ,\mathcal N}(A)\bigr). \end{align}
Proof

Apply the canonical inclusion into the quasi-local algebra to both sides. Both expressions become \(\omega (\iota _\Lambda (A))\); injectivity of the target inclusion gives the equality.

Suppose in addition that \(\omega \) is translation covariant. For \(a\in \mathbb Z\) and \(A\in \mathcal A_\Lambda \),

\begin{align} \iota _{(\Lambda +a)+\mathcal N} \bigl(\omega _{\Lambda +a,\mathcal N} (\tau _{a,\Lambda }(A))\bigr) & =\iota _{(\Lambda +\mathcal N)+a} \bigl(\tau _{a,\Lambda +\mathcal N} (\omega _{\Lambda ,\mathcal N}(A))\bigr). \end{align}

Here \((\Lambda +a)+\mathcal N=(\Lambda +\mathcal N)+a\).

Proof

The left-hand side is \(\omega (\tau _a(\iota _\Lambda (A)))\). Translation covariance moves \(\tau _a\) past \(\omega \), and the finite-region translation formula gives the right-hand side.

Definition 19.10.1.129 Bipartite coordinates for disjoint local algebras

Let \(\Gamma ,\Delta \subset \mathbb Z\) be disjoint finite regions. Writing \(\mathcal C_\Gamma \) and \(\mathcal C_\Delta \) for their configuration spaces, restriction to the two regions gives

\begin{align} \mathcal C_{\Gamma \cup \Delta } & \simeq \mathcal C_\Gamma \times \mathcal C_\Delta , \end{align}

with the \(\Gamma \)-configuration first. Reindexing matrix rows and columns gives a star-algebra equivalence

\begin{align} \beta _{\Gamma ,\Delta }\colon \mathcal A_{\Gamma \cup \Delta } & \simeq M_{\mathcal C_\Gamma \times \mathcal C_\Delta }(\mathbb C). \end{align}

For \(A\in \mathcal A_\Gamma \) and \(B\in \mathcal A_\Delta \),

\begin{align} \beta _{\Gamma ,\Delta }(\iota _{\Gamma ,\Gamma \cup \Delta }(A)) & =A\otimes \mathbf1, \\ \beta _{\Gamma ,\Delta }(\iota _{\Delta ,\Gamma \cup \Delta }(B)) & =\mathbf1\otimes B. \end{align}
Proof

Evaluate both matrices on pairs of configurations. For the first local inclusion, restriction to \(\Gamma \) gives the matrix coefficient of \(A\), while equality of the complementary \(\Delta \)-configurations gives the matrix coefficient of the identity. This is \(A\otimes \mathbf1\). Interchanging the two regions gives \(\mathbf1\otimes B\).

Define the finite local image by

\begin{align} \mathcal R_{\Lambda ,\mathcal N} & =\omega _{\Lambda ,\mathcal N}(\mathcal A_\Lambda ) \subseteq \mathcal A_{\Lambda +\mathcal N}. \end{align}

It is a star-subalgebra star-isomorphic to \(\mathcal A_\Lambda \). If \(\Lambda +\mathcal N=\Gamma \cup \Delta \) with \(\Gamma \cap \Delta =\varnothing \), then

\begin{align} \beta _{\Gamma ,\Delta } (\mathcal R_{\Lambda ,\mathcal N}) & \subseteq M_{\mathcal C_\Gamma \times \mathcal C_\Delta }(\mathbb C) \end{align}

is the bipartite matrix star-subalgebra to which the left and right support-algebra constructions apply. No support algebra is defined in this statement.

Definition 19.10.1.132 Nearest-neighbour pair regions

For \(x\in \mathbb Z\), set

\begin{align} E_x & =\{ 2x,2x+1\} , & L_x & =\{ 2x-1,2x\} , & P_x & =\{ 2x+1,2x+2\} . \label{eq:qca_pair_regions} \end{align}

The order \(L_x,P_x\) is the left–right factor order in equation RR2x of [ GNVW12 , lines 1251–1266 ] . The same four-site localization occurs in [ CPGSV17 , lines 2313–2318 ] .

Lemma 19.10.1.133 Geometry of the nearest-neighbour pair image

For the neighbourhood \(\mathcal N=\{ -1,0,1\} \),

\begin{align} E_x+\mathcal N & =L_x\cup P_x, & L_x\cap P_x & =\varnothing , & P_x & =L_{x+1}. \label{eq:qca_pair_geometry} \end{align}
Proof

Substitute (869). Adding \(-1,0,1\) to the two points of \(E_x\) gives

\begin{align} E_x+\{ -1,0,1\} & =\{ 2x-1,2x,2x+1,2x+2\} =L_x\cup P_x. \end{align}

The largest point of \(L_x\) is \(2x\), while the smallest point of \(P_x\) is \(2x+1\), so the two sets are disjoint. Finally, \(L_{x+1}=\{ 2x+1,2x+2\} =P_x\).

