Tensor Networks with Belief Propagation

• Tensor Networks with Belief Propagation (TNBP) is a family of methods that use BP-style message passing to replace exact tensor contraction with efficient, local approximations.
• Variants like blockBP, generalized BP, and influence-functional BP extend the approach by incorporating block messaging, overlapping regions, and cluster corrections to handle loops and frustration.
• TNBP delivers systematic improvements in simulation accuracy and scalability, though its performance may degrade near criticality or in highly entangled systems.

Searching arXiv for papers on tensor networks with belief propagation and closely related variants to ground the encyclopedia article. arXiv search query: "tensor networks belief propagation blockBP generalized belief propagation cluster expansion influence functional" Tensor Networks with Belief Propagation (TNBP) denotes a family of tensor-network contraction and environment-approximation methods in which belief propagation (BP) or one of its generalizations is used to summarize the effect of the rest of a network on a local region. In the cited literature, the term covers single-site BP for PEPS norm contraction, block and neighborhood coarse-grainings, generalized belief propagation (GBP) on overlapping regions, influence-functional BP for nonequilibrium dynamics, and cluster- or loop-corrected schemes that systematically improve the BP baseline. Across these formulations, the common structure is the same: exact tensor contraction is replaced by local tensor contractions coupled through self-consistent messages, with tree graphs treated exactly and loopy graphs treated approximately or corrected perturbatively (Alkabetz et al., 2020, Guo et al., 2023, Midha et al., 2 Oct 2025, Park et al., 10 Apr 2025).

1. Terminology and scope

The term emerged from the observation that PEPS contraction can be rewritten as inference on an associated graphical model. In that formulation, the double-layer network for $\langle \psi|\psi\rangle$ is mapped to a double-edge factor graph (DEFG), BP messages become matrix-valued objects on doubled virtual bonds, and the resulting approximation is equivalent to the Bethe-Peierls approximation. A central early result is that the simple-update environment approximation is essentially the same approximation as BP on the doubled network, so simple update can be interpreted as a tensor-network realization of Bethe-Peierls mean field (Alkabetz et al., 2020).

In later work, the designation broadened. Some papers use TNBP for any tensor-network contraction method driven by BP-style message passing, whereas one paper uses TNBP specifically for “Tensor Network Influence Functional Belief Propagation,” abbreviated IF-BP, in real-time dynamics on arbitrary graphs (Park et al., 10 Apr 2025). This suggests that TNBP is best understood as a methodological family rather than a single algorithm.

Variant	Message object	Representative paper
Single-site PEPS BP	matrix-valued messages on doubled edges	(Alkabetz et al., 2020)
blockBP	MPS boundary messages between blocks	(Guo et al., 2023)
TN-IF / IF-BP	low-rank MPSs in time	(Park et al., 10 Apr 2025)
GBP for TN contraction	messages between overlapping regions	(Tindall et al., 27 Apr 2026)
Cluster-corrected BP	loop or cluster corrections around a BP vacuum	(Midha et al., 2 Oct 2025)

2. Belief propagation as a tensor-network contraction principle

A general formulation treats a tensor network on a graph $G=(V,E)$ with local tensors $T_v$ and exact contraction

$Z(T)=\star_{v\in V} T_v.$

BP introduces directed edge messages $\mu_{v\to w}\in B_{vw}$ satisfying the fixed-point relation

$$\left(\bigotimes_{n\in N(v)\setminus \{w\}} \mu_{n\to v}\right)\star T_v=\mu_{v\to w},$$

with normalization $\mu_{v\to w}\star \mu_{v\to w}=1$. From these messages one forms local BP factors

$$Z^{(v)}:=\left[\bigotimes_{n\in N(v)}\mu_{n\to v}\right]\star T_v,$$

and the BP approximation

$$Z_0=\prod_{v\in V} Z^{(v)}, \qquad F_0=-\sum_{v\in V}\log Z^{(v)}.$$

In this language, BP is the zeroth-order “vacuum” approximation: exact on loopless graphs, approximate on loopy graphs (Midha et al., 2 Oct 2025).

