Pith. sign in

REVIEW 2 major objections 5 minor 46 references

Generative-enhanced optimization for knapsack problems: an industry-relevant study

T0 review · 2 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that TN-GEO and STN-GEO, two tensor-network generative optimizers, produce solutions of quality comparable to simulated annealing on 60 multi-knapsack instances, while never beating it and costing more.

desk verdict Useful benchmark of TN-GEO on multi-knapsack, but the parity claim with SA is an oracle-envelope result, not a typical-config result. read the letter →

arxiv 2502.04928 v1 pith:VE566CWA submitted 2025-02-07 cs.LG quant-ph

Yelyzaveta Vodovozova , Abhishek Awasthi , Caitlin Jones , Joseph Doetsch , Karen Wintersperger , Florian Krellner , Carlos A. Riofrío This is my paper · ORCID
classification cs.LGquant-ph MSC 90C2781P6868T07
keywords tensornetworksmatrixproductstatesgenerative-enhancedoptimizationmulti-knapsackproblemsimulatedannealingsymmetricquantum-inspiredalgorithmscombinatorial
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Generative-enhanced optimization (GEO) replaces a hand-designed search heuristic with a trained generative model that learns which solutions are cheap. The paper implements two versions — one using a plain matrix product state with an integer encoding, one using a symmetric matrix product state with a binary encoding — and tests them on 60 multi-knapsack instances in which every object must be assigned to a knapsack. The central claim is that both versions find valid solutions of quality comparable to simulated annealing, and that no instance was found where GEO beat simulated annealing. The authors conclude that the method is a viable but more costly alternative for this industrially relevant problem class.

What carries the argument

The machinery is the GEO loop: candidate solutions are selected, assigned softmax probabilities with inverse temperature $\beta$, used to train an MPS generative model via a two-site DMRG-inspired sweep minimizing negative log-likelihood, and then new candidates are drawn by perfect sampling and merged back into the training set. The constraint machinery is the symmetric MPS: imposing U(1) particle-number conservation on the tensor legs forces every sampled bitstring to satisfy the cardinality constraint that each object is assigned to exactly one knapsack. The paper uses two encodings: an integer encoding (each site is a knapsack choice) with an unconstrained MPS, and a binary encoding (one site per object-knapsack pair) with a symmetric MPS; the symmetric version shrinks the feasible search space from $2^{MN}$ to $M^N$.

What would settle it

Run the same 60 instances and search the full 384-configuration grid for any configuration where TN-GEO or STN-GEO finds a strictly better valid solution than the single fixed simulated-annealing protocol; finding even one such instance would contradict the paper's claim that no such situation exists, and finding none would confirm parity in the tested range.

Watch

Extended reading notes

Core claim

The discovery is that an MPS-based generative model can be trained within the GEO loop to concentrate sampling probability on low-cost feasible solutions of a generalized multi-knapsack problem, and that the resulting solver matches simulated annealing in solution quality. The symmetric-MPS variant hard-codes the assignment constraint 'each object goes to exactly one knapsack' via U(1) charge conservation, which cuts the sampled search space from $2^{MN}$ to $M^N$. Across 60 instances, both encodings yielded a ratio of best-found valid solution to exact-optimizer optimum comparable to an ensemble simulated annealer, and both clearly outperformed random sampling on hard instances. The paper additionally reports that GEO never surpassed the simulated-annealing baseline and is substantially more expensive per solution.

Load-bearing premise

The parity claim rests on comparing GEO's best result over 384 per-instance hyperparameter settings, chosen after seeing the outcomes, against a single fixed simulated-annealing protocol; if that asymmetry were removed, GEO's typical performance would likely be lower.

