Matrix-Product Unitaries (MPUs)

• Matrix-Product Unitaries (MPUs) are one-dimensional tensor-network representations of many-body unitary operators that preserve locality and adhere to an entanglement area law.
• They employ transfer matrices and canonical forms to realize structured quantum cellular automata, facilitating symmetry actions and locality-preserving dynamics.
• Recent developments extend MPUs to nonuniform, open-boundary, and fermionic setups, highlighting their versatility in modeling complex quantum evolution.

Searching arXiv for recent and foundational papers on matrix-product unitaries. Matrix-product unitaries (MPUs) are one-dimensional tensor-network representations of many-body unitary operators. In their basic form, an MPU is a matrix-product operator whose contraction yields a unitary on an $N$-site chain, typically for every system size $N$; more general formulations allow open boundaries and site-dependent tensors. Because the operator is assembled by contracting nearest-neighbor bond indices, MPUs provide a natural language for locality-preserving dynamics, quantum cellular automata (QCA), symmetry actions, and tensor-network descriptions of quantum evolution, while still accommodating phenomena that go beyond strict finite-depth circuit structure (Cirac et al., 2017, Styliaris et al., 2024).

1. Definition and basic tensor-network structure

For a qudit chain with local Hilbert spaces $\mathcal H_j\cong \mathbb C^d$, an MPU is given by an MPO of the form

$U\;=\;\sum_{\{i_k,j_k\}=0}^{d-1} \Tr\bigl(M^{[1]\,i_1,j_1}\,M^{[2]\,i_2,j_2}\,\cdots\,M^{[N]\,i_N,j_N}\bigr)\; \ket{i_1\cdots i_N}\bra{j_1\cdots j_N}\,,$

where each $M^{[k]\,i_k,j_k}$ is a $D\times D$ matrix on the auxiliary space, the indices $i_k,j_k$ are the physical input/output labels, and the bond indices are contracted along the chain. In the periodic, translation-invariant setting one usually repeats a single local tensor, while in more general formulations one allows a boundary matrix $b$ and site-dependent tensors $A_k$, writing

$U \;=\; \sum_{i_1,\dots,i_N\atop j_1,\dots,j_N} \Tr\bigl(b\;A_N^{\,i_N j_N}\cdots A_1^{\,i_1 j_1}\bigr)\, \bigl|\,i_1\cdots i_N\bigr\rangle\bigl\langle j_1\cdots j_N\bigr|\,.$

Unitarity requires $N$0, and in the original uniform setting one demands this for every $N$1 (Cirac et al., 2017, Shukla, 1 Feb 2025, Styliaris et al., 2024).

The tensor-network structure immediately bounds operator entanglement. For an $N$2-site MPU with bond dimension $N$3, the operator Schmidt rank across any cut is at most $N$4, so MPUs preserve the one-dimensional entanglement area law. This area-law property is structural rather than algorithmic: the individual local tensors need not themselves be unitary maps on the bond legs, which is precisely why converting an abstract MPU into an explicit quantum circuit is nontrivial (Styliaris et al., 11 Aug 2025).

Several conventions coexist in the literature. Foundational structure theorems treat a uniform tensor repeated on a ring and unitary for all system sizes; later work studies open-boundary MPUs built from a repeated bulk tensor with boundary vectors, and also fully nonuniform MPUs with site-dependent tensors. This distinction matters because strict locality properties, classification statements, and circuit realizability depend sensitively on whether one is in the uniform periodic, open-boundary, nonuniform, or fermionic setting (Şahinoğlu et al., 2017, Styliaris et al., 2024).

2. Locality-preserving structure and the QCA correspondence

In the qudit setting with a single repeated tensor and periodic boundary conditions, MPUs coincide with one-dimensional locality-preserving unitaries, i.e. QCAs, after blocking a finite number of sites. Equivalently, every such MPU has a strict causal cone: for any operator $N$5 supported on site $N$6, the Heisenberg-evolved operator $N$7 is supported within a bounded neighborhood $N$8, with $N$9 independent of system size. Conversely, every finite-bond-dimension 1D QCA can be represented as an MPU after blocking (Cirac et al., 2017, Şahinoğlu et al., 2017).

After sufficient blocking, a simple MPU can be brought into a standard two-layer form

$\mathcal H_j\cong \mathbb C^d$0

where $\mathcal H_j\cong \mathbb C^d$1 and $\mathcal H_j\cong \mathbb C^d$2 are fixed nearest-neighbor unitaries acting on blocked degrees of freedom. In the equivalent fixed-point formulation, the blocked tensor $\mathcal H_j\cong \mathbb C^d$3 satisfies separation and isometry conditions involving positive left and right fixed points $\mathcal H_j\cong \mathbb C^d$4: $\mathcal H_j\cong \mathbb C^d$5 These conditions encode the pull-through structure responsible for strict support preservation (Şahinoğlu et al., 2017).