For PEPS, the construction is usually written on the doubled norm network. Each virtual bond is replaced by a doubled variable $z_i=(x_i,x_i')$, and each local factor is obtained by contracting a tensor and its conjugate over the physical leg,

$G=(V,E)$0

The BP recursion becomes a tensor-network message update over positive semidefinite matrices,

$G=(V,E)$1

and fixed points correspond to extrema of a Bethe free-energy functional on the doubled variables. In this setting, the approximation used in simple update is identical to the Bethe-Peierls approximation implemented by BP (Alkabetz et al., 2020).

The same message tensors can also be used for canonicalization. BP on the norm network yields approximate local environments, and those environments can be converted into the Vidal gauge by square-root factorizations and SVDs on each edge. On trees this reproduces the standard canonical/Vidal form exactly; on loopy graphs it yields an approximate quasi-canonical gauge that improves truncation and simple-update evolution (Tindall et al., 2023).

3. Block, neighborhood, and region generalizations

Plain BP summarizes each bond environment by a single message, which is often too crude in two-dimensional or frustrated systems. A direct extension is block belief propagation, or blockBP, in which the lattice is partitioned into non-overlapping blocks and BP is run on the block graph rather than on individual tensors. The outgoing message from block $G=(V,E)$2 to neighboring block $G=(V,E)$3 is

$G=(V,E)$4

Here the messages are MPSs representing block-boundary environments. Loops inside a block are treated exactly up to the accuracy of the internal contraction, while correlations between distinct block edges are approximated by the product of edgewise MPS messages. For block size $G=(V,E)$5, the method reduces to the earlier BP contraction algorithm equivalent to simple update (Guo et al., 2023).

The infinite-lattice version adopts a repeating unit cell or a self-messaging block and is explicitly multithreadable. In one implementation for infinite lattices, the per-iteration scaling for a block with $G=(V,E)$6 tensors, $G=(V,E)$7 messages, and bond truncation $G=(V,E)$8 is

$G=(V,E)$9

reduced to

$T_v$0

with multi-threading over the messages. For the Kagome lattice, both a symmetry-preserving hexagonal block and a simpler parallelogram block were considered, with the hexagonal block chosen because it preserves the lattice’s sixfold rotational symmetry (Gutman, 2 Mar 2026).

A second line of generalization enlarges the local region around a tensor without necessarily imposing a block partition. Tensor Network Message Passing (TNMP) contracts a neighborhood $T_v$1 exactly or approximately and replaces the rest of the graph by factorized boundary cavity tensors, yielding an interpolation between full local tensor contraction and BP. Its exactness regime is “globally tree-like and locally dense-connected when the dense local graphs have limited treewidth” (Wang et al., 2023). Closely related ideas appear in generalized BP for arbitrary graphical models and loopy tensor networks, where one works with tensor neighborhoods $T_v$2 or overlapping regions and, when necessary, introduces intersection messages to cure the missing-leg problem on arbitrary loopy tensor networks (Hack et al., 2024).

GBP formalizes the region idea through a region graph and the Kikuchi free energy. Parent regions, their overlaps, and Möbius counting numbers $T_v$3 define beliefs $T_v$4 and message tensors between overlapping regions. Ordinary BP is the corner case in which parent regions are single tensors and child regions are pairwise shared indices. Larger regions, such as plaquettes or voxels, explicitly encode short loops and frustration, at increased tensor size and more delicate fixed-point structure (Tindall et al., 27 Apr 2026).

4. Loop expansions, cluster corrections, and stochastic exactification

The most systematic refinement of TNBP starts from the observation that BP gives a distinguished decomposition of each bond into a BP vacuum sector and an orthogonal excitation sector,

$T_v$5

Expanding this identity on all edges yields an exact loop series

$T_v$6

where only generalized loops survive because dangling excitations vanish. The obstruction is combinatorial: disconnected loop collections proliferate too rapidly for naive truncation in $T_v$7 to be systematically reliable (Midha et al., 2 Oct 2025).

