Tensor Network Theory: A formalization blueprint

E Canonical Form Reduction: Supporting Results

This appendix supports Chapter 9. It collects the projection, gauge, blocking, compression, and reindexing arguments used by the main reduction while keeping the source-named canonical-form statements and the principal reductions in the chapter.

Throughout, \(\mathcal{E}_A(X)=\sum _iA^iX(A^i)^\dagger \), \(A^{[p]}\) denotes blocking by \(p\), and \(V^{(N)}(A)_\sigma =\operatorname{tr}(A^{\sigma _1}\cdots A^{\sigma _N})\). Weighted direct sums are written \(\bigoplus _k\mu _kA_k\).

E.1 Similarities, strict splitting, and zero blocks

Similarity invariance and strict dimension decrease justify the induction used in Theorem 9.1.1.2; the final result removes blocks that vanish at every positive length.

Theorem E.1.1 Unitary conjugation preserves the MPV family
#

If \(U\) is a unitary matrix, then \(A\) and \(U^\dagger A U\) generate the same MPV family.

Proof

By cyclicity of the trace, \(\operatorname{tr}((U^\dagger A^{i_1} U) \cdots (U^\dagger A^{i_N} U)) = \operatorname{tr}(A^{i_1} \cdots A^{i_N})\).

Theorem E.1.2 Nontrivial invariant projection gives strict two-block reduction

Under the hypotheses of Theorem 9.1.1, assume in addition that \(P\neq 0\) and \(P\neq \mathbb {1}\). Then there exist \(n,m\) with \(n+m=D\), \(n{\lt}D\), and \(m{\lt}D\), together with block tensors \(A_1,A_2\), such that \(A\) and \(A_1\oplus A_2\) generate the same MPV family in the sense of Definition 2.3.3.

Proof

Apply Theorem 9.1.1. In the diagonal form of \(P\), the \(1\)-eigenspace is nonempty because \(P\neq 0\), and the \(0\)-eigenspace is nonempty because \(P\neq \mathbb {1}\). Hence both block sizes are positive. Since \(n+m=D\), this implies \(n{\lt}D\) and \(m{\lt}D\).

Theorem E.1.3 Nonzero irreducible block decomposition

Every MPS tensor \(A\) admits a finite family of nonzero irreducible blocks \((B_k)_{k=1}^r\), each with at least one nonzero Kraus operator and positive bond dimension. For every \(N\ge 1\) and every \(\sigma \),

\begin{align} V^{(N)}(A)_\sigma & = V^{(N)}\! \left(\bigoplus _{k=1}^r B_k\right)_\sigma . \label{eq:canonical_nonzero_decomp} \end{align}

Moreover, \(\sum _{k=1}^rD_k\le D\).

Proof

Apply the irreducible block decomposition (Theorem 9.1.1.2) and retain the blocks having a nonzero Kraus operator. For every all-zero irreducible block \(C_j\) and every \(N\ge 1\), \(V^{(N)}(C_j)_\sigma =0\), so (??) holds. The dimension inequality follows by restricting the structural identity \(\sum _jD_j=D\) to the retained subset.

E.2 Projector closure and isometric corner decompositions

The results below turn reducing projections into support isometries, compress tensors to their corners, and reconstruct closed-chain traces from an orthogonal family of corners. They support Theorem 9.7.2 and Theorem 9.7.4.

Theorem E.2.1 Invariant projections are reducing under projector closure

Suppose that \(A\) has invariant-projector closure and that \(P\) is an orthogonal projection satisfying \((\mathbb {1}-P)A^iP=0\) for all \(i\). Then \(P\) reduces every tensor matrix: \(PA^i=A^iP\) for all \(i\).

Proof

The lower-corner relation gives \((\mathbb {1}-P)A^i=(\mathbb {1}-P)A^i(\mathbb {1}-P)\). Applying projector closure to \(\mathbb {1}-P\) gives \(A^i(\mathbb {1}-P)=(\mathbb {1}-P)A^i(\mathbb {1}-P)\), hence the upper corner \(PA^i(\mathbb {1}-P)\) also vanishes. Therefore \(PA^i=PA^iP=A^iP\).

Theorem E.2.2 Invariant-projector closure gives a reducing projection

If \(A\) has invariant-projector closure and admits a nontrivial invariant projection, then it admits a nontrivial orthogonal projection commuting with every \(A^i\).

Proof

Apply Theorem E.2.1 to the projection supplied by Definition 9.1.3.