Editorial extensions

If this is right

  • TN-GEO and STN-GEO can be applied to multi-knapsack problems with a documented encoding prescription and reach solution quality comparable to simulated annealing.
  • The symmetric-MPS encoding guarantees that every sampled candidate respects the assignment constraint, so the search space is $M^N$ instead of $2^{MN}$.
  • Small bond dimensions ($\chi=4$) and few training epochs ($N_e=1$\textendash 3) generalize better than larger ones, so overfitting is the main performance risk.
  • On hard instances where random search finds no valid solution, GEO remains competitive with simulated annealing, but at higher computational cost.
  • The paper found no instance where GEO beat simulated annealing, so GEO is not currently a superior replacement for SA on this problem class.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the reported GEO performance is the best over 384 hyperparameter combinations chosen after seeing the results, the parity with simulated annealing is an upper bound; a fixed a priori configuration would likely do worse, so the practical gap to SA is probably larger than the headline claim.
  • The MPS generator can only capture short-range correlations along its one-dimensional chain; the paper's own weight-ordering heuristic works only for small instances, which suggests that the method's advantage, if any, will appear on problems with local structure rather than general industrial instances.
  • A testable extension is to replace the MPS with a tree or comb tensor network to capture longer-range correlations; if parity with SA persists on instances where random search fails, GEO could become useful as a warm-start or constraint-specialized solver.
  • The paper's 'similar quality at higher cost' result implies that the value of TN-GEO for industry is not raw solution quality but the ability to encode hard constraints directly into the sampler; a fair comparison should therefore measure constraint satisfaction and sample diversity, not just best cost.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper proposes two tensor-network-based generative-enhanced optimization (GEO) methods for a generalized multi-knapsack problem with assignment constraints: TN-GEO with an integer encoding and STN-GEO with a binary encoding using symmetric matrix product states. It details the DMRG-inspired training, perfect sampling, and four selection strategies, and benchmarks the methods on 60 instances of size 4–58 objects and 4–10 knapsacks against random sampling and simulated annealing. The main empirical claim is that both GEO variants produce results of similar quality to simulated annealing on these instances, while being more computationally expensive. The paper also reports a hyperparameter study showing smaller bond dimensions and fewer epochs perform better, and discusses limitations related to symmetry preservation, overfitting, and generalization.

Significance. The paper provides a careful implementation-oriented description of TN-GEO/STN-GEO for a constrained industry-relevant problem, including a frank discussion of limitations and an explicit statement that GEO did not outperform simulated annealing. The benchmark set of 60 instances with Gurobi-computed optima and the inclusion of random-sampling and SA baselines are useful reference points. However, the central parity claim is weakened by the per-instance oracle selection of hyperparameters: reported GEO results are maxima over 384 configurations, so the abstract's 'similar quality' is an upper-bound claim rather than a typical-performance claim. The paper's main value is therefore as a methodology study and negative result, not as a demonstration of competitive typical performance.