The remedy is to reorganize the expansion in $T_v$8. After BP normalization, loops are grouped into connected clusters $T_v$9, and only connected clusters contribute to the free-energy expansion,

$Z(T)=\star_{v\in V} T_v.$0

The resulting cluster expansion eliminates the disconnected overcounting inherent in the loop series. A sufficient convergence condition is

$Z(T)=\star_{v\in V} T_v.$1

when loop contributions obey $Z(T)=\star_{v\in V} T_v.$2. Under that condition,

$Z(T)=\star_{v\in V} T_v.$3

The algorithm computes connected clusters up to weight $Z(T)=\star_{v\in V} T_v.$4, with complexity polynomial in $Z(T)=\star_{v\in V} T_v.$5 for fixed $Z(T)=\star_{v\in V} T_v.$6, scaling like $Z(T)=\star_{v\in V} T_v.$7, and each cluster contribution can be evaluated independently (Midha et al., 2 Oct 2025).

The cluster framework extends from partition functions to local observables. For PEPS satisfying a loop-decay condition, BP supplemented by cluster corrections approximates local observables with exponentially small relative error, and connected correlations can be written as sums over connected clusters intersecting both observable regions. In this representation, loop decay necessarily implies exponential decay of connected correlations, which yields a rigorous criterion for when BP can succeed and, conversely, rules out its validity at critical points (Midha et al., 3 Apr 2026).

A different route samples loop corrections rather than truncating them deterministically. In Belief Propagation Loop Monte Carlo (BPLMC), the exact partition function is factorized as

$Z(T)=\star_{v\in V} T_v.$8

for pairwise Markov random fields with symmetric edge potentials. The sampled objects are even-degree subgraphs, updated by XOR with cycles in a cycle basis; umbrella sampling is used to keep the empty-loop sector visible. Within its stated class of models, the estimator is unbiased and has controllable statistical error (Sim et al., 9 Mar 2026).

5. Applications across tensor-network simulation

In finite and infinite PEPS ground-state calculations, blockBP has been used as the environment engine inside imaginary-time evolution. On finite square lattices, blockBP energies for the transverse Ising and antiferromagnetic Heisenberg models match boundary-MPS full-update benchmarks to the printed precision in the reported tables. In the $Z(T)=\star_{v\in V} T_v.$9 transverse Ising model, the average trace distance between blockBP and bMPS two-site reduced density matrices is below $\mu_{v\to w}\in B_{vw}$0 away from criticality and about $\mu_{v\to w}\in B_{vw}$1 near $\mu_{v\to w}\in B_{vw}$2. For a finite $\mu_{v\to w}\in B_{vw}$3 AFH benchmark, the runtime shows near-linear scaling with the number of processes $\mu_{v\to w}\in B_{vw}$4 up to $\mu_{v\to w}\in B_{vw}$5 (Guo et al., 2023).

For the antiferromagnetic Heisenberg model on the infinite Kagome lattice, BlockBP was used to approximate the infinite environment in PEPS imaginary-time evolution. At $\mu_{v\to w}\in B_{vw}$6, the reported ground-state energies are SU: $\mu_{v\to w}\in B_{vw}$7, VU: $\mu_{v\to w}\in B_{vw}$8, and BlockBP: $\mu_{v\to w}\in B_{vw}$9. The same study reports that the magnetization tends toward zero during evolution and that bond energies and negativities exhibit frustration patterns consistent with a spin-liquid-like state (Gutman, 2 Mar 2026).

