{"id":"4ef93b34-1ab3-4d1c-828e-201d8b76f350","arxiv_id":"2411.13364","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"Matrix Product State approximations of the network epidemic steady state can be tuned by bond dimension and outperform second-order mean-field theory for sufficiently large bond dimensions.","lead":"This paper tests whether Matrix Product States, a compression tool from quantum physics, can approximate the exponentially large probability distribution of an epidemic model on a network. It finds that the approximation becomes systematically more accurate with a tunable bond dimension and beats second-order mean-field theory once enough detail is kept, demonstrated on networks up to the Dutch railway network with 55 nodes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (15) states an invalid DMRG objective for finding the steady state of the non-Hermitian generator Q; all MPS results depend on this, so the method description must be clarified or corrected.","rationale":"I read the paper in good faith. The strongest claim is that MPS with bond dimension of order n beats second-order mean-field by orders of magnitude, and the supporting evidence is the DMRG results in Figs. 4 and 5. That evidence is only meaningful if the DMRG actually computes the steady state. The reader's weakest_assumption targets Eq. (15), the stated DMRG objective, and I agree that this is the single most load-bearing point: the variational principle is mathematically invalid for a non-Hermitian generator, so the paper's method description is internally inconsistent with the claim that it finds p_infinity. I differ mildly from the reader's framing in that the chi=1 behavior is not strong evidence of a wrong algorithm, since the mean-field fixed point need not be the best product-state approximation; the real problem is the literal reading of Eq. (15). The numerical results might still be perfectly correct if the stochasticTN implementation solves local linear systems or minimizes a squared residual rather than p^T Q p. Therefore I do not call for rejection; I support the conditional verdict. The proposed concrete test, inspecting and running the released code on a small system where exact diagonalization is available, would settle whether the concern lands. If the code minimizes Eq. (15), the central claims would be unsupported and rejection would be warranted; if the code does something else, the paper needs only a corrected methodological section and a citation. Either way, the current text is not sufficient as a standalone description of the algorithm.","tokens_in":19718,"tokens_out":9424,"duration_ms":98802,"concrete_test":"Run the public stochasticTN DMRG code on a small ER network (n=10, p=0.3) at lambda=0.55, and compare the converged MPS to (i) the exact steady state from full diagonalization and (ii) the global minimizer of p^T Q p over the same MPS manifold, found by direct local optimization of Eq. (15). If the DMRG output matches (ii) or fails to match (i), the stated objective is the implemented algorithm and the concern lands; if it matches (i), Eq. (15) is a misdescription and the numerical results can stand once the text is corrected.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central numerical claim that MPS approximations become accurate for sufficiently large bond dimension depends entirely on the DMRG algorithm converging to the stationary distribution p_infinity satisfying Q p_infinity = 0. But the objective stated in Eq. (15), minimize p_mps^T Q p_mps, is not a valid variational principle for that problem. For a Markov generator Q with column sums zero, p^T Q p = sum_{i != j} Q_ij p_j (p_i - p_j). This quadratic form is generally indefinite and is not minimized at p_infinity; for example, a symmetric two-state chain gives p^T Q p = -(p1-p2)^2, which is maximized at the uniform steady state and minimized at a pure state. Minimizing Eq. (15) therefore drives the MPS toward boundary distributions, not toward the steady state. The paper gives no proof or citation showing that the stationary vector minimizes this functional, and no such proof can exist in general. Consequently, either the implemented algorithm differs from the stated objective, or the numerical results in Figs. 4, 5, 6, 8, 9, 10, and 12 are not produced by the described method. The paper's observation that chi=1 DMRG moves away from the mean-field solution does not resolve this: the mean-field fixed point is not generally the optimal product-state approximation to the exact steady state, so that behavior is consistent with both a correct and an incorrect objective. The open-source stochasticTN code may implement a different algorithm, but as written the paper does not describe a sound procedure.