major comments (2)
  1. [§III-B, Fig. 14 caption, Table IV] The central claim of parity with simulated annealing is established under an asymmetric evaluation protocol. For each of the 60 instances, the reported GEO/STN-GEO result is the maximum over 384 hyperparameter configurations (4 selection strategies × 3 β × 2 α × 4 N_e × 4 χ), each run 10 times, with the configuration selected after observing results (Fig. 14 caption: 'selecting the configuration that maximizes the number of valid solutions at convergence...'). Table IV shows the winning configuration varies from instance to instance (e.g., selection 'all' vs 'best', β ∈ {0.1, 0.01, 0.001}, N_e up to 10, χ up to 32), so no single configuration reproduces the plotted curves. Simulated annealing is run with one fixed protocol (Sec. II-G). Therefore the abstract's statement that TN-GEO and STN-GEO 'produce results of similar quality to simulated annealing' is an upper-envelope claim, not a typical-result claim. A practitioner who commits to a single configuration without instance-specific oracle knowledge is not guaranteed the reported quality, and the paper's own conclusion notes GEO struggles to generalize from random initialization. To support the parity claim, the authors should either report performance for a fixed configuration chosen a priori (e.g., via a separate tuning set or the configuration recommended in Sec. IV), or explicitly reframe the claim as best-over-hyperparameter-search and provide the distribution of results across configurations.
  2. [§II-D (binary and integer cost functions)] The penalty coefficient c_p for the inequality constraints is never given a numerical value; the text only says it is 'big enough' in both encodings. Since the softmax in Eq. (2) depends on β·c_p, this parameter directly controls the effective feasibility pressure, and the paper's own conclusion attributes GEO's failure on larger instances to heavily penalized costs in the softmax (Sec. IV). Without the value of c_p used in the experiments, the benchmark results are not reproducible and the scaling analysis cannot be properly interpreted. Please specify c_p (or the rule for setting it) for each encoding and for the different problem sizes, and state whether it was included in the hyperparameter sweep.
minor comments (5)
  1. [Appendix B, Eq. (15)] In the derivation of the gradient, the third displayed equation reads (Z/Z′ − 2Ψ′(x)/Ψ(x)), but the correct intermediate expression is (2Ψ′(x)/Ψ(x) − Z′/Z). The final Eq. (15) is correct once Z=1 is used, but the incorrect intermediate step will confuse readers and should be fixed.
  2. [§II-G] The SA protocol sets the final temperature T_final = 1, while the initial temperature T_initial is half the standard deviation of the initial ensemble. If T_initial < 1 for some instances, the exponential cooling rate exp((1/n_iter)·ln(T_final/T_initial)) would produce heating rather than cooling. Please clarify the intended schedule or restrict T_initial > T_final.
  3. [§III-A] The text says 'we have observed worse performance on small problem instances for larger values of α (e.g. 0.1)', but Table II lists only α ∈ {0.001, 0.0001}. Please explain where the value 0.1 comes from, or remove it.
  4. [Table IV] The column header 'V R' is not defined in the table or its caption; the meanings of V and R are only introduced in Sec. III-B and Fig. 14. Please add a definition in the table caption.
  5. [§IV] The conclusion states 'we could not find a situation in which GEO produced better results than the standard simulated annealing method', but Fig. 15 shows a case (N=39, M=10) where the binary GEO found a solution while simulated annealing did not. Although the authors call this an outlier, the statement should be qualified to avoid contradiction.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central claim is benchmarked against external references (Gurobi optima and a fixed simulated annealing protocol), and the per-instance best-hyperparameter selection is a fairness limitation, not a definitional reduction.

full rationale

The paper's central claim is that TN-GEO and STN-GEO produce results of similar quality to simulated annealing on 60 externally generated multi-knapsack instances. The numerical targets come from the Gurobi exact solver, and the comparison baseline is an independently specified simulated annealing protocol with its own fixed schedule and Metropolis acceptance criterion. The GEO results are obtained by running many hyperparameter configurations and, per instance, selecting the configuration that maximizes the fraction of valid solutions and then the ratio to the optimal cost (Fig. 14 caption). This is a best-case or upper-envelope benchmark rather than a prediction from a fixed configuration, and it is correctly disclosed in the figure caption; however, selecting the best configuration and then reporting that same configuration's metrics is not a fitted parameter renamed as a prediction, because no claim is made that the selected configuration generalizes to unseen instances and the reported parity could in principle have failed even under this selection. The algorithmic derivation is self-contained: the DMRG-inspired training gradient is derived in Appendix VI-B with explicit tensor diagrams, the symmetric MPS constraint construction follows the cited U(1) symmetry formalism rather than an assertion by the present authors, and the softmax probability assignment and perfect sampling steps are stated as explicit equations. There are no load-bearing self-citations that substitute for independent evidence, and no equation in the paper reduces to its own input by construction. The main concerns raised by the evaluation protocol are benchmark fairness and statistical overfitting to the test instances, which are correctness and reporting risks rather than circularity. I therefore find no significant circularity and score 0.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the GEO training paradigm (softmax-weighted NLL with MPS generators), the U(1)-symmetric encoding for assignment constraints, and a benchmark design that includes exact optimal solutions from Gurobi. The main free parameters are the penalty coefficient and the oracle-selected hyperparameters. No new physical entities are introduced.

