{"id":"772f1e92-5a25-4122-945b-cc8b5e94bfa6","arxiv_id":"2608.02849","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Traceable graph tensor networks can be decomposed, added, and rounded with SVD sweeps, enabling step-truncation PDE integration with lower storage than tensor trains on the tested examples.","lead":"This paper introduces a decomposition and rounding scheme for tensor networks shaped like arbitrary graphs, so long as the graph contains a path visiting every node. It shows the graph format can beat standard tensor train and hierarchical Tucker formats in storage for selected high-dimensional functions and a Fokker-Planck equation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GTN-Round's per-step tolerance control is asserted, not proven: with chord edges the sweep does not yield independent isometric truncations, so the prescribed-tolerance claim for cyclic graphs lacks support.","rationale":"The GTN-SVD construction is the strongest part of the paper: it inherits a proof from FCTN-SVD, and the numerical results are internally coherent. The central enabling operation for time integration is GTN-Round, however, and Section 3.4 provides no theorem. The heuristic tolerance split in Eq. (24) assumes that truncating path-edge SVDs controls the global error even when chords create loops, but the standard TT argument requires the untouched remainder of the network to be isometric. With chord edges that condition fails, so the local singular-value tail can be amplified. The paper's own language ('in practice', 'we claim no optimality') concedes that the bound is not rigorous. Since the abstract and summary explicitly promise rank truncation controlled to a prescribed tolerance at every time step for traceable GTNs, this is load-bearing. The reader's concern about unoptimized TT and HT baselines is legitimate but affects the magnitude of the reported DOF advantage rather than the correctness of the algorithms, so I only partially agree with the reader's choice of weakest assumption. The proposed random rank-inflation test directly checks whether GTN-Round returns an error below the prescribed tolerance on cyclic graphs; if it fails, the tolerance-control claim should be weakened to an empirical observation conditional on topology, and the paper should be revised accordingly. This supports the reader's conditional verdict rather than changing it.","tokens_in":19459,"tokens_out":10160,"duration_ms":106173,"concrete_test":"Take a random 4D tensor exactly represented in the GTN-BB graph of Figure 10 with chord and path ranks around 10-20, embed it in an equivalent representation with every path rank doubled using random orthogonal factors (an exact rank inflation), and apply GTN-Round with eps = 1e-8. Compute the true relative error against the dense tensor, repeating over 100 random draws. If the observed error exceeds eps by more than a factor of 2 in any trial, the Section 3.4 tolerance-control claim is falsified for cyclic graphs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.4 defines GTN-Round as a right-to-left QR orthogonalization followed by a left-to-right truncated-SVD sweep along a traceable path, explicitly leaving chords untouched. The paper then claims an error bound by setting delta = eps ||X~||_F / sqrt((p-1)+c), saying this 'budgets for truncations at the chords as well' (Eq. 24). This does not follow. In TT rounding, each SVD truncation is safe because, after orthogonalization, the remainder of the network is an isometry. In a GTN with chords, cores carry extra legs connecting non-adjacent positions, so the subnetwork to the right of a truncated bond contains loops and is not isometric; the local singular-value tail can be amplified by the loop contraction. The paper explicitly calls the rounding heuristic and says it claims no optimality; the only support for the bound is the empirical statement that the accumulated error stays below eps 'in practice, uniformly across the topologies we tested.' The central abstract claim that step-rounding time integration controls rank truncation to a prescribed tolerance at every time step is therefore not actually established for traceable GTNs with cycles. A second, smaller internal inconsistency: Algorithm 1 computes delta_k with d (the tensor dimension) while Theorem 1 uses p (the number of cores); for the reported GTN-BB runs p = d+2, so the printed pseudocode would use delta_k for k up to p-1 with d-k negative. Both issues should be settled before the tolerance-control claim is accepted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops SVD-based algorithms for tensor networks with arbitrary graph topologies. It