Fundamental Theorem of MPS

• The Fundamental Theorem of Matrix-Product States defines conditions under which distinct MPS tensors yield the same quantum state family through canonical forms and invertible gauge transformations.
• Canonicalization and transfer-operator analysis remove redundancies, ensuring that only the physically meaningful tensor blocks determine state equivalence.
• This structural result underpins the classification of one-dimensional quantum phases, the design of parent Hamiltonians, and extensions to periodic and infinite-dimensional settings.

Searching arXiv for recent and foundational papers on the Fundamental Theorem of Matrix-Product States. The Fundamental Theorem of Matrix-Product States (MPS) is the structural statement that characterizes when two local tensor descriptions generate the same one-dimensional many-body quantum state. In its standard form, the theorem asserts that, once the tensors are brought into an appropriate canonical form, equality of the associated matrix product vectors for all system sizes is equivalent to an invertible gauge transformation acting on the virtual indices; in proportional versions, additional blockwise phases may appear (Cirac et al., 2020). This result places gauge freedom, canonical form, transfer-operator structure, and injectivity at the center of MPS theory. It underlies classification results for one-dimensional phases, symmetry constraints on virtual degrees of freedom, parent-Hamiltonian constructions, and several generalizations, including periodic tensors, ergodic spatially varying tensor sequences, operational rank characterizations, and infinite-dimensional auxiliary spaces (Cuevas et al., 2017, Movassagh et al., 2019, Heikkinen, 18 Feb 2025).

1. Standard statement and formal setting

An MPS, in the translationally invariant periodic setting, is written as

$$|V^{(N)}(A)\rangle=\sum_{i_1,\dots,i_N=1}^d \mathrm{tr}\!\left(A^{i_1}A^{i_2}\cdots A^{i_N}\right)|i_1\cdots i_N\rangle,$$

with local matrices $A^i\in\mathbb C^{D\times D}$ (Cirac et al., 2020). The theorem concerns the inverse problem: given two tensor families $A=\{A^i\}$ and $B=\{B^i\}$, when do they define the same physical state family?

For tensors in canonical form, the fundamental theorem for proportional matrix product vectors states that if $A$ and $B$ generate proportional MPVs for all system sizes $N$, then they have the same number of normal blocks, and after reordering the blocks,

$B^i_k=e^{i\phi_k}X_k A^i_{j_k}X_k^{-1},$

with $X_k$ invertible (Cirac et al., 2020). As an immediate corollary, if the two canonical tensors generate exactly the same MPV for all $N$, then

$A^i\in\mathbb C^{D\times D}$0

so the only remaining ambiguity is a virtual gauge transform (Cirac et al., 2020).

This theorem is often summarized as “same state family if and only if same tensor up to gauge,” but that summary is only correct after the tensors have been reduced to an appropriate canonical representative. Without canonicalization, off-diagonal and upper-triangular redundancies can survive in the raw tensor data while being invisible in the physical state (Cirac et al., 2020).

The theorem is naturally formulated for entire families $A^i\in\mathbb C^{D\times D}$1, not for a single fixed system size. This distinction is essential, because accidental equality at one length need not reflect genuine equivalence of local tensor data. A plausible implication is that the theorem should be regarded as a rigidity statement about the thermodynamic tensor law rather than merely about finite-size wavefunctions.

2. Canonical form, normal tensors, and transfer operators

Canonical form is the mechanism that removes spurious redundancies and isolates the physically meaningful tensor blocks. In the formulation reviewed in (Cirac et al., 2020), a tensor is called normal if its transfer operator is primitive; equivalently, after blocking sufficiently many sites, it has a unique largest eigenvalue and strictly positive left and right fixed points. A general tensor decomposes as

$A^i\in\mathbb C^{D\times D}$2

where each $A^i\in\mathbb C^{D\times D}$3 is normal (Cirac et al., 2020).

The corresponding transfer matrix is

$A^i\in\mathbb C^{D\times D}$4

and its spectral structure governs correlations, canonical normalization, and injectivity (Cirac et al., 2020). For a normal MPS, the leading eigenvalue is unique and the fixed points $A^i\in\mathbb C^{D\times D}$5 are positive definite; after blocking finitely many sites, the tensor becomes injective (Cirac et al., 2020).

