{"id":"a26e906f-9dcc-41c1-bc51-bbc8813d30e8","arxiv_id":"2412.17497","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Densely connected tensor network geometries train to lower infidelity than sparse ones on random quantum states, and a new leaf-contraction trick reduces memory while improving training.","lead":"This paper trains tensor networks with different shapes to copy random quantum states and finds that densely connected layouts reach lower error than sparse chains, even when the chain is big enough to represent the target. The result gives practical guidance for choosing tensor network geometries in quantum machine learning and simulation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The main experiments never specify the surrogate geometry used to generate targets; if that geometry is dense, the reported density advantage could be built into the benchmark rather than discovered.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the surrogate target-generation procedure is underspecified, and if the target distribution is biased toward dense structures, the central claim about connectivity versus capacity collapses. This concern is more fundamental than secondary issues such as missing error bars or best-run selection, because it threatens the external validity of the entire benchmark rather than only the strength of the quantitative claim. The paper's own text supplies the basis for the concern: Sec. II C describes initializing 'a random tensor network with a given structure ansatz' but never identifies that ansatz for the main experiments, and states that recovering the surrogate structure is left to future work. A concrete test comparing MPS-generated, dense-generated, and Haar-random targets would settle whether the observed ordering is a property of gradient training on generic random states or an artifact of target construction. Since the reader already recommended CONDITIONAL acceptance pending clarification and artifacts, this stress-test does not move the verdict; it reinforces the condition.","tokens_in":12751,"tokens_out":7433,"duration_ms":77882,"concrete_test":"Inspect the released code/config, or contact the authors, to determine the surrogate geometry used for the main experiments; if this is unavailable, regenerate the n=16, chi=chiS experiment with three target ensembles: (a) targets from an MPS surrogate, (b) targets from a Star1 surrogate, and (c) uniform Haar-random targets, keeping the optimizer and reporting protocol identical. If the dense-over-sparse ordering does not persist for ensembles (a) or (c), the headline claim is an artifact of the surrogate construction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Sec. V, Figs. 4-5) is that connectivity, not representational capacity, controls gradient trainability. The targets are prepared by contracting a random surrogate tensor network 'with a given structure ansatz' (Sec. II C), but the manuscript never states which surrogate geometry is used in the main n=12 and n=16 experiments, nor whether it is fixed across all model geometries. If the surrogate was a dense or Star-like network, the target states are not neutral 'random states with controlled entanglement' but samples drawn from the dense ansatz's own generative distribution, so the dense model starts with a built-in prior match. Conversely, if the surrogate geometry is matched to each model, then different models are trained on different targets and the cross-geometry comparison in Figs. 4-5 is not well defined. The paper explicitly defers the study of 'encoding different structures into the surrogate' to future work (Sec. II C), which makes this omission consequential. No code or data are released to resolve the ambiguity, so a central empirical conclusion currently rests on an underspecified benchmark.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies gradient-based training of tensor networks with different geometries (MPS, tree-like structures, Star networks, and PEPS) whose task is to encode random quantum states prepared by contracting a random surrogate tensor network into a single dense tensor. The main empirical claim is that more densely connected geometries reach lower infidelities and higher success rates than sparser ones, even when the bond dimension is large enough to represent the target exactly, and that connectivity matters more than total information capacity. The paper also introduces a 'compactification' procedure for tree tensor networks, reports that compact versions train better and use less memory, and presents HPC/GPU timing comparisons. The authors conclude that TN geometry affects trainability beyond representational capacity and connect the density trend to the appearance of barren plateaus.","tokens_in":12983,"tokens_out":4556,"duration_ms":48353,"significance":"If the empirical trend is robust, the paper provides a useful design principle for tensor-network machine learning: for gradient-based training, connectivity can be more important than bond dimension or total parameter count. The compactification observation is a clean, analytically motivated memory-reduction trick for tree networks, and the link to the barren-plateau literature is timely. The main limitation is that the central numerical claim rests on an underspecified benchmark: the surrogate geometry used to generate the target states is never stated for the main experiments, and the reported comparisons do not control for total parameter count or include error bars. The strength of the paper is therefore conditional on additional specification and control experiments, but the core question is well posed and the compactification derivation is a positive contribution.","major_comments":[{"comment":"The surrogate geometry used to