Boundary Matrix Product State Contraction

• Boundary matrix product state contraction is a fixed-point method that approximates dominant eigenvectors of row-to-row transfer matrices using an MPS ansatz with tunable bond dimension.
• It employs an iterative 'power method + truncation' scheme where canonicalization and gauge transformations directly affect the entanglement structure and numerical stability.
• Optimizing the virtual-index gauge redistributes physically relevant information into survivable Schmidt sectors, thereby enhancing accuracy in thermodynamic evaluations.

Searching arXiv for recent and foundational papers on boundary matrix product state contraction. Boundary matrix product state contraction is a fixed-point approach for contracting two-dimensional tensor networks by approximating the dominant eigenvectors of a row-to-row transfer matrix with a matrix product state (MPS). For a square-lattice classical model on an $L\times N$ lattice, the partition function can be written as $Z=\mathrm{Tr}(T^N)$, where $T$ is the transfer matrix between adjacent rows. The boundary-MPS strategy replaces exact manipulation of $T$, whose effective Hilbert space grows as $d^L$, with an MPS ansatz of bond dimension $D$ for the dominant right- and left-fixed points. In non-Hermitian settings, virtual-index gauge degrees of freedom are not merely representational: they can modify the entanglement structure of the fixed-point states and alter how physical information is distributed across Schmidt sectors, with direct consequences for numerical stability and accuracy (Tang et al., 2023).

1. Transfer-matrix formulation of two-dimensional contraction

For a square-lattice classical model with local Boltzmann weights $w_{ij}$, one groups the lattice into rows of $L$ spins $\{s_1,\dots,s_L\}$. The partition function is then represented as

$$Z \;=\; \sum_{\{s_{x,y}=\pm1\}} \prod_{x=1}^L \prod_{y=1}^N w\!\bigl(s_{x,y},s_{x+1,y},s_{x,y+1},s_{x+1,y+1}\bigr) \;=\; \mathrm{Tr}\;\bigl(T^N\bigr),$$

with composite row configurations $Z=\mathrm{Tr}(T^N)$0, $Z=\mathrm{Tr}(T^N)$1, and

$Z=\mathrm{Tr}(T^N)$2

Periodic boundary conditions identify $Z=\mathrm{Tr}(T^N)$3.

Boundary-MPS contraction targets the dominant eigenstructure of $Z=\mathrm{Tr}(T^N)$4. The central object is the dominant right-eigenvector $Z=\mathrm{Tr}(T^N)$5 satisfying

$Z=\mathrm{Tr}(T^N)$6

Once this fixed point is approximated accurately, thermodynamic and local quantities can be extracted from the corresponding transfer-matrix contractions. The method is therefore a boundary reduction of a two-dimensional network to a one-dimensional fixed-point problem (Tang et al., 2023).

2. MPS fixed-point ansatz and iterative contraction

The dominant eigenvector is represented as a periodic or infinite MPS of bond dimension $Z=\mathrm{Tr}(T^N)$7: $Z=\mathrm{Tr}(T^N)$8 with local tensors $Z=\mathrm{Tr}(T^N)$9 carrying physical index $T$0 and virtual indices $T$1. In components,

$T$2

The eigenvalue equation becomes an MPO-MPS eigenproblem,

$T$3

The standard numerical scheme is an iterative fixed-point algorithm described as “power method + truncation.” One initializes a random MPS of bond dimension at most $T$4, applies the transfer MPO to obtain a new MPS, brings the result into left- or right-canonical form by local gauge transformations, performs bondwise singular-value decompositions, truncates to the $T$5 largest Schmidt values, and normalizes the state. The iteration is repeated until the eigenvalue estimate converges. In this workflow, the canonicalization step is not an auxiliary convenience: in non-Hermitian transfer problems it interacts directly with the gauge dependence of the entanglement spectrum and the conditioning of truncation (Tang et al., 2023).

3. Virtual-index gauge freedom and its imaginary-time interpretation

A reversible gauge transformation on the virtual indices is implemented by inserting invertible matrices $T$6 sitewise: $T$7 At the transfer-operator level this induces

$T$8

When $T$9 is taken from the singular-value decomposition $T$0, one has, up to local unitaries,

$T$1

Applying $T$2 is then equivalent to an imaginary-time evolution under a diagonal Hamiltonian $T$3. Large $T$4 tends to project the MPS onto the ground or highest-energy product states of $T$5, thereby suppressing entanglement.

This mechanism clarifies why gauge choice can alter contraction quality even when it leaves the exact partition function invariant. The exact tensor network is gauge-equivalent, but a finite-$T$6 boundary-MPS approximation is not gauge-invariant because truncation depends on the Schmidt spectrum of the chosen representation. A plausible implication is that gauge optimization should be understood as a problem of redistributing physically relevant information into Schmidt sectors that survive finite-bond truncation, rather than as a purely formal canonicalization step (Tang et al., 2023).

4. Entanglement structure, reduced densities, and accuracy

At any bond $T$7, the Schmidt decomposition

$T$8

defines Schmidt values $T$9 and entanglement entropy

$d^L$0

Under a nonunitary gauge transformation, the local Schmidt spectrum changes. In practice, one tracks the one-sided reduced density matrices $d^L$1 and $d^L$2 from the right and left eigen-MPS, together with the double-sided object

$d^L$3

which encodes the physical left-right pairing.