Let \(\omega \) be a star-algebra automorphism of the homogeneous quasi-local algebra which propagates within \(\{ -1,0,1\} \). Write the finite image of \(\mathcal A_{E_x}\) in the ordered \(L_x\times P_x\) coordinates as

\begin{align} S_x & =\beta _{L_x,P_x}\bigl(\omega (\mathcal A_{E_x})\bigr) \subseteq M_{\mathcal C_{L_x}\times \mathcal C_{P_x}}(\mathbb C). \label{eq:qca_site_image} \end{align}

Define

\begin{align} \mathcal R_{2x} & =\operatorname {Spp}_{\mathrm L}(S_x) \subseteq M_{\mathcal C_{L_x}}(\mathbb C), \\ \mathcal R_{2x+1} & =\operatorname {Spp}_{\mathrm R}(S_x) \subseteq M_{\mathcal C_{P_x}}(\mathbb C). \label{eq:qca_RR2x} \end{align}

After the canonical finite-region embeddings, these form one family of star-subalgebras of the quasi-local algebra. Here the matrix embedding is the composition

\begin{align} M_{\mathcal C_\Lambda }(\mathbb C) \cong \mathcal A_\Lambda \longrightarrow \mathcal A. \end{align}

Under this identification, the canonical local inclusion in [ CPGSV17 , lines 2292–2298 ] is exactly the displayed composition. Formula (874) is equation RR2x in [ GNVW12 , lines 1261–1266 ] . Cirac et al. use the homogeneous even factor in [ CPGSV17 , lines 2322–2328 ] ; the odd factor and unified family are supplied by the GNVW construction.

Scope restriction (homogeneous chain): GNVW allow the cell size \(d(y)\) to depend on \(y\), whereas the present definition fixes one positive size \(d\). The site-dependent statement is not claimed here; see [ con26l ] .

For every \(x\in \mathbb Z\), the finite image satisfies

\begin{align} S_x & \subseteq \mathcal R_{2x}\boxtimes \mathcal R_{2x+1} \subseteq M_{\mathcal C_{L_x}\times \mathcal C_{P_x}}(\mathbb C). \label{eq:qca_adjacent_pair_containment} \end{align}

This is the homogeneous, full-matrix instance used in the two-sided containment at [ GNVW12 , lines 1276–1278 ] ; see [ con26l ] . Cirac et al. state only the weaker inclusion in the left support algebra tensored with the unrestricted right physical factor at [ CPGSV17 , lines 2322–2328 ] .

Proof

Apply Theorem 19.10.1.17 to \(S_x\) in the ordered coordinates of (872). Its left support is \(\mathcal R_{2x}\) and its right support is \(\mathcal R_{2x+1}\) by (874), which gives (876).

If \(\widehat{\mathcal R}_y\) denotes the canonical quasi-local copy of \(\mathcal R_y\), then for every \(x\in \mathbb Z\),

\begin{align} \widehat{\mathcal R}_{2x} & \subseteq \mathcal A_{L_x}, & \widehat{\mathcal R}_{2x+1} & \subseteq \mathcal A_{P_x}. \end{align}

More generally, the canonical image of any matrix on a finite region \(\Lambda \) is supported in \(\Lambda \). This is the homogeneous specialization of the localization accompanying equation RR2x in [ GNVW12 , lines 1261–1274 ] ; see [ con26l ] .

Proof

An element of \(\widehat{\mathcal R}_{2x}\) is the canonical image of a matrix in \(\mathcal R_{2x}\subseteq \mathcal A_{L_x}\) and is therefore supported in \(L_x\). The same argument places \(\widehat{\mathcal R}_{2x+1}\) in \(\mathcal A_{P_x}\).

Equation (A1) in lines 2300–2306 of [ CPGSV17 ] defines the local action associated with an MPU by eventual finite-chain conjugation. In lines 2300–2306, eventual constancy is used to obtain a norm-preserving star-algebra automorphism of \(\mathcal A_{\mathrm{loc}}\), together with a translation formula and the support bound \(\mathcal N=[-4D^4,4D^4]\cap \mathbb Z\), and then to extend this automorphism to \(\mathcal A\). The results below isolate the direct-limit and completion argument. They begin with compatible forward and inverse finite-local actions and assume a forward norm formula, a support bound, and a unit-translation formula. The inverse norm formula follows from forward isometry and mutual invertibility. The numerical radius above belongs to the MPU-specific stabilization argument: the generic result below neither proves that radius nor constructs the finite-local actions from an MPU. It also does not provide the blocked circuit or MPU standard form discussed at line 2308.