generate the target states is never specified for the main n=12 and n=16 experiments. The text says targets are prepared by contracting a random tensor network 'with a given structure ansatz' and then explicitly defers 'the study of encoding different structures into the surrogate' to future work. If the surrogate was a dense geometry, the dense models have an unfair prior match; if the surrogate geometry was matched to each model geometry, different models are trained on different targets and the cross-geometry comparison in Figs. 4 and 5 is not well defined. The manuscript must state the surrogate geometry (and its bond dimension) for every experiment, confirm that it is held fixed across all model geometries, and ideally repeat the main comparison with several surrogate geometries to show that the density trend is not an artifact of the target generation procedure. No code or data are released to resolve this ambiguity.","section":"Sec. II C and Sec. III"},{"comment":"The central infidelity-versus-density trend is reported without error bars or seed statistics. Figure 4 shows single-point infidelities for n=12 and Figure 5a shows single-point infidelities for n=16, while Figure 5b reports only the number of trainings crossing a 10^-3 threshold and Figure 7 reports best and median runs for a single n=16 example. Because the main claim is comparative and the text emphasizes that MPS can fail even after 400 trials, the paper should report the distribution of infidelities over random initializations for each geometry and each bond dimension, including the number of trials, the median, and a dispersion measure. Without this, the reader cannot judge whether the density ordering is statistically meaningful or driven by a few lucky runs.","section":"Figs. 4 and 5"},{"comment":"The main comparisons fix the bond dimension (or the size of the largest tensor) rather than the total number of parameters. At fixed bond dimension, denser geometries contain more tensor entries, so the observed advantage could be explained by parameter count rather than by connectivity. The paper's central conclusion that 'the way the TN structure connects the information that it stores is much more important than the total amount of information it can store' requires at least one control experiment with matched total parameter count, e.g., reducing the bond dimension of dense structures so that their total number of trainable parameters is comparable to that of a sparser geometry. Without such a control, the stated distinction between connectivity and information capacity is not established.","section":"Sec. IV and Sec. V"}],"minor_comments":[{"comment":"The fidelity F is defined as an overlap rather than as the absolute square typically used for pure states; please clarify whether the states are taken to be real or whether F denotes the modulus of the overlap, since the loss function in Eq. (4) takes a logarithm of F.","section":"Eq. (3)"},{"comment":"The 'density' of a geometry is quantified by the maximum node distance, but the graph-theoretic definition of distance for PEPS (which contains loops) is only described informally in the figure caption; a precise definition would help reproducibility.","section":"Sec. II B and Fig. 1"},{"comment":"The hardware description contains an incomplete specification: the accelerated-node CPU is described as 'two Intel Sapphire Rapids 8460Y+ with 40 cores at' with the clock speed missing.","section":"Sec. III"},{"comment":"There are typographical errors such as 'compactification ot tree TN' and 'repreentation' in Sec. II B; a careful proofread is needed.","section":"Sec. V"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely question and the compactification part is a nice contribution. However, the main empirical claim is currently built on an underspecified benchmark, and the absence of error bars and parameter-count controls makes the headline conclusion not yet fully supported. I would be comfortable with publication after the surrogate geometry is specified, the statistics are reported, and a matched-parameter control is added. I also note that no code or data are provided, which is particularly limiting given the ambiguity in the target-generation protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read: this is a useful empirical study, but the main claim about density rests on a benchmark whose details are not disclosed. The trend is consistent across n=12, 16, and 20, and the compactification trick is a genuine practical contribution. But the surrogate geometry used to generate target states is never specified in the main experiments, and that matters.\n\nThe new content: a systematic comparison of gradient-based training across MPS, several tree geometries, and PEPS, with a node-distance heuristic for density. The leaf-contraction compactification for tree tensor networks reduces memory and generally improves training speed and final infidelity. That part is concrete and holds up.\n\nWhat is soft: first, the surrogate procedure (Sec. II C) generates targets by contracting a random TN with 'a given structure ansatz', but the geometry of that surrogate is never stated for the n=12 and n=16 runs. If the surrogate was dense, the dense models start with a prior match; if it was MPS-like, the result is more interesting but currently unknown. The paper explicitly defers studying different surrogate structures to future work, which is exactly where this should have been controlled. This is a load-bearing gap for the central claim that connectivity, not capacity, controls trainability.