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a Matrix Product State (MPS) approximation for the steady-state distribution of the epsilon-SIS Markovian spreading model on complex networks. The authors construct an MPO representation of the generator, use a DMRG-type algorithm to find an approximate null vector, and benchmark the resulting MPS against exact diagonalization for n=16 random graphs of four topologies and against kinetic Monte Carlo for a 55-node railway network. They define an entanglement entropy from the MPS singular-value spectrum and use it as a compressibility measure, finding that it saturates for bond dimensions of order n, that it peaks on the endemic side of the phase transition, and that the MPS outperforms second-order mean-field theory for sufficiently large bond dimensions. They also study area/volume scaling of the entanglement entropy and rare-event probabilities.","tokens_in":20038,"tokens_out":9485,"duration_ms":108345,"significance":"If the numerical method is as described, the paper makes a useful contribution: it demonstrates a tunable, tensor-network-based dimensional reduction for non-equilibrium Markovian dynamics on complex networks, with extensive validation against exact diagonalization and MCMC, and it connects MPS compressibility to the physics of the phase transition. The public stochasticTN code, the 100-graph averaging per topology, and the independent railway-network benchmark are concrete strengths that make the central accuracy claims reproducible in principle. The entanglement-entropy saturation criterion for choosing the bond dimension is a practical and falsifiable prescription. However, the paper currently contains a load-bearing ambiguity in the statement of the optimization objective, which must be resolved before the numerical claims can be fully assessed.","major_comments":[{"comment":"The stated DMRG objective is not a valid variational principle for the steady state of a non-Hermitian Markov generator. For a generator Q with column sums zero, p^T Q p = sum_{i != j} Q_ij p_j (p_i - p_j), which is indefinite and is not minimized at the stationary vector satisfying Q p_infinity = 0. For the symmetric two-state chain Q = [[-a,a],[a,-a]], this expression equals -a(p_1-p_2)^2 and is minimized at a pure state, while the steady state is uniform. Thus Eq. (15) cannot justify convergence to Q p_infinity = 0, and the surrounding claim that DMRG searches for a representation close to the exact steady state is unsupported. Every MPS result in Figs. 4-12 depends on this optimization step, so the manuscript must either correct Eq. (15) to the objective actually implemented (e.g., minimization of the residual norm ||Q p||^2 or a linear-system formulation) and describe that algorithm, or provide a valid proof or citation for a variant that is known to work. The observation in Sec. 4.1.1 that chi=1 DMRG moves away from the mean-field solution does not resolve the issue, because that behavior is consistent with both a correct and an incorrect objective. Please also add a convergence test against the exact null vector for the n=16 systems, such as the residual norm ||Q p_mps|| or the overlap with the Lanczos eigenvector, for the DMRG states shown.","section":"Sec. 3.2, Eq. (15)"}],"minor_comments":[{"comment":"The phrase 'the mean-field steady-state solution to (6)' is imprecise, because Eq. (6) is an exact, unclosed evolution equation; the first-order mean-field fixed point should be defined by setting the right-hand side of Eq. (7) to zero.","section":"Sec. 2.2"},{"comment":"There are several typographical errors: 'non-equilibirium' in Sec. 3, 'Fielder' instead of 'Fiedler' in the caption of Fig. 2 and in Sec. 4.1.3, and 'believe propagation' instead of 'belief propagation' in Sec. 5.","section":"Throughout"},{"comment":"The fitted power-law exponents for the area dependence of the entanglement entropy are quoted without fitting ranges or quality measures; given the small range of accessible areas at n=16, reporting fit residuals or confidence intervals would make the area-law claim more robust.","section":"Sec. 4.1.2, Fig. 6"},{"comment":"The statement that rare-event probabilities of order 10^{-14} to 10^{-16} are stable in the singular-value cutoff would be stronger if accompanied by a quantitative convergence measure, such as relative changes between successive cutoffs, rather than visual inspection alone.","section":"Sec. 