The QCA classification enters through an index. In the standard-form approach one defines unfolding ranks $\mathcal H_j\cong \mathbb C^d$6 and $\mathcal H_j\cong \mathbb C^d$7, with $\mathcal H_j\cong \mathbb C^d$8, and sets

$\mathcal H_j\cong \mathbb C^d$9

Equivalent formulations express the GNVW index as a positive rational number obtained from left and right fixed points or from a rank ratio. The index is invariant under blocking, additive under composition and tensor product, and constant under continuous deformations preserving unitarity. Two MPUs are equivalent, after possibly adding ancillas and blocking, if and only if they have the same index (Cirac et al., 2017, Şahinoğlu et al., 2017).

A standard misconception is that the statement “MPUs are QCAs” holds without qualification. The available results support that statement for the uniform periodic qudit setting. Later work shows that it fails for open-boundary MPUs and for fermionic MPUs unless the formalism is generalized appropriately (Styliaris et al., 2024, Piroli et al., 2020).

3. Canonical forms, transfer matrices, and MPU generation criteria

Canonical-form theory for MPUs parallels the canonical form of matrix product states. A tensor can be gauge-transformed into canonical form II so that its transfer operator has left fixed point $U\;=\;\sum_{\{i_k,j_k\}=0}^{d-1} \Tr\bigl(M^{[1]\,i_1,j_1}\,M^{[2]\,i_2,j_2}\,\cdots\,M^{[N]\,i_N,j_N}\bigr)\; \ket{i_1\cdots i_N}\bra{j_1\cdots j_N}\,,$0 and right fixed point a positive invertible $U\;=\;\sum_{\{i_k,j_k\}=0}^{d-1} \Tr\bigl(M^{[1]\,i_1,j_1}\,M^{[2]\,i_2,j_2}\,\cdots\,M^{[N]\,i_N,j_N}\bigr)\; \ket{i_1\cdots i_N}\bra{j_1\cdots j_N}\,,$1. After finite blocking every MPU becomes simple, and in the simple case the local tensor admits a standard-form description that is unique up to a local gauge on the bond spaces. The corresponding fundamental theorem states that two simple MPU tensors generate the same global unitary if and only if their local unitaries differ by such a gauge transformation (Cirac et al., 2017, Franco-Rubio et al., 27 Feb 2025).

A more recent formulation gives a direct transfer-matrix criterion for determining when a local tensor generates a unitary MPO of a given size. For a local tensor $U\;=\;\sum_{\{i_k,j_k\}=0}^{d-1} \Tr\bigl(M^{[1]\,i_1,j_1}\,M^{[2]\,i_2,j_2}\,\cdots\,M^{[N]\,i_N,j_N}\bigr)\; \ket{i_1\cdots i_N}\bra{j_1\cdots j_N}\,,$2, define $U\;=\;\sum_{\{i_k,j_k\}=0}^{d-1} \Tr\bigl(M^{[1]\,i_1,j_1}\,M^{[2]\,i_2,j_2}\,\cdots\,M^{[N]\,i_N,j_N}\bigr)\; \ket{i_1\cdots i_N}\bra{j_1\cdots j_N}\,,$3, and let $U\;=\;\sum_{\{i_k,j_k\}=0}^{d-1} \Tr\bigl(M^{[1]\,i_1,j_1}\,M^{[2]\,i_2,j_2}\,\cdots\,M^{[N]\,i_N,j_N}\bigr)\; \ket{i_1\cdots i_N}\bra{j_1\cdots j_N}\,,$4 and $U\;=\;\sum_{\{i_k,j_k\}=0}^{d-1} \Tr\bigl(M^{[1]\,i_1,j_1}\,M^{[2]\,i_2,j_2}\,\cdots\,M^{[N]\,i_N,j_N}\bigr)\; \ket{i_1\cdots i_N}\bra{j_1\cdots j_N}\,,$5 be the associated transfer matrices. Then $U\;=\;\sum_{\{i_k,j_k\}=0}^{d-1} \Tr\bigl(M^{[1]\,i_1,j_1}\,M^{[2]\,i_2,j_2}\,\cdots\,M^{[N]\,i_N,j_N}\bigr)\; \ket{i_1\cdots i_N}\bra{j_1\cdots j_N}\,,$6 generates a unitary MPO of length $U\;=\;\sum_{\{i_k,j_k\}=0}^{d-1} \Tr\bigl(M^{[1]\,i_1,j_1}\,M^{[2]\,i_2,j_2}\,\cdots\,M^{[N]\,i_N,j_N}\bigr)\; \ket{i_1\cdots i_N}\bra{j_1\cdots j_N}\,,$7 if and only if

