{"id":"952dab48-a19c-4245-9c5a-8ff9c59d17be","arxiv_id":"2411.19100","paper_version":3,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"An infinite matrix-product ansatz for edge messages solves dynamic belief propagation directly in the infinite-time limit, giving accurate steady-state observables and temporal correlations on sparse graphs.","lead":"This paper introduces an analytic approximation for the long-time steady-state dynamics of Markov processes on locally tree-like graphs, encoding edge trajectory probabilities as infinite matrix products. It lets researchers compute steady states and temporal correlations for nonequilibrium models such as epidemics, where Monte Carlo sampling is slow or noisy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Finite-d accuracy rests on an unproven exponential-decay assumption; a tree benchmark would settle whether truncation error is controlled.","rationale":"The reader's weakest assumption is the heuristic exponential decay of temporal correlations that justifies finite-d truncation; I agree this is the least secure element of the central claim. The rest of the chain - the locally tree-like cavity assumption, the matrix-product update algebra, the VUMPS truncation, and the numerical benchmarks - is standard and credible. The truncation error, however, is asserted, not bounded. The proposed test on a degree-2 regular graph isolates the truncation error because a line is a tree, so BP is exact and any discrepancy is solely due to the iMP approximation. If the test shows exponential-in-d convergence at the predicted rate, the concern is resolved; if not, the paper's claim that 'typically provide highly accurate results even for small bond dimensions' would need qualification. Since the paper already flags this limitation and the numerical validation is credible, I do not change the reader's ACCEPT verdict, but I recommend this test as a follow-up to increase confidence.","tokens_in":14943,"tokens_out":12424,"duration_ms":115946,"concrete_test":"Run EDC on a degree-2 regular graph (a line) for the Ising-Glauber model (11), where Glauber's exact solution provides the steady-state autocovariance (12). For fixed J and h in both paramagnetic and ferromagnetic regimes, compute the EDC autocovariance for increasing d and measure the difference from the exact result. Check whether the difference decays exponentially in d with a rate governed by the correlation time, as the iMP truncation argument in Appendix F predicts. If the decay is consistent, the assumption is confirmed; if the error saturates or decays at a different rate, the finite-d accuracy claim lacks quantitative support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, that the fixed point of Eqs. (7)-(10) gives accurate steady-state observables for small bond dimension d, depends on the assertion in Appendices C and F that temporal correlations in the edge messages decay exponentially, so the iMP truncation error decays exponentially in d. This is stated as 'typically' and is not derived from the BP equations. For a generic locally tree-like graph, the edge message is a hidden Markov process whose exact iMP bond dimension is not known and may be infinite; the approximation error is uncontrolled. Near a critical point, where steady-state dynamics are most interesting, Appendix B shows the required d grows as a power law, so the method loses its advantage exactly there. The paper provides credible but empirical evidence; no a posteriori error bound or convergence test isolates the truncation error from the cavity approximation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an analytic method, eternal dynamic cavity (EDC), for the steady-state dynamics of Markov processes on locally tree-like graphs. Edge messages in dynamic belief propagation are approximated by time-translation-invariant infinite matrix-product (iMP) distributions, parametrized by a single d×d×r² tensor for homogeneous regular graphs. The tensor is determined by a fixed-point equation that is solved by iterating the BP update with VUMPS truncation. The method is applied to symmetric and asymmetric Glauber dynamics and to the SIS model, with benchmarks against equilibrium cavity results, Monte Carlo simulations, and known mean-field approximations. The paper also discusses continuous-time and asynchronous dynamics, and provides appendices on the bond dimension, the iMP ansatz, and the expressibility of cyclic matrix products.","tokens_in":15086,"tokens_out":31789,"duration_ms":269172,"significance":"If the central