Lemma E.2.3 Support isometry of an orthogonal projection

Every orthogonal projection \(P\) on \(\mathbb {C}^D\) factors through an isometry: there are \(n\in \mathbb {N}\) and \(V\in \mathbb {C}^{D\times n}\) such that

\begin{align} V^\dagger V & = \mathbb {1}, \notag \\ VV^\dagger & = P, \notag \\ n & = \operatorname{tr}(P). \label{eq:canonical_projection_isometry} \end{align}
Proof

Since \(P\) is Hermitian and idempotent, its eigenvalues are \(0\) and \(1\). Writing \(n=\operatorname{tr}(P)\), unitary diagonalization gives

\begin{align} P & = U\operatorname{diag}(\underbrace{1,\ldots ,1}_{n},0,\ldots ,0)U^\dagger . \notag \end{align}

The matrix \(V\in \mathbb {C}^{D\times n}\) formed by the first \(n\) columns of \(U\) satisfies (??).

Lemma E.2.4 Corner tensors inherit projector closure

Let \(A\) have invariant-projector closure, and let \(V\) be an isometry, \(V^\dagger V=\mathbb {1}\), whose range projection \(VV^\dagger \) commutes with every \(A^i\). Then the corner tensor \(B^i=V^\dagger A^iV\) again has invariant-projector closure.

Proof

Let \(Q\) be an orthogonal projection with \(QB^i=QB^iQ\) for all \(i\). Then \(VQV^\dagger \) is an orthogonal projection. Commutation of \(VV^\dagger \) with each \(A^i\) gives \(V^\dagger A^i=B^iV^\dagger \), so left-invariance of \(Q\) yields

\begin{align} (VQV^\dagger )A^i & =VQB^iV^\dagger =VQB^iQV^\dagger =(VQV^\dagger )A^i(VQV^\dagger ). \notag \end{align}

Projector closure for \(A\) yields \(A^i(VQV^\dagger )=(VQV^\dagger )A^i(VQV^\dagger )\), and compressing both sides between \(V^\dagger \) and \(V\) returns \(B^iQ=QB^iQ\).

Lemma E.2.5 Transfer map of an intertwined corner

Let \(V\) satisfy \(A^iV=VB^i\) for all \(i\). Then, for every \(X\),

\begin{align} \mathcal{E}_A(VXV^\dagger ) & =V\mathcal{E}_B(X)V^\dagger . \label{eq:canonical_corner_transfer} \end{align}

If moreover \(V^\dagger V=\mathbb {1}\), then every eigenvalue of \(\mathcal{E}_B\) is an eigenvalue of \(\mathcal{E}_A\).

Proof

Expanding the left-hand side of (??) termwise, the intertwining \(A^iV=VB^i\) and its adjoint \(V^\dagger (A^i)^\dagger =(B^i)^\dagger V^\dagger \) move both factors past \(V\) and \(V^\dagger \). For the eigenvalue statement, if \(\mathcal{E}_B(X)=\mu X\) with \(X\neq 0\), then \(\mathcal{E}_A(VXV^\dagger )=\mu VXV^\dagger \), and \(VXV^\dagger \neq 0\) because compression with the isometry gives \(V^\dagger (VXV^\dagger )V=X\).

Lemma E.2.6 Matrix product vectors split along an isometric partition

Let \(V_1,\ldots ,V_r\) be isometries, \(V_k^\dagger V_k=\mathbb {1}\), with \(\sum _kV_kV_k^\dagger =\mathbb {1}\), and suppose \(A^iV_k=V_kB_k^i\) for some tensors \(B_1,\ldots ,B_r\). Then \(\mathcal{V}(A)=\mathcal{V}(\bigoplus _{k=1}^rB_k)\).

Proof

The intertwining relation propagates from letters to words, so \(A^wV_k=V_kB_k^w\) for every word \(w\). Inserting the partition of unity \(\sum _kV_kV_k^\dagger =\mathbb {1}\) into the trace and cycling gives

\begin{align} \operatorname{tr}(A^w) & =\sum _{k=1}^r\operatorname{tr}\! \left(V_k^\dagger A^wV_k\right) =\sum _{k=1}^r\operatorname{tr}(B_k^w), \notag \end{align}

which is the matrix product vector of the direct sum.

Let \(E_k:H_k\to \bigoplus _lH_l\) be the canonical inclusion and let \(B=\bigoplus _lB_l\). Then

\begin{align} E_k^\dagger E_k & =\mathbb {1}, \notag \\ BE_k & =E_kB_k, \notag \\ E_k^\dagger B & =B_kE_k^\dagger , \notag \\ E_k^\dagger BE_k & =B_k, \notag \\ B & =\sum _kE_kB_kE_k^\dagger . \label{eq:canonical_direct_summand} \end{align}

In particular, \(B\) commutes with the range projection \(E_kE_k^\dagger \).

Proof

The identities in (??) follow directly from the defining block-diagonal entries. The relations \(BE_k=E_kB_k\) and \(E_k^\dagger B=B_kE_k^\dagger \) imply that \(B\) commutes with \(E_kE_k^\dagger \).