first introduces GTN-SVD, which constructs a decomposition in a user-prescribed graph format by viewing the graph as a subgraph of a complete graph and applying a recursive SVD along unfoldings, with an error bound stated in Theorem 1. It then defines addition, Hadamard product, Frobenius norm, and a rounding procedure, GTN-Round, that operates along a Hamiltonian ('traceable') path through the network while leaving chord edges untouched. The authors apply these algorithms to 6D function approximation and to 4D Fokker-Planck step-truncation time integration, reporting that the barbell GTN format uses substantially fewer degrees of freedom than TT and HT at comparable or better accuracy.","tokens_in":19867,"tokens_out":8331,"duration_ms":78460,"significance":"The paper makes a plausible and useful extension of SVD-constructible low-rank formats beyond TT, HT, and TR to arbitrary graph topologies, with rounding and addition restricted to traceable graphs. The GTN-SVD error bound in Theorem 1 is a genuine contribution, and the addition rule that freezes chord ranks is a sensible mechanism for keeping cyclic networks stable under step truncation. The paper is also honest about the heuristic nature of the rank-splitting rule, the heuristic rounding error control, and the unoptimized TT/HT baselines. If the two load-bearing gaps identified below are resolved, the work would be a significant step toward topology-adaptive tensor computations for high-dimensional PDEs.","major_comments":[{"comment":"The claim that GTN-Round controls the truncation error to a prescribed tolerance at every step is not established. After the QR sweep, a bond on the traceable path has, to its right, a subnetwork that contains chord edges and is therefore generally not an isometry; the discarded singular-value tail can be amplified when the remainder of the network is contracted. Setting delta = eps ||X~||_F / sqrt((p-1)+c) in Eq. (24) does not repair this, because the budget for chord truncations is not tied to any bound on the amplification factor. The paper itself explicitly calls the procedure heuristic and says the error remains below tolerance only 'in practice, uniformly across the topologies we tested.' This matters because the abstract and the output contract of Algorithm 3 assert per-step tolerance control. Please either prove a per-step error bound under explicit assumptions on the chord cores, or remove the prescribed-tolerance claim from the abstract and algorithm contract and add a per-step empirical error diagnostic for the cyclic cases.","section":"§3.4, Eq. (24), Algorithm 3"},{"comment":"Algorithm 1 computes truncation parameters delta_k using the tensor dimension d, while Theorem 1 and its proof use p, the number of cores. For the barbell examples p = d+2, so the loop k = 1,...,p-1 reaches values with d-k < 0 and the square root is undefined. The pseudocode is therefore internally inconsistent and not executable as printed. Please correct the formula to use p, or the quantity actually used in the implementation, and verify that the stated error bound holds for the internal cores.","section":"Algorithm 1 and Theorem 1"},{"comment":"The claimed DOF reductions (e.g., factors 334 and 344 for f1 at eps=1e-8) are measured against tt-toolbox and htucker with default mode orderings and dimension trees. Since TT ranks depend on core ordering and HT ranks depend on the dimension tree, the headline 'substantially fewer degrees of freedom' may be an artifact of not aligning the baselines to the interaction structure of the test functions. The authors acknowledge this, but the conclusion is still stated more strongly than the experiment supports. Please add a sensitivity study with several permuted TT orderings and HT trees, or clearly restrict the claim in the abstract and conclusions to 'default TT/HT implementations.'","section":"§4, before Fig. 8"}],"minor_comments":[{"comment":"In line 10, the SplitRankRule call uses 'rank(U1)' where it should use the current factor U (or U_k); line 11 also has a trailing comma in the reshape list. Please correct these typographical issues.","section":"Algorithm 1"},{"comment":"The text says the barbell's two internal cores are G6 and G7, but the figure appears to label eight cores G1 through G8; please make the core numbering consistent between text and figure.","section":"§4.1, Figure 7"},{"comment":"The entry R_ij = 'rank(i,j)' uses the word 'rank' in a way that could be confused with matrix rank; please define the edge rank explicitly as r_ij before this equation.","section":"Eq. (13)"},{"comment":"The sentence 'Both methods exceed the per-step tolerance eps=1e-9, which is expected' is confusing because the reported final error is an accumulated trajectory error, not a single per-step rounding error; please clarify the distinction.","section":"§4.2.