claims hold, this is a valuable contribution: it provides a way to access steady-state observables and temporal correlations in nonequilibrium Markov processes on sparse graphs without Monte Carlo sampling, and it includes reproducible code and extensive external benchmarks. The derivation of the fixed-point equations from the original transition probabilities, rather than from fitted observables, is a clear strength. The exactness in the large-d limit is supported by an expressibility argument, and the numerical results for small d are impressive, especially for the SIS model where the method outperforms several mean-field approaches.","major_comments":[{"comment":"Equations (G3)-(G6) contain a normalization error and an unjustified step. The cyclic average in Eq. (G3) has T+1 terms, so the prefactor should be 1/(T+1), not 1/T. More importantly, the diagonal blocks of Tr[A(y0)...A(yT)] are products of the form A_s(y0)A_{s+1}(y1)...A_{s-1}(y_T), whereas the QR decomposition of the cyclically shifted message m^T(y_s,...,y_T,y_0,...,y_{s-1}) would naturally produce A_0(y_s)A_1(y_{s+1})...A_T(y_{s-1}). The equality of these two expressions requires an equivariance property of the QR factors that is not stated or proved. Since this appendix is the basis for the claim that the iMP ansatz is exact in the large-d limit, the proof should be repaired or replaced by a precise citation of the known theorem.","section":"Appendix G"},{"comment":"The statement that the finite-d truncation error 'typically decays exponentially' with d is an assumption, not a derived result. The central practical claim of the paper - high accuracy at small bond dimension - rests on this assumption, but no a posteriori error bound or convergence test is provided that isolates the truncation error from the cavity approximation. The near-critical behavior in Appendix B shows that d must grow as a power law, so the method loses its advantage precisely where steady-state dynamics are most interesting. I recommend that the authors either prove a bound under explicit assumptions on the edge-message correlation decay, or add a benchmark on a tree graph (where the cavity assumption is exact) that isolates the truncation error, and state the 'typically' claim as an assumption in the main text.","section":"Appendix C"}],"minor_comments":[{"comment":"The limit of Tr(W^T) is Tr(e π^⊺) or equivalently π^⊺ e, not Tr(π e^⊺) as written; the current expression appears to have the left and right eigenvectors swapped.","section":"Appendix C, Eq. (C4)"},{"comment":"The phrase 'exact in the large-d limit' should be qualified, because Appendix G proves an exact representation only for finite T with bond dimension d ≤ 2r^{T+1}, and the bound diverges as T→∞; the large-d limit of the truncated fixed-point equations is not proven to coincide with the exact infinite-time BP solution.","section":"Abstract and Section VI"},{"comment":"The notation 'F t-1' should read F^{t-1}; as printed, the power is ambiguous.","section":"Appendix F, Eq. (F1)"},{"comment":"There is a typo: 'Altough' should be 'Although'.","section":"Section III"},{"comment":"Please state explicitly how many Monte Carlo samples were used and whether the plotted points are averaged over multiple graph realizations, since the error bars are described only qualitatively.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be a published version (SciPost Phys. 19, 045 (2025)). My main concern is the proof in Appendix G, which is load-bearing for the exactness claim; the normalization error and the unproved block-matrix identity should be fixed. The finite-d accuracy issue is a limitation that the paper partly acknowledges, but it would strengthen the paper to address it with a tree benchmark or an explicit assumption. These are fixable within the scope of the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is a real step forward for cavity-based steady-state dynamics on sparse graphs, and the central construction is sound. The authors take the finite-time matrix-product edge messages from earlier work, push T to infinity, and parametrize time-translation-invariant edge trajectories by a single tensor A. That leaves a fixed-point problem, Eqs. (7)-(10), which they solve with a VUMPS-style fidelity maximization. This is genuinely new relative to the finite-time papers, and it gives access to temporal correlations and correlation times that are hard to get from Monte Carlo. The derivation in Appendices C-G is careful: the cyclic-distribution equivalence in C, the expressibility bound in G, and the boundary-condition discussion in D are all solid. The benchmarks against equilibrium cavity and Monte Carlo are credible, and the SIS comparison against three mean-field methods shows the method does what it claims.