Lemma E.2.8 Canonical hypotheses descend to dependent direct summands

Suppose that, after a change of coordinates, \(A\) is the dependent direct sum of tensors \(B_k\). If \(A\) has invariant-projector closure and no nontrivial periodic vectors, then every \(B_k\) has both properties. The assertion also applies to summands of dimension zero.

Proof

Conjugate the canonical inclusion of the \(k\)-th summand back to the original bond coordinates of \(A\). This is an isometry which intertwines \(B_k\) with \(A\) on both sides, and its range projection commutes with every letter of \(A\). Projector closure therefore descends by exact compression. A periodic vector of \(B_k\) would embed through the same isometry as a periodic vector of \(A\).

Lemma E.2.9 Closed-chain equality from an exact isometric decomposition

Suppose that isometries \(V_k\) satisfy

\begin{align} V_k^\dagger A^i & =(\mu _kB_k^i)V_k^\dagger , \notag \\ A^i & =\sum _kV_k(\mu _kB_k^i)V_k^\dagger . \label{eq:canonical_exact_reconstruction} \end{align}

Then \(A\) and \(\bigoplus _k\mu _kB_k\) have the same matrix product vectors at every positive length.

Proof

Expand the first letter by the reconstruction formula in (??). In each summand, move the remaining word through \(V_k^\dagger \) using the intertwining relation. Cyclicity of the trace places \(V_k^\dagger V_k=\mathbb {1}\) next to the weighted block word, leaving precisely its closed-chain trace.

E.3 Perron gauges and preservation under conjugation

This section records the positive Perron eigenpair, its spectral normalization, and the behavior of irreducibility and eigenvalues under the corresponding gauge.

Lemma E.3.1 Perron eigenpair of a nonzero irreducible tensor

Let \(B\) be an irreducible MPS tensor of positive bond dimension, not all of whose matrices vanish. Then there are \(\rho {\gt}0\) and \(r{\gt}0\) with \(\mathcal{E}_B(\rho )=r\rho \).

Proof

Irreducibility of the tensor makes the transfer map an irreducible completely positive map, and a nonzero matrix makes it nonzero, so it does not annihilate any nonzero PSD matrix. The Perron–Frobenius theorem for positive maps (Theorem 6.7.2) gives a nonzero PSD eigenvector with eigenvalue \(r{\gt}0\), and irreducibility upgrades it to a positive-definite one ( [ Wol12 , Theorem 6.3(2) ] ).

Lemma E.3.2 Spectral normalization of an irreducible block

Let \(B\) be an irreducible MPS tensor of positive bond dimension with \(\mathcal{E}_B(\rho )=r\rho \) for some \(\rho {\gt}0\) and \(r{\gt}0\), and suppose every eigenvalue of \(\mathcal{E}_B\) of modulus \(r\) equals \(r\). Then \(r^{-1/2}B\) is a normal tensor.

Proof

The rescaling identity \(\mathcal{E}_{r^{-1/2}B}(X)=r^{-1}\mathcal{E}_B(X)\) gives \(\mathcal{E}_{r^{-1/2}B}(\rho )=r^{-1}r\rho =\rho \), so \(\rho \) is a positive-definite fixed point of \(\mathcal{E}_{r^{-1/2}B}\). The rescaled map remains irreducible and completely positive, so Theorem 6.10.1 identifies its spectral radius with the eigenvalue one. The eigenvalues of modulus one of \(\mathcal{E}_{r^{-1/2}B}\) are the eigenvalues of \(\mathcal{E}_B\) of modulus \(r\), divided by \(r\); by hypothesis the only such eigenvalue is one, so the peripheral eigenvalue set is \(\{ 1\} \). Irreducibility is unchanged by a nonzero rescaling.

The support projection \(\pi \) of a positive semidefinite matrix \(\rho \) factors as \(\pi =\rho W\) for a spectral pseudo-inverse \(W\), witnessing that the range of \(\pi \) is the range of \(\rho \). For a rectangular matrix \(Y\), the support projection \(\pi \) of \(YY^\dagger \) fixes \(Y\), \(\pi Y=Y\); and whenever \(AY=YG\) for some matrix \(G\)—that is, \(A\) maps the range of \(Y\) into itself—the one-sided invariance \((\mathbb {1}-\pi )A\pi =0\) holds.