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically interesting and honest about its limitations, but the two main risks are the unsupported per-step tolerance guarantee for GTN-Round and the comparison against unoptimized TT/HT baselines. The Algorithm 1/Theorem 1 p-versus-d inconsistency is easily fixable. I would not recommend rejection, but the authors should either prove or explicitly reframe the rounding claim before the paper is accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: this is a solid algorithmic paper with one real gap. The traceable-path rounding for cyclic tensor networks is new and useful, but the central claim that the rounding controls error to a prescribed tolerance at every time step is heuristic for graphs with chords—there is no proof, and the paper says so itself. The numerical comparison also runs TT/HT without structure optimization, so the DOF reductions are suggestive, not definitive.\n\nWhat's actually new: GTN-SVD turns FCTN-SVD into a user-specified topology by trivializing edges in a rank adjacency matrix; more interesting, Sections 3.2–3.4 define addition, Hadamard product, and rounding for traceable GTNs. The rounding reduces to a TT-style sweep along the Hamiltonian path and leaves chords as extra legs, with a deliberately conservative threshold. For cyclic graphs the sweep's remainder contains loops, so the standard isometry argument doesn't hold; the paper explicitly calls the bound heuristic and supports it only numerically. That is a real gap, and it matters because the abstract's 'rank truncation controlled to a prescribed tolerance at every time step' is the headline.\n\nSmaller but real: Algorithm 1 prints δ_k with d (tensor dimension) while Theorem 1 uses p (number of cores). For the barbell topologies p=d+2, so the pseudocode would take a square root of a negative number for the last cores. Probably a typo, but it needs fixing. Also, the abstract says 'comparable or better accuracy,' yet the η=1 Fokker–Planck run shows GTN-BB at 1.07e-7 final relative error versus TT at 6.23e-9—worse by a factor of about 17. The authors mention this in text but not in the abstract.\n\nThe baseline issue is the one the reader flagged: they use tt-toolbox and htucker with default orderings and say so, and even note structure optimization could improve TT/HT. The factors of 166–382 in DOF are still large, but the test functions are deliberately chosen with interactions matching the barbell topology, so generalizability is untested.\n\nWho should read it: anyone working on tensor formats for high-dimensional PDEs. The idea of using a Hamiltonian path to port TT machinery to cyclic graphs is worth having around. As a referee I'd want code, a corrected Algorithm 1, a tested optimized-TT or at least a rank-optimized comparison, and either a proof of the rounding bound or a qualifier in the abstract ('empirically controlled'). With those changes it would be a good paper. Worth sending to review, but not ready as is.","headline":"A useful new algorithmic idea for tensor network computation, but the central tolerance-control claim for cyclic graphs is heuristic and the numerical comparisons use unoptimized baselines; worth a serious referee but needs revision.","tokens_in":20296,"tokens_out":3369,"would_cite":true,"duration_ms":31799,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A69"],"pacs":[],"model":"deepseek-v4-flash","headline":"SVD-based decomposition, addition, and rounding are defined for tensor networks on any graph topology that admits a Hamiltonian path, enabling step-truncation integration of high-dimensional PDEs directly in graph format.","keywords":["tensor networks","graph tensor networks","SVD decomposition","tensor rounding","tensor train","hierarchical Tucker","Fokker-Planck equation","step-truncation time integration"],"falsifier":"Re-run the six-dimensional f1 decomposition at $\\varepsilon = 10^{-8}$ after optimizing the tensor-train core ordering (e.g., by permutation search) and compare the resulting degrees of freedom with the 47,264 reported for the barbell graph. If the optimized TT representation comes within a small factor of that count, the claimed 334x storage advantage is an artifact of using the default un-optimized TT ordering.","tokens_in":19278,"feed_emoji":"🔗","tokens_out":14273,"duration_ms":116117,"temperature":0.7,"pith_summary":"This paper claims that a tensor network shaped as any graph with a Hamiltonian path, not just the chains of