\n\nThe soft spot is exactly where the paper is least explicit: the claim that small d works relies on temporal correlations in edge messages decaying fast enough, stated in Appendix C as \"typically\" exponential with no bound. There is no a posteriori error estimate, and near critical points the required d grows as a power law (Appendix B), so the method is least reliable where steady-state dynamics are most interesting. The stress-test note suggesting a tree benchmark is a good one; the current evidence is empirical and convincing but not proof. That said, this is a standard situation for cavity methods, and the authors acknowledge it.\n\nThe paper is not a paradigm shift, but it is a substantive, useful extension with reproducible code. I would send it to a serious referee. The referee should press on the truncation error question and ask for a convergence test that isolates truncation error from the cavity approximation, but this is not a desk-reject.\n\nI'd bring it to reading group and would cite it if I were working on dynamic message passing.","headline":"Solid iMP steady-state extension of matrix-product BP with credible numerics; truncation error control is heuristic but standard.","tokens_in":15577,"tokens_out":1997,"would_cite":true,"duration_ms":19084,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes that steady-state dynamics on locally tree-like graphs can be encoded in a single infinite matrix-product tensor per edge, computed by a fixed-point belief-propagation iteration, exact as the tensor's size grows.","keywords":["nonequilibrium steady state","Markov processes on graphs","dynamic cavity method","infinite matrix products","belief propagation","locally tree-like graphs","Ising-Glauber dynamics","SIS epidemic model"],"falsifier":"Run the SIS model on an infinite degree-3 random regular graph at $\\lambda/\\rho$ near 0.55, compute the proposed method's autocovariances and transfer-matrix gap for bond dimensions $d=10,20,40$, and compare with long Monte Carlo runs. If the inferred correlation time saturates at a finite $d$-independent value while the Monte Carlo autocorrelation continues to show a power-law tail, the claim that finite-$d$ iMP messages capture the true steady-state temporal correlations would be contradicted.","tokens_in":14733,"feed_emoji":"♾️","tokens_out":13754,"duration_ms":134440,"temperature":0.7,"pith_summary":"The paper seeks to make the steady-state dynamics of Markov processes on locally tree-like graphs analytically tractable. It claims that in the infinite-time limit, the edge messages of dynamic belief propagation — conditional probabilities for infinite trajectories on neighboring vertices — can be represented as infinite matrix products, and that for homogeneous regular graphs a single $d\\times d\\times r^2$ tensor $A$ determines the whole steady-state trajectory distribution. The tensor is obtained as the fixed point of a belief-propagation update with truncation, exact as the bond dimension $d\\to\\infty$ and typically accurate already for small $d$. This matters because steady-state observables and, especially, temporal correlations in recurrent-state nonequilibrium models such as the SIS epidemic and asymmetric Glauber dynamics are extremely hard to estimate by sampling, while the proposed method computes them directly and extends to continuous-time dynamics.","feed_headline":"One tensor encodes steady-state dynamics on tree-like graphs","feed_subtitle":"Belief-propagation fixed point yields observables and correlations that Monte Carlo sampling cannot resolve.","key_machinery":"The load-bearing object is the infinite matrix-product (iMP) edge message, an infinite product of $d\\times d$ matrices $A(x_i^t,x_j^t)$ over time steps $t$, one matrix for each pair of states of the two endpoint variables. It represents the joint probability of two infinite trajectories in the cavity system and is time-translation invariant by construction. The paper's eternal dynamic cavity equations close on this object: equation (7) builds an updated message from local transition factors and neighboring messages, equation (8) is an exact SVD/QR split, equation (9) recombines the pieces into the new tensor, and equation (10) is the truncation step that projects back to bond dimension $d$ using a variational fidelity-per-time-step criterion. The fixed