\n\nSecond, the empirical reporting is light: no error bars, several figures show best runs rather than distributions, and comparisons are at fixed bond dimension without controlling for total parameter count. MPS with the same chi has far fewer parameters than Star or PEPS, so part of the advantage may be parameter count rather than geometry per se. This is not fatal, because the density ordering persists at chi = chi_S, but it needs addressing.\n\nThird, no code or data are released, so the surrogate ambiguity cannot be resolved by inspection.\n\nOn the positive side, the paper is clearly written, cites the relevant barren-plateau and tree-TN literature, and the compactification derivation is sound. The trend is consistent across sizes, which suggests the effect is real, even if its cause is not fully pinned down.\n\nWho should read it: anyone working on tensor-network quantum machine learning or variational TN training. It deserves a serious referee round, but with requests for the surrogate specification, error bars or full distributions, and ideally code/data release. I would accept it conditionally, not as is.","headline":"A useful empirical study with a genuine compactification trick, but the central density claim is undercut by an undisclosed surrogate geometry in the target-generation step.","tokens_in":13433,"tokens_out":3142,"would_cite":false,"duration_ms":28992,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Denser tensor network geometries train faster and fit random quantum states better than sparse ones, even when the sparse network has enough memory to represent the target exactly.","keywords":["tensor networks","gradient-based training","geometry","bond dimension","infidelity","barren plateaus","matrix product states","compact tensor networks"],"falsifier":"Generate the same training benchmark using targets that are exactly Haar-random statevectors, and separately using targets produced by contracting an explicitly sparse surrogate network such as a low-bond-dimension MPS, then compare dense and sparse model geometries; if sparse models match dense ones on either target family, the density advantage is an artifact of the target-generation procedure rather than a property of gradient training.","tokens_in":12574,"feed_emoji":"🕸️","tokens_out":2440,"duration_ms":26914,"temperature":0.7,"pith_summary":"This paper studies gradient-based training of tensor networks used as machine-learning models. It trains several network geometries, from a one-dimensional chain to star-shaped trees and a two-dimensional grid, to encode the same random quantum state, and measures the final infidelity. The central claim is that denser connectivity, meaning shorter paths between tensors, leads to lower infidelity and higher success rates, and that this advantage persists even when the bond dimension is large enough to represent the target state exactly. The authors conclude that how a tensor network connects the information it stores matters more than how much information it can store. They also introduce a compactification procedure for tree-shaped networks that reduces memory and improves training, and they document practical speedups from GPU acceleration on a large supercomputer for the largest system sizes.","feed_headline":"Denser tensor networks train better even with full memory","feed_subtitle":"How a network connects its tensors, not just how many it stores, decides gradient-training success on random quantum states.","key_machinery":"The surrogate training protocol is the central tool: a random tensor network with a chosen geometry and bond dimension is contracted into a single dense tensor, producing a target quantum state with controlled entanglement but no visible network structure; the model tensor network is then trained against this single surrogate through one contraction per step rather than through statistical sampling. This isolates the effect of geometry from sampling artifacts. The paper also uses a compactification procedure, contracting the outer branches of a tree network until a bond of maximal dimension is reached, which lowers total memory while slightly increasing contraction cost; node-to-node distance in the network graph serves as the operational measure of density.","core_discovery":"For gradient-based training of tensor networks to encode random quantum states, network geometry dominates representational capacity. When the target state is generated with a fixed amount of controlled entanglement and the model network is trained by minimizing a log-infidelity loss with L-BFGS-B and automatic differentiation, the final infidelity decreases as the network becomes more densely connected, as measured by the maximum distance between nodes. Dense structures such as the star geometry and PEPS reach low infidelity at low bond dimension, while MPS and large-diameter trees need to reach the maximal Schmidt bond dimension and, in some cases, still fail to converge. Increasing the bond dimension beyond what is needed for exact representation does not improve the infidelity and, for sparse structures, lowers the fraction of successful trainings. The compact version of a tree network, obtained by contracting peripheral bonds with dimension below the maximum, both reduces total memory and trains to better infidelity in fewer iterations.","pith_inferences":["A testable extension would be to generate surrogate targets from deliberately sparse networks, such as low-bond-dimension MPS, and check whether dense geometries still win; if the density advantage disappears for sparse targets, the benefit is tied to matching geometry to target entanglement rather than to density per se.","The result suggests that for classical machine-learning tensor train (MPS-style) models, reshaping the data layout to shorten effective paths, for example through feature rearrangement, could improve gradient training without changing the total parameter count.","The observed