The importance of canonical form is not only technical. It makes the theorem meaningful by separating physical equivalence from presentation artifacts. In this gauge-fixed setting, the theorem says that virtual gauge freedom is the only freedom left. When the transfer map is unital, the similarity transform can be chosen unitary (Cirac et al., 2020).

The same transfer-operator logic appears in closely related settings. In stochastic-process MPS, the infinite-chain transfer matrix is

$A^i\in\mathbb C^{D\times D}$6

and ergodicity is equivalent to nondegeneracy of the dominant transfer-matrix eigenvalue (Yang et al., 2018). In ergodic, spatially varying MPS, a sequence of channel products converges exponentially to a rank-one replacement process, yielding boundary-independent bulk limits and exponential clustering (Movassagh et al., 2019). These results are not the standard fundamental theorem, but they reinforce the central role of transfer fixed points and canonicalization.

3. Injective, non-injective, periodic, and boundary-sensitive variants

The cleanest form of the theorem occurs for injective MPS. In that case, the map from virtual to physical degrees of freedom is one-to-one, the tensor becomes unique up to a single gauge transform, and the parent Hamiltonian has a unique gapped ground state (Cirac et al., 2020). This is the regime in which the theorem functions as a full classification statement for uniform one-dimensional tensor descriptions.

Non-injective tensors require a block decomposition. They correspond to symmetry-breaking or degenerate situations, and their ground space is spanned by the states associated with the normal blocks (Cirac et al., 2020). The theorem then becomes blockwise: equal state families imply matched normal blocks, phases, and similarities rather than a single global gauge.

A central limitation of older canonical-form formulations is that they are naturally adapted to primitive, aperiodic tensors and therefore obscure non-trivial periodicity. The paper “Irreducible forms of Matrix Product States: Theory and Applications” extends the standard framework to arbitrary translationally invariant MPS, including periodic states such as antiferromagnetic patterns (Cuevas et al., 2017). It introduces the irreducible form

$A^i\in\mathbb C^{D\times D}$7

where each block defines an irreducible completely positive map of spectral radius $A^i\in\mathbb C^{D\times D}$8 (Cuevas et al., 2017).

For a periodic block with period $A^i\in\mathbb C^{D\times D}$9, the peripheral spectrum is

$A=\{A^i\}$0

and the tensor admits a cyclic projector decomposition satisfying

$A=\{A^i\}$1

together with

$A=\{A^i\}$2

(Cuevas et al., 2017). The resulting generalized theorem states that if two tensors in irreducible form generate the same MPS, then they are related not only by similarity transforms but also by a diagonal phase matrix $A=\{A^i\}$3 built from roots of unity: $A=\{A^i\}$4 (Cuevas et al., 2017). In the aperiodic case $A=\{A^i\}$5, $A=\{A^i\}$6 is trivial and the standard theorem is recovered.

Boundary conditions affect how the theorem is expressed but not its conceptual content. With periodic boundaries one works with trace-form MPVs; with open boundaries, boundary vectors enter explicitly and the canonical/gauge structure is encoded through Schmidt decompositions and boundary fixed points (Cirac et al., 2020). A plausible implication is that the theorem is best understood as a statement about the virtual bulk law, with boundary data acting as external linear functionals rather than altering the equivalence principle itself.

4. Symmetries, phases, parent Hamiltonians, and renormalization

One of the most important uses of the fundamental theorem is the translation of physical symmetries into virtual relations. For an injective canonical MPS invariant under $A=\{A^i\}$7, invariance implies

$A=\{A^i\}$8

and the theorem makes this relation necessary, not merely sufficient (Cirac et al., 2020). The virtual operators obey

$A=\{A^i\}$9

which produces projective representations classifying one-dimensional symmetry-protected topological phases by $B=\{B^i\}$0, while physical phase factors are classified by $B=\{B^i\}$1 in the presence of time-reversal or reflection (Cirac et al., 2020).

The theorem also explains degeneracies and edge modes. Non-trivial projective virtual symmetry enforces entanglement-spectrum degeneracies and protected edge degrees of freedom, as in AKLT-type states (Cirac et al., 2020). In this sense, the virtual gauge structure is not an arbitrary redundancy but the locus where phase information is stored.

For parent Hamiltonians, the theorem clarifies why the ground-space structure is encoded directly in the canonical block decomposition. Injective MPS have frustration-free parent Hamiltonians with a unique ground state, whereas block-injective MPS yield a ground space equal to the span of the block states (Cirac et al., 2020). The theorem supplies the uniqueness input needed to infer that no other tensor data can generate the same local ground-space family except those related by the admissible virtual transformations.