$U\;=\;\sum_{\{i_k,j_k\}=0}^{d-1} \Tr\bigl(M^{[1]\,i_1,j_1}\,M^{[2]\,i_2,j_2}\,\cdots\,M^{[N]\,i_N,j_N}\bigr)\; \ket{i_1\cdots i_N}\bra{j_1\cdots j_N}\,,$8

This criterion replaces earlier fixed-point checks by a trace condition on transfer matrices and applies to all uniform MPUs of size $U\;=\;\sum_{\{i_k,j_k\}=0}^{d-1} \Tr\bigl(M^{[1]\,i_1,j_1}\,M^{[2]\,i_2,j_2}\,\cdots\,M^{[N]\,i_N,j_N}\bigr)\; \ket{i_1\cdots i_N}\bra{j_1\cdots j_N}\,,$9 (Shukla, 1 Feb 2025).

When generation holds for sufficiently many sizes, transfer-matrix spectra rigidify. Specifically, if $M^{[k]\,i_k,j_k}$0 generates an MPU for all $M^{[k]\,i_k,j_k}$1 up to $M^{[k]\,i_k,j_k}$2, then both $M^{[k]\,i_k,j_k}$3 and $M^{[k]\,i_k,j_k}$4 have exactly one nonzero eigenvalue, equal to $M^{[k]\,i_k,j_k}$5. Consequently, for some $M^{[k]\,i_k,j_k}$6,

$M^{[k]\,i_k,j_k}$7

and this rank-one structure yields a finite causal cone. In that sense, locality preservation emerges naturally from all-$M^{[k]\,i_k,j_k}$8 generation (Shukla, 1 Feb 2025).

The transfer-matrix viewpoint also clarifies “unitary only for selected sizes” phenomena. One example has $M^{[k]\,i_k,j_k}$9 but $D\times D$0, so

$D\times D$1

The corresponding MPO is unitary only for odd $D\times D$2 and is not locality-preserving. Earlier work described analogous odd-chain examples as carrying a formal “fractional” rank-ratio index outside the original GNVW classification. These examples delimit the regime in which standard MPU-QCA equivalence holds (Shukla, 1 Feb 2025, Şahinoğlu et al., 2017).

4. Fermionic MPUs and generalized auxiliary traces

For fermionic chains, the tensor-network formalism must respect the global $D\times D$3 grading. A local fermionic tensor $D\times D$4 carries parity

$D\times D$5

and consistency requires each local tensor to be even: $D\times D$6 with $D\times D$7 the parity operator on the auxiliary space. Composition of fermionic MPOs introduces additional sign factors coming from graded reordering, so many bosonic proofs do not transfer unchanged (Piroli et al., 2020).

The fermionic case differs sharply from the qudit case in two ways. First, fermionic MPUs with periodic boundary conditions do not necessarily have a strict causal cone; explicit examples have transfer matrices with additional nonzero subleading eigenvalues and can move information arbitrarily far. Second, not all fermionic QCAs can be represented as ordinary fermionic MPUs. The canonical counterexample is Majorana-mode translation, $D\times D$8, whose correct tensor representation requires inserting an odd operator $D\times D$9 into the auxiliary trace rather than using the plain periodic trace (Piroli et al., 2020).

This motivates generalized fermionic MPUs of the form

$i_k,j_k$0

where $i_k,j_k$1 has definite parity. Two cases are distinguished. Type I uses $i_k,j_k$2. Type II uses

$i_k,j_k$3

and captures Majorana translation. Locality preservation is characterized by a graded analogue of simpleness, requiring for blocked tensors that the transfer matrix satisfies

$i_k,j_k$4

Up to appending inert ancillary fermionic degrees of freedom, every representative of this family is a fermionic QCA and vice versa (Piroli et al., 2020).

The associated fermionic index theorem defines maps $i_k,j_k$5 and $i_k,j_k$6, sets

$i_k,j_k$7

and proves that $i_k,j_k$8 is rational for Type I and $i_k,j_k$9 times rational for Type II, invariant under blocking and inert ancillas, additive under composition and tensor product, locally constant under continuous deformations, and complete for equivalence up to circuits, blocking, and ancillas. A graded canonical form for fermionic MPS, unique up to even gauge transformations and permutation of blocks, underlies the proof (Piroli et al., 2020).