Definition 19.10.1.137 Compatible finite-local automorphism data

For each finite region \(\Lambda \), let

\begin{align} f_\Lambda ,g_\Lambda \colon \mathcal A_\Lambda & \longrightarrow \mathcal A_{\mathrm{loc}} \end{align}

be star-algebra homomorphisms such that, whenever \(\Lambda \subseteq \Gamma \),

\begin{align} f_\Gamma \circ \iota _{\Lambda ,\Gamma } & =f_\Lambda , & g_\Gamma \circ \iota _{\Lambda ,\Gamma } & =g_\Lambda . \end{align}

Let \(f,g\colon \mathcal A_{\mathrm{loc}}\to \mathcal A_{\mathrm{loc}}\) be the maps induced by these compatible families. Compatible finite-local automorphism data require

\begin{align} g(f_\Lambda (A)) & =\iota _\Lambda (A), & f(g_\Lambda (A)) & =\iota _\Lambda (A) \end{align}

for every \(A\in \mathcal A_\Lambda \).

The compatible families induce star-algebra homomorphisms

\begin{align} f,g\colon \mathcal A_{\mathrm{loc}} & \longrightarrow \mathcal A_{\mathrm{loc}}. \end{align}
Proof

Apply the star-algebra universal property to each compatible family.

For every finite region \(\Lambda \) and \(A\in \mathcal A_\Lambda \),

\begin{align} f(\iota _\Lambda (A)) & =f_\Lambda (A), & g(\iota _\Lambda (A)) & =g_\Lambda (A). \end{align}
Proof

These are the evaluation identities supplied by the universal property.

Compatible finite-local automorphism data determine a star-algebra equivalence

\begin{align} F\colon \mathcal A_{\mathrm{loc}} & \simeq \mathcal A_{\mathrm{loc}} \end{align}

whose forward and inverse maps are \(f\) and \(g\), respectively.

Proof

Induct on finite-region representatives in the algebraic direct limit. The two inverse assumptions give

\begin{align} g(f(\iota _\Lambda (A))) & =g(f_\Lambda (A))=\iota _\Lambda (A), \\ f(g(\iota _\Lambda (A))) & =f(g_\Lambda (A))=\iota _\Lambda (A). \end{align}

Thus \(f\) and \(g\) are mutually inverse.

For every \(A\in \mathcal A_{\mathrm{loc}}\),

\begin{align} F(A) & =f(A), & F^{-1}(A) & =g(A). \end{align}
Proof

Both identities follow from the defining forward and inverse maps of \(F\).

Definition 19.10.1.142 Forward local norm hypothesis

Assume \(d{\gt}0\). The forward local norm hypothesis is

\begin{align} \lVert f_\Lambda (A)\rVert & =\lVert A\rVert \end{align}

for every finite region \(\Lambda \) and every \(A\in \mathcal A_\Lambda \).

Under the forward local norm hypothesis, every \(A\in \mathcal A_{\mathrm{loc}}\) satisfies

\begin{align} \lVert F(A)\rVert & =\lVert A\rVert . \end{align}
Proof

Induct on finite-region representatives. For \(A=\iota _\Lambda (A_\Lambda )\),

\begin{align} \lVert F(A)\rVert & =\lVert f_\Lambda (A_\Lambda )\rVert =\lVert A_\Lambda \rVert =\lVert A\rVert . \end{align}

Under the forward local norm hypothesis, for all \(A,B\in \mathcal A_{\mathrm{loc}}\),

\begin{align} \lVert F(A)-F(B)\rVert & =\lVert A-B\rVert , & \lVert F^{-1}(A)-F^{-1}(B)\rVert & =\lVert A-B\rVert . \end{align}
Proof

Apply the forward norm identity to \(A-B\) to prove that \(F\) is an isometry. Since \(F\circ F^{-1}\) is the identity, the right-inverse law then makes \(F^{-1}\) an isometry as well.

Lemma 19.10.1.145 Inverse algebraic norm formula

Under the forward local norm hypothesis, every \(A\in \mathcal A_{\mathrm{loc}}\) satisfies

\begin{align} \lVert F^{-1}(A)\rVert & =\lVert A\rVert . \end{align}
Proof

Apply the inverse isometry to the distance from \(A\) to \(0\) and use \(F^{-1}(0)=0\).