In nonequilibrium dynamics, IF-BP extends the tensor network influence functional to tree lattices, heavy-hex lattices, and arbitrary sparse graphs. On trees the method is exact up to MPS truncation; on loopy graphs it adds a BP approximation on top of temporal-MPS compression. For the heavy-hex kicked Ising model, at $$\left(\bigotimes_{n\in N(v)\setminus \{w\}} \mu_{n\to v}\right)\star T_v=\mu_{v\to w},$$0 Trotter steps the maximum error is reported below $$\left(\bigotimes_{n\in N(v)\setminus \{w\}} \mu_{n\to v}\right)\star T_v=\mu_{v\to w},$$1 at $$\left(\bigotimes_{n\in N(v)\setminus \{w\}} \mu_{n\to v}\right)\star T_v=\mu_{v\to w},$$2 and below $$\left(\bigotimes_{n\in N(v)\setminus \{w\}} \mu_{n\to v}\right)\star T_v=\mu_{v\to w},$$3 at $$\left(\bigotimes_{n\in N(v)\setminus \{w\}} \mu_{n\to v}\right)\star T_v=\mu_{v\to w},$$4. At $$\left(\bigotimes_{n\in N(v)\setminus \{w\}} \mu_{n\to v}\right)\star T_v=\mu_{v\to w},$$5, IF-BP with $$\left(\bigotimes_{n\in N(v)\setminus \{w\}} \mu_{n\to v}\right)\star T_v=\mu_{v\to w},$$6 agrees well with MIX TN and iPEPS benchmarks, and for times up to $$\left(\bigotimes_{n\in N(v)\setminus \{w\}} \mu_{n\to v}\right)\star T_v=\mu_{v\to w},$$7 it continues to produce physically sensible magnetization dynamics while iPEPS shows unphysical late-time behavior. The same work reports that the temporal entanglement entropy grows logarithmically with time in the heavy-hex kicked Ising model, and that a cluster expansion in loop width $$\left(\bigotimes_{n\in N(v)\setminus \{w\}} \mu_{n\to v}\right)\star T_v=\mu_{v\to w},$$8 systematically improves quenches in the 2D transverse-field Ising model; at $$\left(\bigotimes_{n\in N(v)\setminus \{w\}} \mu_{n\to v}\right)\star T_v=\mu_{v\to w},$$9, $\mu_{v\to w}\star \mu_{v\to w}=1$0 and $\mu_{v\to w}\star \mu_{v\to w}=1$1 agree well with sparse Pauli dynamics up to about $\mu_{v\to w}\star \mu_{v\to w}=1$2, with errors below $\mu_{v\to w}\star \mu_{v\to w}=1$3 (Park et al., 10 Apr 2025).

BP-based gauging and truncation have also been used outside PEPS contraction proper. In the graph tensor-network quantum annealer (GTQA), BP is repeatedly invoked to restore an approximate Vidal gauge after gate applications on sparse graph tensor networks. Reported simulations reach up to $\mu_{v\to w}\star \mu_{v\to w}=1$4 qubits and $\mu_{v\to w}\star \mu_{v\to w}=1$5 two-qubit gates on random 3-regular graphs. On a $\mu_{v\to w}\star \mu_{v\to w}=1$6-qubit non-degenerate QUBO instance, the recovered solution matches the best of $\mu_{v\to w}\star \mu_{v\to w}=1$7 heuristic solvers on $\mu_{v\to w}\star \mu_{v\to w}=1$8 out of $\mu_{v\to w}\star \mu_{v\to w}=1$9 bits, with an objective value only $$Z^{(v)}:=\left[\bigotimes_{n\in N(v)}\mu_{n\to v}\right]\star T_v,$$0 worse than the best heuristic; on a $$Z^{(v)}:=\left[\bigotimes_{n\in N(v)}\mu_{n\to v}\right]\star T_v,$$1-vertex MaxCut instance, a reported cut of $$Z^{(v)}:=\left[\bigotimes_{n\in N(v)}\mu_{n\to v}\right]\star T_v,$$2 versus a best heuristic value of $$Z^{(v)}:=\left[\bigotimes_{n\in N(v)}\mu_{n\to v}\right]\star T_v,$$3 gives $$Z^{(v)}:=\left[\bigotimes_{n\in N(v)}\mu_{n\to v}\right]\star T_v,$$4 (Luchnikov et al., 2024).

6. Regimes of validity, convergence issues, and fundamental limits

The principal attraction of TNBP is algorithmic. BP is exact on trees, efficient on arbitrary graphs, highly parallelizable, and naturally compatible with tensor contractions. Block, neighborhood, and region variants expose a tunable hierarchy: plain BP is the cheapest level; blockBP and GBP retain more short-loop structure; cluster corrections yield a systematic expansion; and stochastic loop sampling can recover exactness within a restricted model class (Hack et al., 2024, Tindall et al., 27 Apr 2026, Sim et al., 9 Mar 2026).