Proof

In the spectral decomposition \(\rho =U\operatorname{diag}(\lambda )U^\dagger \), take \(W=U\operatorname{diag}(\lambda ^{+})U^\dagger \), where \(\lambda ^{+}\) inverts the positive eigenvalues and vanishes on the kernel. Then \(\rho W=U\operatorname{diag}(\mathbf{1}_{\lambda {\gt}0})U^\dagger =\pi \). From \(\pi (YY^\dagger )=YY^\dagger \), the matrix \(((\mathbb {1}-\pi )Y)((\mathbb {1}-\pi )Y)^\dagger \) vanishes, hence \(\pi Y=Y\). Finally,

\begin{align} (\mathbb {1}-\pi )A\pi & =(\mathbb {1}-\pi )AY(Y^\dagger W) =(\mathbb {1}-\pi )Y\, G(Y^\dagger W) =0. \notag \end{align}
Theorem E.3.4 Irreducibility under rescaled conjugation

Let \(B\) be an irreducible tensor, let \(X\) be invertible, and let \(c\ne 0\). Then the rescaled conjugated tensor with letters \(cX^{-1}B^vX\) is irreducible. In particular, the rescaled unital gauge of an irreducible tensor is irreducible; this is the normalization of [ CPGSV16 , Section 2.3 ] applied to an irreducible corner.

Proof

An invariant orthogonal projection \(P\notin \{ 0,\mathbb {1}\} \) of the conjugated tensor makes the columns of \(Y=XP\) span a subspace invariant under every letter of \(B\): \(B^vY=YG_v\), where \(G_v=X^{-1}B^vXP\). The support projection of \(YY^\dagger \) is then a nontrivial invariant orthogonal projection of \(B\) by Lemma E.3.3: it is nonzero because \(Y\ne 0\), and it is not the identity because any nonzero vector in the kernel of \(P\) transports through \((X^\dagger )^{-1}\) to the kernel of \(YY^\dagger \). This contradicts irreducibility of \(B\).

Theorem E.3.5 Unital Schwarz data for an irreducible corner

Let \(B\) be an irreducible tensor with a positive definite transfer eigenvector, \(\mathcal{E}_B(\rho )=r\rho \), where \(\rho {\gt}0\) and \(r{\gt}0\). Then the rescaled unital gauge \(K^v=r^{-1/2}\rho ^{-1/2}B^v\rho ^{1/2}\) is a unital Kraus family, is an irreducible tensor with an irreducible transfer map, and its adjoint map has a positive definite fixed point. These are the hypotheses of the cyclic decomposition of the peripheral spectrum, [ Wol12 , Theorem 6.6 ] .

Proof

Unitality is Theorem 9.8.2; irreducibility is Theorem E.3.4. The adjoint of a unital family is trace-preserving, so its transfer map is a channel and has a positive semidefinite fixed point by the Cesàro argument; irreducibility of the adjoint map upgrades the fixed point to a positive definite one.

Lemma E.3.6 Eigenvalue transport to the unital gauge

In the setting of Theorem E.3.5, every transfer eigenvalue \(\mu \) of \(B\) becomes the transfer eigenvalue \(\mu /r\) of the rescaled unital gauge: the gauge conjugates the transfer map by the congruence with \(\rho ^{1/2}\) and divides it by \(r\). In particular, the peripheral eigenvalues \(e^{2\pi iq/p}\) described in [ CPGSV16 , Section 2.3 ] arise from the transfer eigenvalues of modulus \(r\).

Proof

If \(\mathcal{E}_B(Z)=\mu Z\), then \(Z'=\rho ^{-1/2}Z\rho ^{-1/2}\) is nonzero and satisfies \(\mathcal{E}_K(Z')=(\mu /r)Z'\) by direct computation.

E.4 Blocking identities and primitive stability

These identities control transfer-map powers, weighted direct sums, and positive-length MPV equalities under repeated blocking. They are used in the common-length reduction of Theorem 9.3.1.4.

Blocking by a length \(p\) commutes with weighted direct sums up to the \(p\)th power of each weight: for nonzero weights \((\mu _k)\) and blocks \((B_k)\),

\begin{align} \left(\bigoplus _k\mu _kB_k\right)^{[p]} & \sim \bigoplus _k\mu _k^{\, p}B_k^{[p]}, \label{eq:canonical_block_weight_power} \end{align}

where \(\sim \) is equality of MPV families. When two block families are reblocked to a common length \(p=m_ke_k\), the \(p\)-blocked alphabet is identified with the \(e_k\)-fold tensor power of the \(m_k\)-blocked alphabet, so each sector reblocked through its own period agrees with the directly \(p\)-blocked block.

Theorem E.4.1 Common blocking period for a block family

Let \((B_k)_{k=1}^r\) be a finite family of positive-bond-dimension MPS tensors, each admitting a blocking period \(p_k\) making its transfer map primitive. Then \(p=\operatorname{lcm}(p_1,\ldots ,p_r)\) is a common period: \(\mathcal{E}_{B_k^{[p]}}\) is primitive for every \(k\).

Proof

Each \(p_k\) divides \(p\), so primitivity of \(\mathcal{E}_{B_k^{[p_k]}}\) carries over to \(\mathcal{E}_{B_k^{[p]}}\) via the divisibility identity for transfer-map powers.