5. Symmetries, cohomological data, and gauging

For a finite group $b$0 with on-site representation $b$1, a $b$2-symmetric MPU satisfies

$b$3

In the standard form, symmetry is implemented on the virtual spaces by projective representations $b$4 and $b$5, obeying

$b$6

with $b$7 a $b$8-valued $b$9-cocycle. Under equivalence generated by blocking, adjoining symmetric ancillas in arbitrary representations, and continuous symmetric deformation, symmetric MPUs are classified by exactly two invariants: the chiral index and the cohomology class $A_k$0. If the on-site representation is fixed, one obtains additional symmetry-protected indices

$A_k$1

which refine the classification and are stable against disorder (Gong et al., 2018).

A complementary operator-algebraic framework treats MPUs as locally decomposable group actions on the quasi-local algebra. In that language, one writes each action as a finite-link unitary conjugation composed with left- and right-half-chain automorphisms, extracts unitaries $A_k$2 on a half-chain, and obtains a scalar obstruction $A_k$3 satisfying the pentagon identity, hence defining a class

$A_k$4

When the action preserves split states $A_k$5, one also obtains twisted $A_k$6-cocycles $A_k$7. The pair of cohomological data $A_k$8 furnishes a complete invariant in that setting, and the resulting dictionary was worked out explicitly against the finite-size MPS/MPO formalism (Rubio et al., 2024).

MPU symmetries of matrix product states admit a defect formalism extending the well-known on-site case. Truncating a global MPU creates defect tensors at the endpoints of the truncation; these defects can be moved by local unitaries

$A_k$9

and fused by local fusion operators $U \;=\; \sum_{i_1,\dots,i_N\atop j_1,\dots,j_N} \Tr\bigl(b\;A_N^{\,i_N j_N}\cdots A_1^{\,i_1 j_1}\bigr)\, \bigl|\,i_1\cdots i_N\bigr\rangle\bigl\langle j_1\cdots j_N\bigr|\,.$0. Under the block-independence condition, one can gauge the symmetry by promoting defects to gauge degrees of freedom on bonds, obtaining an MPS of the same bond dimension with commuting local Gauss-law projectors. If block independence fails, a modified state-level gauging still produces a locally symmetric MPS, but the resulting gauge constraints need not commute on different sites (Franco-Rubio et al., 27 Feb 2025).

These symmetry results show that MPUs support several distinct kinds of topological data. In the locality-preserving on-site symmetric setting, the invariants are an index and a $U \;=\; \sum_{i_1,\dots,i_N\atop j_1,\dots,j_N} \Tr\bigl(b\;A_N^{\,i_N j_N}\cdots A_1^{\,i_1 j_1}\bigr)\, \bigl|\,i_1\cdots i_N\bigr\rangle\bigl\langle j_1\cdots j_N\bigr|\,.$1-cocycle class; in the operator-algebraic group-action setting, a $U \;=\; \sum_{i_1,\dots,i_N\atop j_1,\dots,j_N} \Tr\bigl(b\;A_N^{\,i_N j_N}\cdots A_1^{\,i_1 j_1}\bigr)\, \bigl|\,i_1\cdots i_N\bigr\rangle\bigl\langle j_1\cdots j_N\bigr|\,.$2-cocycle class and twisted $U \;=\; \sum_{i_1,\dots,i_N\atop j_1,\dots,j_N} \Tr\bigl(b\;A_N^{\,i_N j_N}\cdots A_1^{\,i_1 j_1}\bigr)\, \bigl|\,i_1\cdots i_N\bigr\rangle\bigl\langle j_1\cdots j_N\bigr|\,.$3-cocycles appear naturally; and in MPS applications, these invariants control defect fusion and gauging. A plausible implication is that “MPU symmetry” is best understood as a hierarchy of structures rather than a single classification problem.