failure of MPS at maximal bond dimension hints that the loss landscape, not representational power, is the bottleneck; quantifying the spectral properties of the Hessian for different geometries would make the mechanism precise.","Because compactification effectively increases density, the same procedure could be applied to loop-containing geometries such as PEPS, though the paper only demonstrates it for tree networks."],"forward_implications":["If geometry rather than capacity controls trainability, then practitioners should choose tensor-network ansatze whose connectivity mirrors the presumed correlation structure of the target data, rather than merely increasing bond dimension.","Barren-plateau effects in tensor-network training should be viewed as a continuous function of network density: MPS sits at the sparse extreme, while star and PEPS geometries are largely immune in the tested regime.","The compactification procedure gives a memory reduction for tree tensor networks with many leaves, and in the tested cases it also improves training quality and speed, so it can be used to raise the achievable precision under a fixed memory budget.","Because the best geometry depends on the entanglement structure of the task, automatic structural optimization of tree tensor networks, referenced in the paper, becomes a natural complement to density-oriented training.","For large enough system sizes, GPU-accelerated nodes yield a time and energy advantage over CPU-only runs, provided double precision is used; single precision degrades the achievable infidelity by one to two orders of magnitude."],"supporting_citations":[{"why":"Provides the tensor-network notation and the MPS contraction formula that the paper uses to define all geometries.","marker":"[22]"},{"why":"Supplies the variational MPS and PEPS framework and the bond-dimension-to-entanglement correspondence the paper relies on.","marker":"[9]"},{"why":"Defines the projected entangled pair states geometry that serves as one of the dense comparison structures.","marker":"[7]"},{"why":"Establishes the prior link between tensor-network structure and barren plateaus that the paper extends to a density dependence.","marker":"[15]"},{"why":"Reports the presence and absence of barren plateaus in tensor-network machine learning, the baseline the MPS failure is compared against.","marker":"[16]"},{"why":"The library used for tensor-network contractions and for integrating automatic differentiation into the training loop.","marker":"[50]"},{"why":"Provides the automatic differentiation and GPU execution backend used for the gradient-based training and the acceleration comparison.","marker":"[51]"}],"fun_headline_variants":["Dense tensor network geometry boosts gradient training fidelity","How tensor network shape decides training success","Density beats sparsity in tensor network training","Tightly connected tensor nets train better","Network geometry, not size, key to tensor training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The surrogate targets are meant to be representative random states with controlled entanglement, but the paper does not specify the geometry or bond dimension of the surrogate network used to generate the main targets, so the density advantage could partly reflect the hidden structure of the targets themselves.","fun_headline_variants_meta":{"raw":{"variants":["Dense tensor network geometry boosts gradient training fidelity","How tensor network shape decides training success","Density beats sparsity in tensor network training","Tightly connected tensor nets train better","Network geometry, not size, key to tensor training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000153,"raw_usage":{"total_tokens":1181,"prompt_tokens":892,"completion_tokens":289,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":220}},"tokens_in":508,"tokens_out":289,"duration_ms":3567,"temperature":1.0,"reasoning_tokens":220,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:27:44.519569+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate the same training benchmark using targets that are exactly Haar-random statevectors, and separately using targets produced by contracting an explicitly sparse surrogate network such as a low-bond-dimension MPS, then compare dense and sparse model geometries; if sparse models match dense ones on either target family, the density advantage is an artifact of the target-generation procedure rather than a property of gradient training.","supporting_citations":[{"cited_title":"Antenna” structure that reduces the distance between sites with respect to an MPS, with- out allowing tensors with more than 3 virtual indices, whereas“Balanced","cited_arxiv_id":null,"evidence_quote":"Provides the tensor-network notation and the MPS contraction formula that the paper uses to define all geometries."},{"cited_title":"Verstraete, V","cited_arxiv_id":null,"evidence_quote":"Supplies the variational MPS and PEPS framework and the bond-dimension-to-entanglement correspondence the paper relies on."},{"cited_title":"Cervero Mart ´ ın, K","cited_arxiv_id":null,"evidence_quote":"Establishes the prior link between tensor-network structure and barren plateaus that the paper extends to a density dependence."},{"cited_title":"size of largest tensor","cited_arxiv_id":null,"evidence_quote":"Reports the presence and absence of barren plateaus in tensor-network machine learning, the baseline the MPS failure is compared against."},{"cited_title":"On the Trainability and Classical Simulability of Learning Matrix Product States Variationally","cited_arxiv_id":"2409.10055","evidence_quote":"The library used for tensor-network contractions and for integrating automatic differentiation into the training loop."}],"review_version":1}