Theorem E.4.2 Common primitive blocking for irreducible left-canonical blocks

Let \((B_k)_{k=1}^r\) be a finite family of positive-bond-dimension, left-canonical, tensor-irreducible MPS tensors. Then there is a single blocking length \(p{\gt}0\) such that \(\mathcal{E}_{B_k^{[p]}}\) is primitive for every \(k\).

Proof

Apply Theorem 9.9.2.1 to each block, then apply Theorem E.4.1 to the resulting finite family of positive periods. Their least common multiple is positive and makes every blocked transfer map primitive.

Lemma E.4.3 Adjoint primitivity equivalence
#

Let \(\mathcal{E}\) be a finite-dimensional completely positive map. Then \(\mathcal{E}\) is primitive if and only if its Frobenius adjoint \(\mathcal{E}^\dagger \) is primitive. Equivalently, the eigenvalues \(\lambda \) of \(\mathcal{E}\) with \(|\lambda |=1\) form the singleton \(\{ 1\} \) if and only if the same holds for \(\mathcal{E}^\dagger \), because adjunction sends every eigenvalue to its complex conjugate.

Lemma E.4.4 Primitivity under blocking by a multiple

Let \(A\) have positive bond dimension and let \(p_0,p\ge 1\) with \(p_0\mid p\). If \(\mathcal{E}_{A^{[p_0]}}\) is primitive, then \(\mathcal{E}_{A^{[p]}}\) is primitive.

Let \(A\) be left-canonical, tensor-irreducible, and have primitive transfer map. Then \(A\) has primitive fixed-point data \((\rho ,P_\rho ,N)\) with \(\rho {\gt}0\) and \(N^n\to 0\). For every blocking length \(P\), the tensor \(A^{[P]}\) is left-canonical, its transfer map is \(\mathcal{E}_A^P\), and \(\rho \) is fixed by this blocked map. For every \(n\ge 1\),

\begin{align} \mathcal{E}_A^n & =P_\rho +N^n. \label{eq:canonical_primitive_fixed_point_power} \end{align}

Finally, a channel with a positive-definite fixed point whose positive semidefinite fixed points are all proportional is irreducible as a map; irreducibility of a transfer map implies tensor irreducibility.

Proof

Peripheral primitivity, tensor irreducibility, and left-canonical normalization first give primitive fixed-point data, and tensor irreducibility makes its positive semidefinite fixed point positive definite. Blocking preserves the left-canonical identity, replaces the transfer map by its positive power, and therefore preserves the fixed point. The defining complementary-power identity gives (??). The channel fixed-point criterion then gives map irreducibility from uniqueness of positive semidefinite fixed points, and the final implication converts map irreducibility to tensor irreducibility.

Let \(A\) be left-canonical, irreducible, primitive, of positive bond dimension. For every \(P{\gt}0\), the blocked tensor \(A^{[P]}\) is irreducible: there is no orthogonal projection \(Q\) with \(Q\neq 0\), \(Q\neq \mathbb {1}\) such that \((\mathbb {1}-Q)(A^{[P]})^\alpha Q=0\) for every \(P\)-blocked index \(\alpha \).

Proof

Lemma E.4.5 gives primitive fixed-point data with \(\rho {\gt}0\). In particular, if \(P_\rho (X)=\operatorname{tr}(X)\rho /\operatorname{tr}(\rho )\) and \(N:=\mathcal{E}_A-P_\rho \), then

\begin{align} \mathcal{E}_A & =P_\rho +N, \notag \\ N^n & \longrightarrow 0. \notag \end{align}

The primitive fixed-point theorem supplies this complementary-power convergence.

The blocked transfer map is the corresponding positive power:

\begin{align} \mathcal{E}_{A^{[P]}}(X) & =\mathcal{E}_A^P(X). \label{eq:canonical_block_transfer_power} \end{align}

Hence \(\rho \) remains a positive-definite fixed point after blocking. Let \(\sigma \ge 0\) be any fixed point of \(\mathcal{E}_A^P\), set

\begin{align} c & =\frac{\operatorname{tr}(\sigma )}{\operatorname{tr}(\rho )}, \notag \\ \sigma ’ & =\sigma -c\rho . \notag \end{align}

Then \(\operatorname{tr}(\sigma ')=0\), so \(P_\rho (\sigma ')=0\), while \(\mathcal{E}_A^P(\sigma ')=\sigma '\). Since the fixed-point projection and the complement annihilate one another, iteration gives, for every \(k\ge 1\),

\begin{align} \sigma ’ & =\mathcal{E}_A^{Pk}(\sigma ’) =N^{Pk}(\sigma ’). \notag \end{align}

But \(N^{Pk}(\sigma ')\to 0\) as \(k\to \infty \), hence \(\sigma '=0\). Thus every positive semidefinite fixed point of the blocked map is proportional to \(\rho \).