Renormalization-group fixed points are likewise expressed tensorially. Blocking followed by an isometry leaves the tensor invariant up to gauge at a fixed point, and for MPS zero correlation length is equivalent to

$B=\{B^i\}$2

(Cirac et al., 2020). This identifies the theorem as one of the algebraic engines behind real-space RG for one-dimensional phases.

The irreducible-form extension sharpens the symmetry statement in the periodic case. If

$B=\{B^i\}$3

then there exists a diagonal unitary $B=\{B^i\}$4 commuting with the tensor and a unitary $B=\{B^i\}$5 such that

$B=\{B^i\}$6

(Cuevas et al., 2017). The extra invisible phase matrix is precisely the periodic correction missed by the primitive canonical form.

5. Operational, ergodic, and stochastic reinterpretations

Several later works do not restate the algebraic theorem verbatim but provide structurally analogous results that illuminate its scope.

In “Testing matrix product states,” the relevant MPS characterization is operational rather than gauge-theoretic: an $B=\{B^i\}$7-partite pure state is an MPS of bond dimension $B=\{B^i\}$8 iff every contiguous prefix reduced density matrix has rank at most $B=\{B^i\}$9 (Soleimanifar et al., 2022). The paper gives the tensor definition

$A$0

and then works mainly with the Schmidt-rank formulation (Soleimanifar et al., 2022). This is not the classical fundamental theorem, since canonical forms and gauge freedom are absent, but it captures the same rigidity from intrinsic state data: MPS structure is completely encoded in bipartite bond ranks.

The same paper proves that MPS membership can be tested with

$A$1

copies and that, for $A$2,

$A$3

copies are necessary (Soleimanifar et al., 2022). This suggests that the fundamental structure detected by the theorem is globally constrained even when expressible through local cut ranks.

In “Matrix Product States for Quantum Stochastic Modelling,” the theorem-like content is a process-to-MPS correspondence: a stochastic process represented by an $A$4-machine with transition matrix $A$5 is fully represented by an iMPS with site matrices

$A$6

and the process is ergodic iff the iMPS transfer matrix is nondegenerate (Yang et al., 2018). The right fixed point $A$7 encodes overlaps of the internal quantum states $A$8, and the left fixed point $A$9 encodes stationary causal-state probabilities $B$0 (Yang et al., 2018). This does not classify all tensors yielding the same state family, but it identifies a structural equivalence between predictive causal-state models and MPS representations.

The same work proves that the Schmidt coefficients of the q-sample across any bipartition are the square roots of the eigenvalues of

$B$1

so that the entanglement entropies of the q-sample equal the simulator’s memory entropies,

$B$2

and

$B$3

(Yang et al., 2018). A plausible implication is that, in this setting, the “fundamental theorem” becomes an equivalence between virtual bond structure and predictive memory structure.

In “An ergodic theorem for quantum processes with applications to matrix product states,” the analog of canonicalization is a gauge normalization for ergodic spatially varying tensors,

$B$4

which yields a trace-preserving transformed channel

$B$5

with $B$6 (Movassagh et al., 2019). This is explicitly compared to putting an MPS into canonical form, but for an ergodic, spatially varying tensor sequence rather than a single uniform tensor. The long product of channels approaches a rank-one map

$B$7

with exponential accuracy (Movassagh et al., 2019). The article does not provide an equivalence-class theorem, but it generalizes the roles of canonical fixed points, boundary independence, and exponential clustering.

6. Extensions beyond the standard finite-dimensional theorem

The most direct extension of the theorem’s existence side to infinite-dimensional Hilbert spaces appears in “The Canonical Forms of Matrix Product States in Infinite-Dimensional Hilbert Spaces” (Heikkinen, 18 Feb 2025). There the central theorem states:

$B$8

The proof replaces finite-dimensional linear algebra with the singular value decomposition of compact operators and the Schmidt decomposition for separable Hilbert spaces (Heikkinen, 18 Feb 2025). Under the Hilbert-space isomorphism

$B$9

a state becomes a Hilbert-Schmidt operator with an SVD

$N$0

yielding the infinite-dimensional Schmidt form and then an exact idMPS by recursive decomposition (Heikkinen, 18 Feb 2025).