4.2, Fig. 11"}],"recommendation":"major_revision","confidential_remarks":"The main issue is a method-description gap rather than an observed numerical failure. If the authors can confirm or correct the actual objective used in the DMRG implementation and add residual or overlap tests against exact null vectors for the small systems, the paper should be suitable for publication. The public code makes this verification feasible. I do not see a reason for rejection at this stage, provided the variational-principle issue is fixed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about this one. First, the paper builds a genuinely practical MPS tool for the steady state of epsilon-SIS on networks, and it comes with unusually careful benchmarks: exact diagonalization up to n=20, MCMC on a 55-node railway network, and a public code. Second, the method as written is not what produces the numbers: the DMRG objective in Eq. (15), minimizing p^T Q p, is not a valid way to find the stationary distribution of a non-Hermitian generator. That has to be fixed before the paper can be taken at face value.\n\nThe new content is real. The entanglement entropy diagnostic, and the observation that it peaks on the endemic side of the transition, is a useful addition to the complex-systems toolbox. So is the systematic accuracy comparison: for bond dimension around system size, the MPS beats second-order mean-field by orders of magnitude near criticality. The rare-event probabilities down to 10^-14 on the railway network are a nice demonstration. The Fiedler-vector ordering trick is sensible and reduces MPO bond dimension. Credit where due: the benchmarks are extensive, the truncation-error power law is a useful practical guide, and the numerical claims are validated against independent methods.\n\nThe soft spot is load-bearing. Eq. (15) as stated is not a variational principle for the stationary state. For a Markov generator, p^T Q p is indefinite; for a symmetric two-state chain it equals -(p1-p2)^2, which is maximized, not minimized, at the uniform steady state. Minimizing it drives the MPS toward pure configurations, not toward Qp=0. The paper's observation that chi=1 DMRG moves away from the mean-field solution is consistent with a wrong objective and does not fix the problem. Either the code minimizes something else (for example, the squared norm of the residual ||Q p||^2) or the results in Figures 4-12 are not produced by the described algorithm. The authors need to state the actual objective, or correct Eq. (15), and show that their algorithm converges to Qp=0.\n\nThe rest of the paper holds up. The area-law scaling in the inactive phase and the volume dependence in the endemic phase are interesting but secondary. The 'edge of chaos' language is hand-wavy, but the underlying observation is solid.\n\nWho should read this: anyone working on tensor-network methods for stochastic processes on networks, and complex-systems researchers who need full steady-state distributions for moderate-size networks. It deserves a serious referee, but with a request for a major revision on the DMRG objective. I'd conditionally accept it myself, and I'd want to see the stochasticTN code compared against Eq. (15).","headline":"A useful MPS tool for network SIS steady states, but the stated DMRG objective in Eq. (15) is not a valid way to find a stationary distribution and must be fixed before the method description can be trusted.","tokens_in":20546,"tokens_out":4609,"would_cite":true,"duration_ms":49856,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["89.75.Fb","05.10.-a","02.70.-c"],"model":"deepseek-v4-flash","headline":"Matrix product states compress the $2^n$-dimensional steady state of a network epidemic model, and with bond dimension of order $n$ the compressed distribution beats second-order mean-field theory by orders of magnitude.","keywords":["matrix product states","epsilon-SIS model","epidemic spreading","Markov processes","entanglement entropy","mean-field approximation","dimensional reduction","complex networks"],"falsifier":"A decisive check is to run the identical DMRG minimization of $p^T Q p$ on a small Markov chain with a non-symmetric generator and a known stationary vector; if the converged MPS does not match the exact stationary distribution (or if the $\\chi=1$ fixed point differs from the mean-field solution, as the paper reports), then the optimization objective is not the right fixed-point condition and the accuracy claims need to be re-evaluated.","tokens_in":19522,"feed_emoji":"🦠","tokens_out":8935,"duration_ms":84214,"temperature":0.7,"pith_summary":"The