Blocking preserves the left-canonical identity, so Theorem 4.7.2 makes the blocked transfer map a channel. Its positive-definite fixed point and the uniqueness just proved imply map irreducibility by Lemma E.4.5. The final implication in that lemma then shows that \(A^{[P]}\) is tensor-irreducible.

Theorem E.4.7 Extra blocking of primitive irreducible sector blocks

Let \(A\) have positive bond dimension and be trace-preserving, primitive, and tensor-irreducible. For every \(k{\gt}0\),

\begin{align} \sum _\alpha ((A^{[k]})^\alpha )^\dagger (A^{[k]})^\alpha & =\mathbb {1}, \notag \\ \sigma _\partial (\mathcal{E}_{A^{[k]}}) & =\{ 1\} , \label{eq:canonical_extra_blocking} \end{align}

and \(A^{[k]}\) is tensor-irreducible. This \(k\) is a later common-refinement or injectivity length, distinct from the period-removal length that produced the sector block.

Proof

Trace preservation follows from blocking a trace-preserving tensor, primitivity from the transfer-power identity (??), and irreducibility from Theorem E.4.6.

Lemma E.4.8 Blocking the weighted block sum

For nonzero weights \((\mu _k)_{k=1}^r\), blocks \((B_k)_{k=1}^r\), and any \(p\ge 1\), the blocking identity (??) holds, where \(\sim \) denotes generation of the same MPV family at every length.

Lemma E.4.9 Positive-length equality survives blocking

If two tensors have the same MPV coefficients at every positive length, then so do their blocked tensors after any positive blocking length.

Lemma E.4.10 Positive-length blocking of a weighted block sum

If \(A\) and \(\bigoplus _k\mu _kB_k\) have the same MPV coefficients at every positive length, then after any positive blocking length \(p\), \(A^{[p]}\) and \(\bigoplus _k\mu _k^{\, p}B_k^{[p]}\) have the same MPV coefficients at every positive length.

Lemma E.4.11 Positive-length equality of powered block sums

If \(\bigoplus _j\mu _j^AA_j\) and \(\bigoplus _k\mu _k^BB_k\) have the same MPV coefficients at every positive length, then after any positive common blocking length \(p\), \(\bigoplus _j(\mu _j^A)^pA_j^{[p]}\) and \(\bigoplus _k(\mu _k^B)^pB_k^{[p]}\) have the same MPV coefficients at every positive length.

Lemma E.4.12 Nonzero weights survive blocking
#

If \(\mu \in \mathbb {C}\) is nonzero and \(p\ge 1\), then \(\mu ^p\neq 0\); blocking preserves the nonzero-weight hypothesis of a weighted block decomposition.

Lemma E.4.13 Iterated blocking at the transfer-map level

For any MPS tensor \(A\) and \(m,n\ge 0\),

\begin{align} \mathcal{E}_{(A^{[m]})^{[n]}} & = \mathcal{E}_{A^{[mn]}}. \label{eq:canonical_iterated_blocking} \end{align}

E.5 Support compression and cyclic-sector families

Support isometries first realize the cyclic corners. The remaining results place finitely many cyclic-sector families over one blocked alphabet and identify their weighted nonzero parts.

Theorem E.5.1 Supported corner compression with a support isometry

Let \(A\) be a tensor normalized on the support of an orthogonal projection \(P\), so that \(PA^iP=A^i\) and \(\sum _i(A^i)^\dagger A^i=P\). Then there are a corner tensor \(C\), a linear isomorphism \(\varphi :M_{\dim C}(\mathbb {C})\xrightarrow {\sim }PM_{D}(\mathbb {C})P\), and a rectangular isometry \(V:\mathbb {C}^{\dim C}\to \mathbb {C}^D\) such that

\begin{align} V^\dagger V & =\mathbb {1}, \notag \\ VV^\dagger & =P, \notag \\ \varphi (X) & =VXV^\dagger . \label{eq:canonical_support_isometry} \end{align}

Moreover, \(\varphi (C^i)=A^i\) for every physical letter, and for every word \(\sigma =(\sigma _1,\ldots ,\sigma _N)\),

\begin{align} V^{(N)}(C)_\sigma & = \operatorname{tr}\! (PA^{\sigma _1}\cdots A^{\sigma _N}). \label{eq:canonical_corner_trace} \end{align}
Proof

Diagonalize the projection \(P\) by a unitary change of basis. The inclusion of the support coordinates gives a rectangular isometry \(V\) satisfying (??). Compress the letters of \(A\) to this support and let \(\varphi \) be the corresponding corner isomorphism. The support relation \(PA^iP=A^i\) gives the corner-letter identity. For each word \(\sigma =(\sigma _1,\ldots ,\sigma _N)\), cyclicity of the trace and \(VV^\dagger =P\) give

\begin{align} V^{(N)}(C)_\sigma & = \operatorname{tr}(C^{\sigma _1}\cdots C^{\sigma _N}) = \operatorname{tr}(V^\dagger A^{\sigma _1}\cdots A^{\sigma _N}V) = \operatorname{tr}(PA^{\sigma _1}\cdots A^{\sigma _N}), \notag \end{align}

which proves (??).