The resulting coefficient arrays take the form

$N$1

with countably infinite auxiliary indices (Heikkinen, 18 Feb 2025). Left-, right-, mixed-, and fully canonical forms are defined by the corresponding infinite-dimensional normalization conditions. This is not the standard gauge-equivalence theorem, but it extends a foundational component often associated with the theoremic core of MPS: exact representability and canonicalization by iterative Schmidt decompositions.

A different infinite-dimensional direction appears in “Exact bosonic Matrix Product States (and holography),” which derives an exact MPS for bosonic lattice ground states using a one-dimensional Euclidean path integral with a function-space bond variable (Janik, 2018). For a bosonic lattice theory with Euclidean action

$N$2

the local tensor becomes a functional

$N$3

defined by a path integral (Janik, 2018). In the harmonic-chain case, the transfer matrix is exactly diagonalized, and the auxiliary spectrum forms two Fock spaces of continuous modes (Janik, 2018). This is constructive rather than classificatory: no uniqueness-up-to-gauge theorem is proved, but the work shows that exact MPS structure survives well beyond finite bond dimension.

A related structural generalization appears in “Wick’s theorem for matrix product states,” which proves that for a generic translation-invariant MPS or cMPS in the thermodynamic limit, all $N$4-point correlation functions are determined by sufficiently low-order ones, specifically by two- and three-point functions in the generic case $N$5 (Hübener et al., 2012). The transfer data are encoded in the poles of the Z- or Laplace-transformed correlators, e.g.

$N$6

for discrete MPS (Hübener et al., 2012). This is again not the fundamental theorem in the strict gauge-equivalence sense, but it is a theorem of state rigidity derived from the same finite transfer-spectrum structure.

The geometric complement to the fundamental theorem is provided by the MPS analogue of Thouless’ theorem. For a left-canonical reference MPS $N$7, the transformed local tensors

$N$8

parametrize the entire fixed-bond-dimension MPS manifold, provided the reference has no vanishing Schmidt values (Wouters et al., 2013). This result does not identify when two tensor sets define the same state; instead, it gives explicit non-redundant coordinates on the manifold modulo gauge, thereby complementing the fundamental theorem’s quotient-space viewpoint.

7. Scope, misconceptions, and relation to higher-dimensional tensor networks

A common misconception is that the fundamental theorem is simply the statement $N$9. In fact, that formula is valid only under precise hypotheses: canonical or irreducible form, equality of the entire family of states for all lengths, and appropriate injective or block-normal assumptions (Cirac et al., 2020, Cuevas et al., 2017). Outside that regime, extra phases, block decompositions, or periodic root-of-unity structures may appear.

A second misconception is that canonical form is only a numerical convenience. The cited literature treats it as the prerequisite for meaningful uniqueness. Without canonical form, upper-triangular or off-diagonal sectors can be physically silent while altering the raw tensor data (Cirac et al., 2020).

A third misconception is that the theorem fully generalizes to projected entangled pair states (PEPS). For normal PEPS, a corresponding theorem exists: if two normal PEPS tensors generate the same state on sufficiently large lattices, then they are related by a virtual gauge transformation on boundary legs,

$B^i_k=e^{i\phi_k}X_k A^i_{j_k}X_k^{-1},$0

with $B^i_k=e^{i\phi_k}X_k A^i_{j_k}X_k^{-1},$1 on an $B^i_k=e^{i\phi_k}X_k A^i_{j_k}X_k^{-1},$2 torus (Cirac et al., 2020). However, the review emphasizes that the full higher-dimensional theory is far less complete. For semi-injective or MPO-injective PEPS, equivalence is implemented by matrix product operators and pulling-through relations rather than by a simple gauge transform, and a fully general PEPS fundamental theorem is not expected because the corresponding problem is undecidable in full generality (Cirac et al., 2020).

The one-dimensional theorem is therefore unusually strong. It combines canonicalization, transfer-operator Perron–Frobenius structure, injectivity after blocking, and virtual gauge equivalence into a complete algebraic framework. Later developments extend one or another of these ingredients—periodicity, ergodic spatial variation, stochastic predictive structure, operational cut-rank characterizations, exact infinite-dimensional representations—but the classical theorem remains the central uniqueness statement: after removal of redundant blocks, an MPS is determined by its tensor up to virtual gauge, and many of the principal applications of MPS theory flow from that fact (Cirac et al., 2020, Cuevas et al., 2017).