free parameters (2)
  • penalty coefficient cp
    Used in both encodings' cost functions to enforce inequality constraints; the paper only says 'big enough' (Sec. II-D). The results depend on this value, which is never reported.
  • hyperparameter configuration (beta, alpha, Ne, chi, selection strategy) = beta=0.1/0.01/0.001, alpha=0.001/0.0001, Ne=1/3/5/10, chi=4/8/16/32, selection in {all,best,symmetric,best symmetric}
    Scanned in Sec. II-G and selected per instance as the best in Table IV; the central parity result uses these oracle-selected values.
assumptions (5)
  • standard math SVD/QR decomposition of a U(1)-symmetric tensor preserves the symmetry block structure
    Invoked in Sec. II-E and Appendix VI-C to keep the STN symmetric during training.
  • standard math Perfect sampling from a right-canonical MPS with unit norm yields exact samples from the Born distribution
    Used in Sec. II-F; relies on Ferris and Vidal [22].
  • domain assumption The softmax-weighted NLL loss with DMRG-style gradient descent makes the MPS generator assign increasing probability to low-cost solutions
    This is the core GEO mechanism, inherited from [11]; the paper calls the generalization 'somewhat unexpected' (Sec. IV), so it is empirical rather than proven.
  • domain assumption A sufficiently large penalty coefficient cp makes infeasible solutions unattractive while preserving the optimum
    The cost functions in Sec. II-D include cp max(0,...) penalties; no value or rule for cp is given, and the results depend on it.
  • ad hoc to paper The generated benchmark instances are representative of industry-relevant multi-knapsack problems
    Instances are constructed by random weights/values and then reduced by Gurobi if unassigned objects remain (Sec. II-G), which may exclude hard instances with tight assignment constraints.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generative-enhanced optimization for knapsack problems: an industry-relevant study." pith.science (2026). https://pith.science/paper/VE566CWA

@misc{pith2026250204928,
  author       = {Pith},
  title        = {Pith review of: Generative-enhanced optimization for knapsack problems: an industry-relevant study},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VE566CWA}},
  note         = {Machine review of arXiv:2502.04928}
}
read the original abstract

Optimization is a crucial task in various industries such as logistics, aviation, manufacturing, chemical, pharmaceutical, and insurance, where finding the best solution to a problem can result in significant cost savings and increased efficiency. Tensor networks (TNs) have gained prominence in recent years in modeling classical systems with quantum-inspired approaches. More recently, TN generative-enhanced optimization (TN-GEO) has been proposed as a strategy which uses generative modeling to efficiently sample valid solutions with respect to certain constraints of optimization problems. Moreover, it has been shown that symmetric TNs (STNs) can encode certain constraints of optimization problems, thus aiding in their solution process. In this work, we investigate the applicability of TN- and STN-GEO to an industry relevant problem class, a multi-knapsack problem, in which each object must be assigned to an available knapsack. We detail a prescription for practitioners to use the TN-and STN-GEO methodology and study its scaling behavior and dependence on its hyper-parameters. We benchmark 60 different problem instances and find that TN-GEO and STN-GEO produce results of similar quality to simulated annealing.

Figures

Figures reproduced from arXiv: 2502.04928 by the authors.