point of this iteration yields the steady-state beliefs from which single-time observables, $n$-point correlations, and correlation times are read off via the principal eigenspaces of the associated transfer matrices.","core_discovery":"On the paper's own terms, the discovery is that taking the $T\\to\\infty$ limit of the dynamic cavity equations does not require tracking an ever-growing trajectory catalog: time-translation invariant edge messages $m_A(x_i,x_j)=\\cdots A(x_i^t,x_j^t)A(x_i^{t+1},x_j^{t+1})\\cdots$ are encoded by one tensor $A\\in\\mathbb{R}^{d\\times d\\times r^2}$ per edge type. The fixed point of equations (7)--(10) determines $A$ through a local update that composes the single-step transition factors with Kronecker products of neighboring messages, splits the result by an exact matrix decomposition, and truncates back to bond dimension $d$ by maximizing the fidelity per time step. The paper argues this representation becomes exact as $d\\to\\infty$, that finite-$d$ solutions are typically highly accurate, and demonstrates the claim numerically on symmetric and non-reciprocal Ising-Glauber dynamics on regular graphs and on the SIS model, reproducing known exact solutions and matching Monte Carlo data while resolving correlations that Monte Carlo cannot.","pith_inferences":["One consequence the paper does not develop: because correlation times are obtained from ratios of leading transfer-matrix eigenvalues, the same fixed point could provide direct estimates of dynamic critical exponents near absorbing transitions, without the finite-time scaling extrapolations that Monte Carlo requires.","The same fixed-point structure should apply to other recurrent-state processes on locally tree-like graphs, such as voter models, exclusion processes, kinetically constrained models, and opinion dynamics; the paper lists these as future applications but presents no test, so this is a plausible extension rather than a demonstrated result.","For heterogeneous or disordered graphs, the single-tensor parametrization would become a distribution of tensors evolved by population dynamics; the paper states this is straightforward but does not implement it, and whether the truncation remains stable under disorder averaging is an open question.","Since the iMP transfer matrices carry the temporal correlation spectrum, one could attempt a finite-$d$ extrapolation of the leading eigenvalue ratios to locate phase transitions; this is an inference about a numerical strategy, not a claim in the paper."],"forward_implications":["For homogeneous regular graphs, the whole steady-state trajectory distribution is contained in one $d\\times d\\times r^2$ tensor, so observables and correlation functions can be studied analytically rather than by long simulations.","Temporal autocovariances that decay exponentially and are overwhelmed by Monte Carlo sampling error at large time lags can be evaluated directly from the transfer matrix of the iMP representation.","The method transfers to continuous-time and asynchronous dynamics: taking the discrete-time step $\\Delta t\\to0$ does not increase the computational cost, because the representation still uses a single tensor per time step.","For the SIS model, the steady-state infectious fraction is computed across the endemic transition with accuracy that the paper reports as better than three mean-field closures (recurrent dynamic message passing, individual-based mean field, and the cavity master equation).","Near dynamic phase transitions the required bond dimension grows as a power law, matching the analogy with quantum many-body matrix-product methods; away from criticality moderate $d$ suffices."],"supporting_citations":[{"why":"Provides the individual-based mean-field approximation that the SIS comparison must beat.","marker":"[5]"},{"why":"Provides the recurrent dynamic message-passing baseline for the SIS steady-state comparison.","marker":"[7]"},{"why":"Provides the cavity master equation baseline for the SIS steady-state comparison.","marker":"[8]"},{"why":"Supplies the finite-time matrix-product edge-message construction that the infinite-time iMP ansatz extends.","marker":"[10]"},{"why":"Establishes the matrix-product approximation to the dynamic cavity method and its polynomial cost scaling.","marker":"[11]"},{"why":"Gives the matrix-product belief-propagation framework for stochastic dynamics and the comparison methods used for SIS.","marker":"[12]"},{"why":"Provides the variational uniform matrix-product optimization used to truncate the updated edge messages