tensor-train or the trees of hierarchical Tucker, can be decomposed, added, and rounded using SVD-based sweeps. That closes the step-truncation workflow: a time-dependent high-dimensional PDE solution can be advanced and recompressed to a prescribed tolerance entirely inside the graph format, never forming the dense tensor. On multivariate test functions and a four-dimensional Fokker-Planck equation, a barbell-shaped graph matches or beats tensor-train and hierarchical-Tucker accuracy while needing far fewer degrees of freedom, sometimes by factors of hundreds. The authors present the advantage as empirical evidence rather than a theorem, and they flag that their baselines use default, un-optimized arrangements.","feed_headline":"382x fewer coefficients: barbell graphs beat tensor trains","feed_subtitle":"A traceable graph network solves 4D Fokker-Planck with per-step error under tolerance and a fraction of the storage","key_machinery":"The load-bearing objects are the rank adjacency matrix $\\mathbf{R}$, whose diagonal entries are physical mode sizes and whose off-diagonal entries are edge ranks (1 for absent edges), and the traceable (Hamiltonian) path through the network, a path that visits every core exactly once. GTN-SVD builds on FCTN-SVD and uses the SplitRankRule to allocate a truncated-SVD rank among the outgoing bonds of each core. The traceable path turns the graph into a TT-like chain with chords; GTN-Round then runs a QR orthogonalization sweep from right to left and a truncated-SVD compression sweep from left to right along that path, budgeting the threshold $\\delta = \\varepsilon \\|\\widetilde{X}\\|_F/\\sqrt{(p-1)+c}$ over path edges and chords. Addition block-concatenates the path edges while max-concatenating the chords, so chord ranks do not grow during repeated addition.","core_discovery":"The central discovery is that any simple graph tensor network is isomorphic to a fully-connected tensor network whose missing edges have bond dimension one, which lets SVD-based decomposition (GTN-SVD) extend to arbitrary user-specified topologies. For computation, the paper shows that a traceable path through the graph orders the cores into a TT-like spine, and that this spine makes two further operations possible: addition, which block-concatenates path edges and overlays chords at the larger of their two ranks, and rounding, which sweeps right-to-left with QR orthogonalization and left-to-right with truncated SVD. Because chords are frozen during addition while path edges are compressed each step, iterative step-truncation remains stable on cyclic networks. The paper reports that this graph-format integrator solves a 4D Fokker-Planck equation with comparable accuracy to TT and HT at lower degrees of freedom and lower overall runtime.","pith_inferences":["Editorial extension: the chord-freezing addition rule suggests a general recipe for any cyclic format: identify a spanning path, concatenate along it, and overlay every other edge, which may let tensor-ring and other looped formats round stably over many steps.","Editorial extension: if the GTN advantage is robust, tensor-network structure search should treat the graph topology itself as the optimization variable, not just the mode ordering inside a fixed format, making GTN a natural target format for such searches.","Editorial extension: the rounding error budget is deliberately conservative because chords are never compressed; a sharper per-chord criterion could lower the achieved ranks without exceeding the stated tolerance, a directly testable improvement.","Editorial extension: because chord ranks freeze during addition, the per-step cost of a traceable GTN integrator should stay nearly constant as the number of steps grows, unlike TT where costs fall as ranks decay; long-time runs could verify whether this changes the crossover point seen in the diffusion experiment."],"forward_implications":["Any graph topology that admits a Hamiltonian path becomes usable for computation: tensors can be decomposed into it, added, and rounded entirely in the format.","Step-truncation time integration of high-dimensional PDEs can run directly in graph format with rank truncation controlled to a prescribed relative tolerance at every step.","In the tested cases, the traceable barbell graph attains comparable or better accuracy than tensor-train and hierarchical Tucker while using dramatically fewer degrees of freedom (a 382x reduction for a 4D function at $\\varepsilon=10^{-8}$; a 166x reduction at the first Fokker-Planck step).","Repeated addition is stable on cyclic traceable networks because path edges are compressed each step while chord edges are