6. Beyond strict QCA: boundaries, circuit complexity, and random ensembles

Open-boundary and nonuniform MPUs exhibit behavior not captured by the original QCA correspondence. A translation-invariant open-boundary MPU can fail to preserve locality exactly. The canonical example is the multi-control-$U \;=\; \sum_{i_1,\dots,i_N\atop j_1,\dots,j_N} \Tr\bigl(b\;A_N^{\,i_N j_N}\cdots A_1^{\,i_1 j_1}\bigr)\, \bigl|\,i_1\cdots i_N\bigr\rangle\bigl\langle j_1\cdots j_N\bigr|\,.$4 gate on $U \;=\; \sum_{i_1,\dots,i_N\atop j_1,\dots,j_N} \Tr\bigl(b\;A_N^{\,i_N j_N}\cdots A_1^{\,i_1 j_1}\bigr)\, \bigl|\,i_1\cdots i_N\bigr\rangle\bigl\langle j_1\cdots j_N\bigr|\,.$5 qubits,

$U \;=\; \sum_{i_1,\dots,i_N\atop j_1,\dots,j_N} \Tr\bigl(b\;A_N^{\,i_N j_N}\cdots A_1^{\,i_1 j_1}\bigr)\, \bigl|\,i_1\cdots i_N\bigr\rangle\bigl\langle j_1\cdots j_N\bigr|\,.$6

which has operator-Schmidt rank $U \;=\; \sum_{i_1,\dots,i_N\atop j_1,\dots,j_N} \Tr\bigl(b\;A_N^{\,i_N j_N}\cdots A_1^{\,i_1 j_1}\bigr)\, \bigl|\,i_1\cdots i_N\bigr\rangle\bigl\langle j_1\cdots j_N\bigr|\,.$7 across any cut and an open-boundary bond-dimension-$U \;=\; \sum_{i_1,\dots,i_N\atop j_1,\dots,j_N} \Tr\bigl(b\;A_N^{\,i_N j_N}\cdots A_1^{\,i_1 j_1}\bigr)\, \bigl|\,i_1\cdots i_N\bigr\rangle\bigl\langle j_1\cdots j_N\bigr|\,.$8 MPU representation, yet has infinite light-cone radius and can create arbitrarily long-range correlations. More generally, for nonuniform or open-boundary MPUs, unitarity is governed by recursive conditions on the span of the traceless transfer operators and on the boundary data. In the semi-simple subclass, periodic-boundary MPUs are forced to factorize as

$U \;=\; \sum_{i_1,\dots,i_N\atop j_1,\dots,j_N} \Tr\bigl(b\;A_N^{\,i_N j_N}\cdots A_1^{\,i_1 j_1}\bigr)\, \bigl|\,i_1\cdots i_N\bigr\rangle\bigl\langle j_1\cdots j_N\bigr|\,.$9

so the nontrivial beyond-QCA phenomena are genuinely tied to boundary or nonuniform structure (Styliaris et al., 2024).

The same work establishes a bijection between MPUs and locally maximally entanglable states through Choi vectorization. For a nonuniform MPU $N$00, its vectorized state $N$01 is an MPS of bond dimension $N$02, and one can express

$N$03

with each $N$04 an isometry. This shows that general MPUs need not reduce to depth-$N$05 local circuits on a fixed local register, even though they retain controlled operator entanglement (Styliaris et al., 2024).

Recent circuit-complexity results partially close that gap. For a large class of open-boundary MPUs built from a repeated bulk tensor and satisfying a bond-nondegeneracy assumption, one can construct a polynomial-depth quantum circuit using local isometries and a binary tree of merging steps implemented through oblivious amplitude amplification. In the uniform-bulk case the depth satisfies

$N$06

while in the nonuniform case one has

$N$07

with $N$08 the smallest nonzero Schmidt value of the normalized Choi state. QCA correspond to $N$09, giving $N$10. This does not turn every MPU into a shallow circuit, but it shows that a broad bond-nondegenerate class is polynomially realizable despite the nonunitarity of the local tensors (Styliaris et al., 11 Aug 2025).

Random matrix-product unitaries provide a different perspective on complexity. In the random MPU ensemble with bond dimension $N$11, polynomial $N$12 already suffices to reproduce Haar values of higher-order OTOCs for local, finite-trace observables: $N$13 The same ensemble forms an approximate unitary $N$14-design in frame-potential sense, with

$N$15

for $N$16. However, for local traceless observables, reproducing Haar-like OTOCs requires $N$17, i.e. volume-law bond dimension. This separates approximate-design behavior from the stronger notion of freeness probed by OTOCs (Dowling et al., 31 Jul 2025).

Taken together, these developments sharpen the scope of the MPU formalism. Uniform periodic qudit MPUs support a complete locality-preserving theory with canonical forms and an index. Open-boundary, nonuniform, random, and fermionic variants preserve the tensor-network advantages of MPO representations but no longer collapse to the same notion of QCA. The modern theory of MPUs is therefore not a single classification theorem, but a collection of closely related regimes linked by transfer-matrix structure, symmetry data, and operator-entanglement constraints.