Figure 1
Figure 1. GEO pipeline. The process begins with the initial pairs of observations and their costs (step 0) and proceeds by selecting [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Diagram of a Matrix Product State with L = 6 elements. The MPS representation allows for an efficient description of the quantum state, as the number of parameters scales polynomially with the system size L, instead of exponentially as in the general case. This is useful for approximating many￾body quantum states with limited entanglement classically. Note that any quantum state can be expressed as an MPS, but the M… view at source ↗
Figure 3
Figure 3. Schematic of the charge conservation for a symmetric [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Schematic of the charge conservation in a MPS with [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: MPS with integer encoding for a knapsack problem [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 9
Figure 9. Figure 9: Update of T [i] with reshaped U. = T [i+1] ki ni+1 ki+1 ΛV † χ ni+1 ki+1 ⃗b ⃗b [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]
Figure 7
Figure 7. Figure 7: Updating the merged tensor at i = 2. 4. Decomposing via SVD: Decompose the updated tensor T [i,i+1] via SVD. First, reshape T [i,i+1] into a matrix T˜ by combining the legs ki−1 and ni into a single index, and ki+1 and ni+1 into another: T˜ dki−1+ni,dki+1+ni+1 = T [i,i…
Figure 8
Figure 8. Figure 8: SVD of the reshaped merged tensor with truncation [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 11
Figure 11. Figure 11: Example of computing the gradient of the NLL loss w.r.t. [PITH_FULL_IMAGE:figures/full_fig_p008_11.png]
Figure 12
Figure 12. Figure 12: This figure illustrates the optimization outcomes for a problem instance involving [PITH_FULL_IMAGE:figures/full_fig_p010_12.png]
Figure 13
Figure 13. Figure 13: Heatmaps displaying optimization performance across various parameter settings for integer and binary encoding [PITH_FULL_IMAGE:figures/full_fig_p012_13.png]
Figure 14
Figure 14. Figure 14: Each method is run 10 times for every problem size (Integer and Binary also use different hyper-parameters from [PITH_FULL_IMAGE:figures/full_fig_p013_14.png]
Figure 15
Figure 15. Figure 15: Subset of tested problem instances in which random [PITH_FULL_IMAGE:figures/full_fig_p013_15.png]
Figure 16
Figure 16. Figure 16: Tensor diagram of Ψ(x). ∂L ∂T[i,i+1] = X x∈T p(x) ·  1 P(x) · ∂P(x) ∂T[i,i+1]  = X x∈T p(x) · [PITH_FULL_IMAGE:figures/full_fig_p015_16.png]
Figure 17
Figure 17. Figure 17: Tensor diagram of ∂Ψ(x) ∂T[2,3] . To simplify this equation further, we use the fact that the MPS has the canonical center at i, and that the MPS is normalized during the whole optimization process. Using these properties we conclude that Z = 1 and Z ′ = T [i,i+1], wh…
Figure 18
Figure 18. Figure 18: Tensor diagram of Z of a mixed canonical MPS. Z ′ = = [PITH_FULL_IMAGE:figures/full_fig_p015_18.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 23 canonical work pages

  1. [1]

    Martello and P

    S. Martello and P. Toth, Knapsack problems: algorithms and computer implementations. USA: John Wiley & Sons, Inc., 1990

    work page 1990

  2. [2]

    A quantum approximate optimization algorithm,

    E. Farhi, J. Goldstone, and S. Gutmann, “A quantum approximate optimization algorithm,” arXiv preprint arXiv:1411.4028 , 2014

    arXiv 2014

  3. [3]

    Evidence of scaling advantage for the quantum approximate optimization algorithm on a classically intractable problem,

    R. Shaydulin, C. Li et al. , “Evidence of scaling advantage for the quantum approximate optimization algorithm on a classically intractable problem,” Science Advances , vol. 10, no. 22, p. eadm6761,

    work page

  4. [4]

    Quantum speedups in solving near- symmetric optimization problems by low-depth qaoa,

    A. Montanaro and L. Zhou, “Quantum speedups in solving near- symmetric optimization problems by low-depth qaoa,” 2024. [Online]. Available: https://arxiv.org/abs/2411.04979

    arXiv 2024

  5. [5]