paper claims that the exponentially large steady-state distribution of a Markovian epidemic model on a network can be faithfully compressed into a matrix product state whose bond dimension $\\chi$ is a tunable knob. At $\\chi$ of order the number of nodes $n$, the compressed distribution is more accurate than second-order mean-field theory, and at larger $\\chi$ it beats it by several orders of magnitude, on the four random-graph families tested and on a 55-node railway network. It introduces the Shannon entropy of the singular-value spectrum ('entanglement entropy') as a measure of how compressible the distribution is, and finds that this quantity peaks on the endemic side of the epidemic transition, not exactly at the variance peak. A sympathetic reader would care because this offers a systematic, accuracy-controlled alternative to moment-closure approximations for a whole class of Markovian complex systems, and a computable indicator of when low-dimensional effective descriptions should fail.","feed_headline":"Tensor-network states beat mean-field for epidemic models","feed_subtitle":"Bond dimension tunes accuracy, and compressibility peaks just past the epidemic threshold.","key_machinery":"The object that carries the argument is the matrix product state representation of the probability vector, $p_{i_1\\dots i_n}\\approx \\sum_{\\chi_1,\\dots,\\chi_{n-1}} U^1_{i_1\\chi_1}U^2_{\\chi_1 i_2\\chi_2}\\cdots U^n_{\\chi_{n-1}i_n}$, with bond dimensions $\\chi$ controlling the number of singular values kept between adjacent nodes. The infinitesimal generator of the $\\epsilon$-SIS process is written exactly as a matrix product operator of bond dimension $2+2m$ (with $m$ the number of edges), using the local operators $\\hat q^{si}$, $\\hat q^{is}$ and the infection projector $\\hat n$. The steady state is then approximated by sweeping the DMRG algorithm to minimize $p^TQp$, truncating singular values below $\\delta$ or above $\\chi_{\\max}$; reordering nodes by the graph Laplacian's Fiedler vector reduces the MPO bond dimension and computation time. The compressibility measure is the Shannon entropy of the squared singular values across a bond, $S_{\\mathrm{EE}}(s)=-\\sum_k (s_k^2/|s|^2)\\log(s_k^2/|s|^2)$, which the paper uses to quantify when the distribution is or is not reducible.","core_discovery":"The central discovery is that the steady-state probability vector of the $\\epsilon$-SIS process on an arbitrary network, a vector of length $2^n$, is highly compressible as an MPS away from a narrow window near criticality, and that the compression error is controlled by the truncated singular values. Concretely, the paper shows that with a bond dimension comparable to the system size, $\\chi \\sim O(n)$, the single-node infection probabilities $\\langle \\hat n_i\\rangle$ computed from the MPS are already more accurate than those from a second-order mean-field closure, and that increasing $\\chi$ further reduces the error by several orders of magnitude. It also finds that the entanglement entropy of the MPS peaks just after the phase transition, on the disordered side, and that the bond dimension at which this entropy saturates coincides with the point where the MPS outperforms second-order mean-field. On the 55-node railway network, the MPS matches Monte Carlo results within MCMC error bars and resolves probabilities of rare events that Monte Carlo would need enormous samples to see.","pith_inferences":["Beyond epidemic models, the same construction—an exact MPO for a local Markov generator plus DMRG sweeping—applies to any network Markov process whose transitions are local in the tensor-product basis, so opinion dynamics, SIR/SEIR, and contact processes should be compressible in the same way.","The claim that entanglement entropy peaks just on the endemic side of the transition is made for one specific node ordering; if the peak location is truly ordering-independent, it could serve as a generic nonequilibrium criticality marker, and testing it with tree tensor networks or different orderings would sharpen the claim.","The paper's rare-event probabilities suggest a potentially powerful use: with truncation thresholds as calibrated control parameters, the MPS could estimate large-deviation probabilities where direct Monte Carlo is exponentially costly, a use the paper mentions but does not explore quantitatively."],"forward_implications":["For any of the four random-graph topologies tested, choosing $\\chi$ of order $n$ makes the MPS approximation of node infection rates more accurate than second-order