The numerical results reported for local gauges $d^L$4 show that $d^L$5 and that the Schmidt values are squeezed as $d^L$6 increases. However, the same transformation can push the physically relevant components into small Schmidt sectors that are then truncated away. This directly contradicts a common simplification according to which reduced entanglement automatically improves fixed-point contraction. In the square-lattice Ising benchmarks, the variance

$d^L$7

decreases as gauge strength grows, yet the physical accuracy deteriorates (Tang et al., 2023).

A second misconception is that one-sided diagnostics suffice. The reported discrepancy between $d^L$8, $d^L$9, and $D$0 shows that non-Hermitian contraction problems can be well-conditioned from one side and ill-conditioned from the physically relevant biorthogonal pairing. This suggests that boundary-MPS contraction of non-Hermitian transfer matrices is intrinsically biorthogonal, even when the algorithm is implemented through one-sided updates.

5. Model examples and the role of normality

Two numerical examples illustrate how gauge structure controls contraction behavior.

Model	Representation or gauge	Reported numerical behavior
Square-lattice Ising at criticality	$D$1 or $D$2	$D$3 falls $D$4; free-energy relative error rises rapidly with $D$5 in single-direction TEBD; VUMPS is more robust but still degrades
Triangular-lattice antiferromagnetic Ising	$D$6 normal versus $D$7 non-normal; $D$8	direct boundary-MPS on $D$9 fails; $w_{ij}$0; for $w_{ij}$1 convergence is restored

In the square-lattice critical Ising case, the two local gauges $w_{ij}$2 and $w_{ij}$3 both reduce entanglement entropy, with the decay faster for $w_{ij}$4. Yet the free-energy relative error

$w_{ij}$5

rises rapidly with $w_{ij}$6 in single-direction TEBD. VUMPS truncation is more robust, but still degrades as $w_{ij}$7 grows. The combination of lower entropy and lower variance with worse thermodynamic estimates establishes that gauge-induced disentangling is not equivalent to improved contraction quality (Tang et al., 2023).

In the triangular-lattice antiferromagnetic Ising problem, two tensor-network representations of the same partition function are related by a non-local gauge transformation. The normal representation $w_{ij}$8 encodes all local symmetries, whereas the non-normal representation $w_{ij}$9 leads to failure of direct boundary-MPS contraction: the power method does not converge and the variance blows up. A nonlocal MPO gauge,

$L$0

interpolates $L$1 to $L$2 through

$L$3

For $L$4, the boundary MPS converges as for $L$5. In the same interpolation, the entanglement spectra of $L$6 evolve from complex, characteristic of the non-normal point, to positive real when $L$7. The central-charge fit from entanglement scaling,

$L$8

returns $L$9 only near the normal point. The concrete numerical lesson is that normality is not a peripheral algebraic property: it is a controlling factor for the conditioning of fixed-point contraction (Tang et al., 2023).

6. Gauge criteria, best practices, and adjacent transfer-matrix frameworks

The operational guidance extracted from these results is explicit. Whenever possible, the transfer MPO should be gauged toward normality by maximizing

$\{s_1,\dots,s_L\}$0

Equivalent guidance in the same spirit is to impose the minimal canonical form, to preserve physical spatial symmetries so that the transfer operator is block-diagonal or normal, and to favor variational truncation schemes such as VUMPS or tangent-space methods over a simple one-sided SVD truncation. Monitoring both one-sided reduced densities $\{s_1,\dots,s_L\}$1, $\{s_1,\dots,s_L\}$2 and the double-sided $\{s_1,\dots,s_L\}$3 is part of the recommended diagnostic set; large discrepancies indicate ill-conditioning. For non-Hermitian MPOs produced by nonlocal gauge mischoices, one should search for a small-bond MPO $\{s_1,\dots,s_L\}$4 that restores normality (Tang et al., 2023).

Related transfer-matrix constructions in other settings clarify the broader significance of boundary-state contraction, while remaining technically distinct from the two-dimensional classical-network problem. In boundary-driven quantum chains, non-Hermitian and often non-diagonalisable families of commuting transfer operators arise from non-unitary irreducible representations of Yang-Baxter algebra, and contraction is reorganized through a double-auxiliary transfer matrix and a projected diagonal subspace (Prosen, 2015). In the algebraic Bethe-ansatz treatment of the open XXZ chain, exact Bethe states with open boundaries are written as standard open-chain MPS with fixed end-vectors, and norms, overlaps, and one-point functions reduce to powers of a local transfer operator $\{s_1,\dots,s_L\}$5 contracted against boundary tensors (Mei et al., 2016). For inhomogeneous MPS in left-canonical CPTP gauge, the auxiliary transfer maps $\{s_1,\dots,s_L\}$6 define right-tail products $\{s_1,\dots,s_L\}$7, and decay of the centered trace-Dobrushin coefficient $\{s_1,\dots,s_L\}$8 yields a unique auxiliary right boundary sequence, quantitative boundary stability, and correlation bounds governed by the same auxiliary product coefficients (Pathirana, 30 Apr 2026).

Taken together, these results place boundary matrix product state contraction within a broader transfer-operator framework in which fixed points, boundary data, and contraction properties are inseparable from the chosen gauge. In the specific context of two-dimensional tensor-network contraction, the decisive point is that virtual-index gauge freedom changes the numerically accessible entanglement structure of the transfer-matrix eigenstates. The practical problem is therefore not only to approximate the dominant fixed point, but also to represent it in a gauge where truncation preserves the physically relevant sectors.