The main failure modes are equally consistent across the literature. Plain loopy BP is uncontrolled: convergence and uniqueness are not guaranteed, and accuracy deteriorates when long-range correlations, strong frustration, or short loops dominate. In PEPS language, the approximation is local and mean-field-like, so it can miss nonlocal entanglement. The rigorous cluster theory sharpens this picture by showing that the loop-decay hypothesis required for exponential convergence implies exponential clustering of correlations, so critical points lie outside the formal validity regime (Alkabetz et al., 2020, Midha et al., 3 Apr 2026).

GBP improves loop handling but introduces its own caveats. For norm networks, simple BP tends to preserve positive semidefiniteness of messages, whereas more general GBP updates do not generally preserve PSD-ness. The reported consequences include non-real or unbounded Kikuchi free energies and convergence failures when local tensors contain negative or complex entries. In positive classical or positive norm-network settings, plaquette- or voxel-based regions can substantially improve energy and entropy estimates, but the method is markedly more fragile in sign-rich random tensor networks (Tindall et al., 27 Apr 2026).

There are also explicit complexity barriers. For random-circuit out-of-time-order correlators on Willow-like square-lattice hardware, one study concluded that TNBP is not a feasible classical simulator. The combination of dense 2D connectivity, loopy BP environments, and Schrödinger-picture entanglement growth leads to rapidly increasing PEPS bond dimensions $$Z^{(v)}:=\left[\bigotimes_{n\in N(v)}\mu_{n\to v}\right]\star T_v,$$5 and boundary-MPS bond dimensions $$Z^{(v)}:=\left[\bigotimes_{n\in N(v)}\mu_{n\to v}\right]\star T_v,$$6. On two 23-qubit experimental circuits with $$Z^{(v)}:=\left[\bigotimes_{n\in N(v)}\mu_{n\to v}\right]\star T_v,$$7 and $$Z^{(v)}:=\left[\bigotimes_{n\in N(v)}\mu_{n\to v}\right]\star T_v,$$8 two-qubit gates, using $$Z^{(v)}:=\left[\bigotimes_{n\in N(v)}\mu_{n\to v}\right]\star T_v,$$9 and $$Z_0=\prod_{v\in V} Z^{(v)}, \qquad F_0=-\sum_{v\in V}\log Z^{(v)}.$$0, the reported TNBP signal-to-noise ratios are approximately $$Z_0=\prod_{v\in V} Z^{(v)}, \qquad F_0=-\sum_{v\in V}\log Z^{(v)}.$$1 and $$Z_0=\prod_{v\in V} Z^{(v)}, \qquad F_0=-\sum_{v\in V}\log Z^{(v)}.$$2, respectively, versus larger values for TNMC and for the experiment. The same paper estimates $$Z_0=\prod_{v\in V} Z^{(v)}, \qquad F_0=-\sum_{v\in V}\log Z^{(v)}.$$3 for a $$Z_0=\prod_{v\in V} Z^{(v)}, \qquad F_0=-\sum_{v\in V}\log Z^{(v)}.$$4-qubit OTOC case and $$Z_0=\prod_{v\in V} Z^{(v)}, \qquad F_0=-\sum_{v\in V}\log Z^{(v)}.$$5 for a $$Z_0=\prod_{v\in V} Z^{(v)}, \qquad F_0=-\sum_{v\in V}\log Z^{(v)}.$$6-qubit OTOC$$Z_0=\prod_{v\in V} Z^{(v)}, \qquad F_0=-\sum_{v\in V}\log Z^{(v)}.$$7 case, with associated memory demands deemed infeasible (Bermejo et al., 16 Apr 2026).

Taken together, the literature defines TNBP not as a single algorithm but as a layered contraction paradigm. BP supplies the tree-exact local environment; tensor-network structure specifies what is contracted at each update; blocks, neighborhoods, and overlapping regions enlarge the exact local computation; cluster or stochastic loop corrections repair the loopy-graph error; and rigorous results identify exponential loop decay as the key small parameter behind success. Within that regime, TNBP provides an efficient and systematically improvable contraction framework; outside it, especially near criticality or in highly entangled dense 2D circuit dynamics, its limitations are structural rather than merely implementation-dependent (Midha et al., 2 Oct 2025, Midha et al., 3 Apr 2026).