Theorem E.5.2 Commuting projections give sector isometries

Let a left-canonical tensor \(A\) commute with a family of orthogonal projections \((P_k)\) summing to \(\mathbb {1}\). Then \(A\) decomposes into compressed sector tensors \((C_k)\), with a unit-weight direct-sum MPV decomposition, corner isomorphisms \(\varphi _k:M_{\dim C_k}(\mathbb {C})\xrightarrow {\sim }P_kM_{D}(\mathbb {C})P_k\), and rectangular isometries \(V_k\) satisfying

\begin{align} V_k^\dagger V_k & =\mathbb {1}, \notag \\ V_kV_k^\dagger & =P_k, \notag \\ \varphi _k(X) & =V_kXV_k^\dagger . \label{eq:canonical_sector_isometries} \end{align}

The sector letters obey \(\varphi _k(C_k^i)=P_kA^iP_k\).

Proof

Apply Theorem E.5.1 to the compression of \(A\) on each sector \(P_k\). Orthogonality and \(\sum _kP_k=\mathbb {1}\) give the direct-sum decomposition. For every word \(\sigma =(\sigma _1,\ldots ,\sigma _N)\), the partition of unity and (??) give

\begin{align} V^{(N)}(A)_\sigma & = \operatorname{tr}\! \left(\left(\sum _kP_k\right) A^{\sigma _1}\cdots A^{\sigma _N}\right) = \sum _k\operatorname{tr}(P_kA^{\sigma _1}\cdots A^{\sigma _N}) = \sum _kV^{(N)}(C_k)_\sigma . \notag \end{align}

This is the unit-weight matrix-product-vector identity.

Theorem E.5.3 Adjoint-fixed projections give sector isometries

Let \(A\) be left-canonical and let \((P_k)\) be orthogonal projections summing to \(\mathbb {1}\) and fixed by the adjoint transfer map. Then the same sector decomposition and support isometries of Theorem E.5.2 exist. In particular, the adjoint fixed-point condition \(\mathcal{E}_A^\dagger (P_k)=P_k\) gives \(P_kA^i=A^iP_k\), which gives the corner-letter identity \(\varphi _k(C_k^i)=P_kA^iP_k\).

Proof

The fixed-point identity \(\mathcal{E}_A^\dagger (P_k)=P_k\), together with left-canonicality \(\sum _i(A^i)^\dagger A^i=\mathbb {1}\), places \(P_k\) in the multiplicative domain of \(\mathcal{E}_A^\dagger \), forcing \(P_kA^i=A^iP_k\) for every letter \(A^i\). The result then follows from Theorem E.5.2.

If \(A\) and \(B\) generate the same MPV family at every positive length, then after removing the all-zero blocks and applying the trace-preserving gauge there are weighted nonzero-block tensors \(\bigoplus _j\mu _j^AA_j\) and \(\bigoplus _k\mu _k^BB_k\) such that, for every \(N\ge 1\) and \(\sigma \),

\begin{align} V^{(N)}(A)_\sigma & = V^{(N)}\! \left(\bigoplus _j\mu _j^AA_j\right)_\sigma , \label{eq:canonical_nonzero_part_a} \\ V^{(N)}(B)_\sigma & = V^{(N)}\! \left(\bigoplus _k\mu _k^BB_k\right)_\sigma . \label{eq:canonical_nonzero_part_b} \end{align}

The two nonzero parts agree:

\begin{align} V^{(N)}\! \left(\bigoplus _j\mu _j^AA_j\right)_\sigma & = V^{(N)}\! \left(\bigoplus _k\mu _k^BB_k\right)_\sigma . \label{eq:canonical_nonzero_parts_equal} \end{align}

Each nonzero block has primitive irreducible cyclic sectors: for each \(j\), there is \(m_j\ge 1\) with a unit-weight sector decomposition \(A_j^{[m_j]}\sim \bigoplus _{u=1}^{m_j}C^A_{j,u}\), and similarly for \(B_k\).

Proof

Apply Theorem 9.2.1.1 separately to \(A\) and \(B\). Its positive-length conclusion gives (??) and (??). Chaining these identities through \(\mathcal{V}^+(A)=\mathcal{V}^+(B)\) gives (??). Finally apply Theorem 9.3.1.3 to every trace-preserving irreducible nonzero-weight block.