mean-field theory, and larger $\\chi$ improves on it by orders of magnitude.","The bond dimension at which the entanglement entropy saturates is a practical indicator of the $\\chi$ needed to beat mean-field, giving a data-free way to set the approximation's accuracy.","In the inactive and endemic regimes even small $\\chi$ gives good accuracy, while near the entanglement-entropy peak the approximation needs large $\\chi$, so the hardest systems are those just past the epidemic threshold.","The method provides the full $2^n$ distribution in compressed form, so rare-event probabilities such as the chance that exactly $k$ nodes are infected can be extracted at negligible extra cost, with probabilities as low as $10^{-14}$ stable over truncation thresholds.","For the 55-node railway network, MPS density predictions agree with the Monte Carlo reference within the MCMC error bars, while first- and second-order mean-field deviate systematically."],"supporting_citations":[{"why":"Defines the $\\epsilon$-SIS model with spontaneous infection, the process whose steady state is compressed.","marker":"[36]"},{"why":"Derives the node-based mean-field hierarchy from the master equation, giving the first- and second-order baselines compared against.","marker":"[10]"},{"why":"Provides the pair/triple closure used to close the second-order mean-field equations.","marker":"[6]"},{"why":"Adapts DMRG to stochastic systems, the algorithmic basis for finding the steady state as an MPS.","marker":"[23]"},{"why":"Supplies the MPS construction and DMRG review that the compression and sweeping are built on.","marker":"[15]"},{"why":"Prior tensor-network/ALS treatment of epidemic spreading models that this work extends to steady-state distributions on complex networks.","marker":"[24]"}],"fun_headline_variants":["MPS compresses epidemic steady states effectively","Bond dimension tunes accuracy beyond mean-field","Entropy spike reveals edge of chaos in spreading","Tensor networks beat mean-field in epidemic models","Compressible epidemic states: a tensor network approach"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that running DMRG sweeps to minimize $p^T Q p$ produces the steady state of the non-Hermitian Markov generator $Q$; the paper supplies no proof or citation for this, and for a generic Markov chain the quadratic form is not minimized at the stationary distribution, so all MPS results inherit this assumption.","fun_headline_variants_meta":{"raw":{"variants":["MPS compresses epidemic steady states effectively","Bond dimension tunes accuracy beyond mean-field","Entropy spike reveals edge of chaos in spreading","Tensor networks beat mean-field in epidemic models","Compressible epidemic states: a tensor network approach"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000269,"raw_usage":{"total_tokens":1641,"prompt_tokens":984,"completion_tokens":657,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":588}},"tokens_in":600,"tokens_out":657,"duration_ms":7505,"temperature":1.0,"reasoning_tokens":588,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:30:47.010142+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive check is to run the identical DMRG minimization of $p^T Q p$ on a small Markov chain with a non-symmetric generator and a known stationary vector; if the converged MPS does not match the exact stationary distribution (or if the $\\chi=1$ fixed point differs from the mean-field solution, as the paper reports), then the optimization objective is not the right fixed-point condition and the accuracy claims need to be re-evaluated.","supporting_citations":[{"cited_title":"1103/PhysRevE.86.016116","cited_arxiv_id":null,"evidence_quote":"Defines the $\\epsilon$-SIS model with spontaneous infection, the process whose steady state is compressed."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Derives the node-based mean-field hierarchy from the master equation, giving the first- and second-order baselines compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the pair/triple closure used to close the second-order mean-field equations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Adapts DMRG to stochastic systems, the algorithmic basis for finding the steady state as an MPS."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the MPS construction and DMRG review that the compression and sweeping are built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior tensor-network/ALS treatment of epidemic spreading models that this work extends to steady-state distributions on complex networks."}],"review_version":1}