back to bond dimension $d$.","marker":"[22]"},{"why":"Supplies the tangent-space truncation algorithms that keep the infinite matrix product at fixed bond dimension.","marker":"[24]"},{"why":"Supplies the Perron-Frobenius theorem used to show the time-cyclic distribution reproduces steady-state marginals in the infinite-time limit.","marker":"[49]"},{"why":"Gives the theorem that cyclic edge messages have exact uniform matrix-product representations, the expressibility basis of the iMP ansatz.","marker":"[53]"}],"fun_headline_variants":["Tensor encodes Markov steady states on tree-like graphs","Belief propagation yields accurate steady-state dynamics","Matrix products solve graph dynamics without sampling","One tensor captures temporal correlations beyond Monte Carlo","Analytic steady-state dynamics for locally tree-like graphs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The practical accuracy for small $d$ rests on the assumption that correlations between distant time steps die out quickly, so cutting the infinite product of matrices down to a moderate size discards little probability; near a phase transition, correlations become long-ranged and the needed cut size grows, weakening that advantage.","fun_headline_variants_meta":{"raw":{"variants":["Tensor encodes Markov steady states on tree-like graphs","Belief propagation yields accurate steady-state dynamics","Matrix products solve graph dynamics without sampling","One tensor captures temporal correlations beyond Monte Carlo","Analytic steady-state dynamics for locally tree-like graphs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000269,"raw_usage":{"total_tokens":1628,"prompt_tokens":959,"completion_tokens":669,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":601}},"tokens_in":575,"tokens_out":669,"duration_ms":9220,"temperature":1.0,"reasoning_tokens":601,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:32:16.177248+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the SIS model on an infinite degree-3 random regular graph at $\\lambda/\\rho$ near 0.55, compute the proposed method's autocovariances and transfer-matrix gap for bond dimensions $d=10,20,40$, and compare with long Monte Carlo runs. If the inferred correlation time saturates at a finite $d$-independent value while the Monte Carlo autocorrelation continues to show a power-law tail, the claim that finite-$d$ iMP messages capture the true steady-state temporal correlations would be contradicted.","supporting_citations":[{"cited_title":"Van Mieghem, J","cited_arxiv_id":null,"evidence_quote":"Provides the individual-based mean-field approximation that the SIS comparison must beat."},{"cited_title":"Shrestha, S","cited_arxiv_id":null,"evidence_quote":"Provides the recurrent dynamic message-passing baseline for the SIS steady-state comparison."},{"cited_title":"Ortega, D","cited_arxiv_id":null,"evidence_quote":"Provides the cavity master equation baseline for the SIS steady-state comparison."},{"cited_title":"Barthel, C","cited_arxiv_id":null,"evidence_quote":"Supplies the finite-time matrix-product edge-message construction that the infinite-time iMP ansatz extends."},{"cited_title":"Barthel,The matrix product approximation for the dynamic cavity method, Journal of Statistical Mechanics: Theory and Experiment2020, 013217 (2020)","cited_arxiv_id":null,"evidence_quote":"Establishes the matrix-product approximation to the dynamic cavity method and its polynomial cost scaling."},{"cited_title":"Crotti and A","cited_arxiv_id":null,"evidence_quote":"Gives the matrix-product belief-propagation framework for stochastic dynamics and the comparison methods used for SIS."},{"cited_title":"Zauner-Stauber, L","cited_arxiv_id":null,"evidence_quote":"Provides the variational uniform matrix-product optimization used to truncate the updated edge messages back to bond dimension $d$."},{"cited_title":"Vanhecke, M","cited_arxiv_id":null,"evidence_quote":"Supplies the tangent-space truncation algorithms that keep the infinite matrix product at fixed bond dimension."},{"cited_title":"Gantmakher,The Theory of Matrices(Chelsea Publishing Company, New York, 1959), Vol","cited_arxiv_id":null,"evidence_quote":"Supplies the Perron-Frobenius theorem used to show the time-cyclic distribution reproduces steady-state marginals in the infinite-time limit."},{"cited_title":"Perez-Garcia, F","cited_arxiv_id":null,"evidence_quote":"Gives the theorem that cyclic edge messages have exact uniform matrix-product representations, the expressibility basis of the iMP ansatz."}],"review_version":1}