Under the forward local norm hypothesis, the algebraic equivalence extends to a star-algebra equivalence

\begin{align} \overline F\colon \mathcal A & \simeq \mathcal A \end{align}

of the quasi-local completion.

Proof

The forward norm formula makes \(F\) an isometry, and the inverse law makes \(F^{-1}\) an isometry. Apply the completion extension for star-algebra equivalences to these two continuous maps.

Let \(j\colon \mathcal A_{\mathrm{loc}}\to \mathcal A\) be the canonical completion map. For every \(A\in \mathcal A_{\mathrm{loc}}\),

\begin{align} \overline F(j(A)) & =j(F(A)), & \overline F^{-1}(j(A)) & =j(F^{-1}(A)). \end{align}
Proof

These are the canonical-image formulas for the continuous extension of a star-algebra equivalence and for its inverse.

For every finite region \(\Lambda \) and \(A\in \mathcal A_\Lambda \),

\begin{align} \overline F(j_\Lambda (A)) & =j(f_\Lambda (A)), & \overline F^{-1}(j_\Lambda (A)) & =j(g_\Lambda (A)). \end{align}
Proof

Substitute \(j_\Lambda =j\circ \iota _\Lambda \) into the dense evaluation formulas and use \(F(\iota _\Lambda (A))=f_\Lambda (A)\) and \(F^{-1}(\iota _\Lambda (A))=g_\Lambda (A)\).

Definition 19.10.1.149 Finite-local support bound

For a finite neighborhood \(\mathcal N\subset \mathbb Z\), the forward family has support bound \(\mathcal N\) if

\begin{align} f_\Lambda (A) & \in \iota _{\Lambda +\mathcal N} (\mathcal A_{\Lambda +\mathcal N}) \end{align}

for every finite region \(\Lambda \) and every \(A\in \mathcal A_\Lambda \).

If the forward family has support bound \(\mathcal N\), then the completed automorphism \(\overline F\) propagates within \(\mathcal N\).

Proof

Let \(X=j_\Lambda (A)\) be supported in \(\Lambda \). Choose \(B\in \mathcal A_{\Lambda +\mathcal N}\) such that \(f_\Lambda (A)=\iota _{\Lambda +\mathcal N}(B)\). Then

\begin{align} \overline F(X) & =j(f_\Lambda (A)) =j_{\Lambda +\mathcal N}(B), \end{align}

so the image is supported in \(\Lambda +\mathcal N\).

Definition 19.10.1.151 Finite-local unit-translation formula

The forward family commutes with unit translation if, for every finite region \(\Lambda \) and every \(A\in \mathcal A_\Lambda \),

\begin{align} f_{\Lambda +1}(\tau _{1,\Lambda }(A)) & =\tau _1(f_\Lambda (A)). \end{align}

If the finite-local unit-translation formula holds, then every \(A\in \mathcal A_{\mathrm{loc}}\) satisfies

\begin{align} F(\tau _1(A)) & =\tau _1(F(A)). \end{align}
Proof

Induct on finite-region representatives and write \(A=\iota _\Lambda (A_\Lambda )\). Then

\begin{align} F(\tau _1(A)) & =f_{\Lambda +1}(\tau _{1,\Lambda }(A_\Lambda )) =\tau _1(f_\Lambda (A_\Lambda )) =\tau _1(F(A)). \end{align}

If the finite-local unit-translation formula holds, then the completed automorphism \(\overline F\) is translation covariant.

Proof

By Theorem 19.10.1.61, it suffices to prove that \(\overline F\) commutes with translation by one site. Induct on the completion. On every \(A\in \mathcal A_{\mathrm{loc}}\),

\begin{align} \overline F(\overline\tau _1(j(A))) & =j(F(\tau _1(A))) =j(\tau _1(F(A))) =\overline\tau _1(\overline F(j(A))). \end{align}

The equality set is closed because both composites are continuous, so the identity extends from the canonical dense subalgebra to the completion.

Let \(d{\gt}0\). Suppose the compatible forward family preserves the local operator norm, has a finite support bound, and satisfies the finite-local unit-translation formula. Then the completed automorphism \(\overline F\) is a one-dimensional QCA.

Proof

The support hypothesis supplies a finite forward propagation neighborhood, while the unit-translation formula supplies translation covariance. These are exactly the two defining properties of a one-dimensional QCA.