frozen, so chord ranks do not inflate over time.","The authors claim the gains as empirical, not universal: there is no theorem that graph formats beat chain or tree formats in general."],"supporting_citations":[{"why":"Supplies the FCTN-SVD recursion and SplitRankRule that GTN-SVD extends; its error bound underpins Theorem 1.","marker":"[17]"},{"why":"Defines the tensor-train format and its SVD-based rounding, which the path-based addition and rounding operations adapt.","marker":"[12]"},{"why":"Defines the hierarchical Tucker format and serves as the HT baseline in the numerical comparisons.","marker":"[13]"},{"why":"Contributes the max-concatenation addition idea for tensor ring that the chord-overlay treatment adapts.","marker":"[19]"},{"why":"Defines step-truncation time integration, the workflow the GTN integrator implements.","marker":"[18]"},{"why":"Provides the default tensor-train implementation used as a baseline in the experiments.","marker":"[21]"},{"why":"Provides the default hierarchical-Tucker implementation used as a baseline in the experiments.","marker":"[22]"},{"why":"Documents that TT ranks depend on core ordering, grounding the caveat about un-optimized baselines.","marker":"[23]"},{"why":"Documents fundamental limitations of rounding tensor formats with loops, motivating the heuristic traceable-path rounding.","marker":"[20]"}],"fun_headline_variants":["Graph tensor networks go cyclic: SVD works on any topology","Traceable graphs make tensor trains generalize to cycles","Cyclic tensor networks: SVD rounding with per-step tolerance","GTN-SVD: arbitrary graphs, addition, rounding for PDEs","Fokker-Planck on graph formats: fewer DOF, same accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported factor-of-hundreds savings compare the new graph formats against off-the-shelf tensor-train and hierarchical-Tucker codes with their default variable orderings; an optimized reordering of those baselines could shrink or erase the advantage.","fun_headline_variants_meta":{"raw":{"variants":["Graph tensor networks go cyclic: SVD works on any topology","Traceable graphs make tensor trains generalize to cycles","Cyclic tensor networks: SVD rounding with per-step tolerance","GTN-SVD: arbitrary graphs, addition, rounding for PDEs","Fokker-Planck on graph formats: fewer DOF, same accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000623,"raw_usage":{"total_tokens":2831,"prompt_tokens":838,"completion_tokens":1993,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":454,"completion_tokens_details":{"reasoning_tokens":1904}},"tokens_in":454,"tokens_out":1993,"duration_ms":11709,"temperature":1.0,"reasoning_tokens":1904,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:57:40.110963+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the six-dimensional f1 decomposition at $\\varepsilon = 10^{-8}$ after optimizing the tensor-train core ordering (e.g., by permutation search) and compare the resulting degrees of freedom with the 47,264 reported for the barbell graph. If the optimized TT representation comes within a small factor of that count, the claimed 334x storage advantage is an artifact of using the default un-optimized TT ordering.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the FCTN-SVD recursion and SplitRankRule that GTN-SVD extends; its error bound underpins Theorem 1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the tensor-train format and its SVD-based rounding, which the path-based addition and rounding operations adapt."},{"cited_title":"Grasedyck, Hierarchical singular value decomposit ion of tensors, SIAM Journal on Matrix Analysis and Applications 31 (4) (2010) 2029–2054","cited_arxiv_id":null,"evidence_quote":"Defines the hierarchical Tucker format and serves as the HT baseline in the numerical comparisons."},{"cited_title":"On Algorithms for and Computing with the Tensor Ring Decomposition","cited_arxiv_id":"1807.02513","evidence_quote":"Contributes the max-concatenation addition idea for tensor ring that the chord-overlay treatment adapts."},{"cited_title":"Rodgers, A","cited_arxiv_id":null,"evidence_quote":"Defines step-truncation time integration, the workflow the GTN integrator implements."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the default tensor-train implementation used as a baseline in the experiments."},{"cited_title":"Kressner, C","cited_arxiv_id":null,"evidence_quote":"Provides the default hierarchical-Tucker implementation used as a baseline in the experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents that TT ranks depend on core ordering, grounding the caveat about un-optimized baselines."}],"review_version":1}