Lemma E.5.5 Common reblocking of cyclic-sector families

Given a finite family \((B_k)_{k=1}^r\), each with primitive irreducible cyclic sectors, there is a common blocking length \(p\) such that each \(B_k^{[p]}\) admits a unit-weight direct-sum decomposition over a single common blocked alphabet \(d^{[p]}\) into left-canonical blocks with primitive transfer maps, tensor irreducibility, and positive bond dimensions.

Lemma E.5.6 Common reblocking at a prescribed multiple

With hypotheses as in Lemma E.5.5, and given any common multiple \(p\) of the per-block period-removal lengths \((m_k)\), the common reblocking can be performed at the prescribed length \(p\).

Lemma E.5.7 Structural properties of the common reblocked sectors

Each sector tensor in the common reblocked cyclic-sector family of Lemma E.5.5 is trace-preserving, \(\sum _i(B^i)^\dagger B^i=\mathbb {1}\), has primitive transfer map, \(\sigma _\partial (\mathcal{E}_B)=\{ 1\} \), is tensor-irreducible, and has positive bond dimension.

Lemma E.5.8 Common-alphabet flattening

With \(p=m_ke_k\) for each \(k\), the iterated blocked tensor \((B_k^{[m_k]})^{[e_k]}\) agrees with \(B_k^{[p]}\) in MPV family, after the canonical identification of the \(p\)-blocked alphabet \(d^{[p]}\) with the \(e_k\)-fold tensor power of the \(m_k\)-blocked alphabet \(d^{[m_k]}\).

Lemma E.5.9 Direct blocking in the common alphabet

For every original block \(B_k\), direct blocking at the common length \(p=m_ke_k\) agrees, as an MPV family, with the same block transported to the common blocked alphabet. Write \(\widetilde{B}_k\) for this common-alphabet image. Then

\begin{align} V^{(N)}(B_k^{[p]})_\sigma & = V^{(N)}(\widetilde{B}_k)_\sigma . \notag \end{align}
Lemma E.5.10 Reindexed nonzero-part identity

The powered nonzero-part decomposition \(\bigoplus _k\mu _k^pB_k^{[p]}\) carries over under the common-alphabet identification of Lemma E.5.8: the common-alphabet sectors have the same nonzero-part decomposition as the direct \(p\)-blocked tensors.

E.6 Prepared normal-form consequences

The first theorem derives the overlap condition from a prepared primitive family. The second does not construct primitive blocks from an arbitrary tensor: it only orders a family that already satisfies all primitive, irreducible, trace-preserving, nonzero-weight, and positive-dimension hypotheses.

Theorem E.6.1 Self-overlap is derived in normal canonical form

If \((\mu _k,A_k)\) satisfies the normal canonical form conditions, then \(O_{A_kA_k}(N)\to 1\) for every block \(k\).

Proof

Fix a block \(k\). Irreducibility and the TP normalization give a nonzero positive semidefinite fixed point \(\rho \) of \(\mathcal{E}_{A_k}\) with fixed-point projector \(P\); primitivity gives the complementary gap \(\rho _{\operatorname{spec}}(\mathcal{E}_{A_k}-P){\lt}1\) (Theorem 4.11.3). Theorem 7.11.2 then yields \(O_{A_kA_k}(N)\to 1\).

Theorem E.6.2 Reordering prepared primitive blocks into normal canonical form

Suppose \(A\) generates the same MPV family as \(\bigoplus _k\mu _kA_k\), and that each \(A_k\) is irreducible, satisfies \(\sum _i(A_k^i)^\dagger A_k^i=\mathbb {1}\), and has primitive transfer map, all \(\mu _k\neq 0\), and all bond dimensions are positive. Then there exist a blocking length \(p{\gt}0\), weights \((\nu _k)\), and blocked tensors \((B_k)\) such that \(A^{[p]}\) generates the same MPV family as \(\bigoplus _k\nu _kB_k\), and \((\nu _k,B_k)\) satisfies the normal canonical form conditions. The weight moduli need not be distinct; equal-modulus blocks are allowed.

Proof

The given blocks are already primitive, so take \(p=1\). Reorder the blocks by non-increasing weight modulus, with ties allowed, and set \(\nu _k:=\mu _k\) and \(B_k:=A_k\) in that order. Irreducibility, TP normalization, primitivity, nonzero weights, and positive bond dimensions are preserved, so the reordered family satisfies Definition 9.9.2 and generates the same MPV family as \(A\).

Remark E.6.3 Scope of Theorem E.6.2
#

Theorem E.6.2 assumes all six block hypotheses in advance and produces the normal canonical form by reordering. It is not the arbitrary-input existence theorem of  [ CPGSV16 , Section 2.3 ] ; that primitive-block reduction, including its structural bond-dimension bound, is Theorem 9.11.1.2 in Section 9.11.1.