    A fast quantum mechanical algorithm for database search,

    L. K. Grover, “A fast quantum mechanical algorithm for database search,” in Proceedings of the Twenty-Eighth Annual ACM Symposium on Theory of Computing , ser. STOC ’96. New York, NY , USA: Association for Computing Machinery, 1996, p. 212–219. [Online]. Available: https://doi.org/10.1145/237814.237866

    arXiv 1996

  6. [6]

    Opening the black box inside grover’s algorithm,

    E. M. Stoudenmire and X. Waintal, “Opening the black box inside grover’s algorithm,” Phys. Rev. X , vol. 14, p. 041029, Nov

    work page

  7. [7]

    Quantum computing in the nisq era and beyond,

    J. Preskill, “Quantum computing in the nisq era and beyond,” Quantum, vol. 2, p. 79, 2018. [Online]. Available: https://doi.org/10.22331/ q-2018-08-06-79

    work page 2018

  8. [8]

    Available: https://link.aps.org/doi/10.1103/PhysRevX

    [Online]. Available: https://link.aps.org/doi/10.1103/PhysRevX. 14.041029

    doi:10.1103/physrevx

Show all 46 references
  1. [9]

    Quantum fourier transform has small entanglement,

    J. Chen, E. Stoudenmire, and S. R. White, “Quantum fourier transform has small entanglement,” PRX Quantum , vol. 4, p. 040318, Oct 2023. [Online]. Available: https://link.aps.org/doi/10.1103/ PRXQuantum.4.040318

    2023

  2. [10]

    A quantum-inspired approach to exploit turbulence structures,

    N. Gourianov, M. Lubasch et al. , “A quantum-inspired approach to exploit turbulence structures,” Nature Computational Science , vol. 2, no. 1, pp. 30–37, Jan 2022. [Online]. Available: https: //doi.org/10.1038/s43588-021-00181-1

    2022 doi

  3. [11]

    Enhancing combinatorial optimization with classical and quantum generative models,

    J. Alcazar, M. Ghazi Vakili et al. , “Enhancing combinatorial optimization with classical and quantum generative models,” Nature Communications, vol. 15, no. 1, p. 2761, Mar 2024. [Online]. Available: https://doi.org/10.1038/s41467-024-46959-5 16

    2024 doi

  4. [12]

    CMA-ES/pycma on Github,

    N. Hansen, Y . Akimoto, and P. Baudis, “CMA-ES/pycma on Github,” Zenodo, DOI:10.5281/zenodo.2559634, Feb. 2019. [Online]. Available: https://doi.org/10.5281/zenodo.2559634

    2019 doi

  5. [13]

    Cinelli, M

    L. Cinelli, M. Marins et al. , Variational Methods for Machine Learning with Applications to Deep Networks . Springer International Publishing, 2021. [Online]. Available: https://books.google.de/books? id=N5EtEAAAQBAJ

    2021

  6. [14]

    Protes: Probabilistic optimization with tensor sampling,

    A. Batsheva, A. Chertkov et al. , “Protes: Probabilistic optimization with tensor sampling,” 2023. [Online]. Available: https://arxiv.org/abs/ 2301.12162

    2023 arXiv

  7. [15]

    Variational quantum algorithms,

    M. Cerezo, A. Arrasmith et al. , “Variational quantum algorithms,” Nature Reviews Physics , vol. 3, no. 9, pp. 625–644, 2021. [Online]. Available: https://doi.org/10.1038/s42254-021-00348-9

    2021 doi

  8. [16]

    Generative adversarial networks,

    I. J. Goodfellow, J. Pouget-Abadie et al. , “Generative adversarial networks,” 2014. [Online]. Available: https://arxiv.org/abs/1406.2661

    2014 arXiv

  9. [17]

    Practical overview of image classification with tensor-network quantum circuits,

    D. Guala, S. Zhang et al. , “Practical overview of image classification with tensor-network quantum circuits,” Scientific Reports , vol. 13, no. 1, p. 4427, Mar 2023. [Online]. Available: https://doi.org/10.1038/ s41598-023-30258-y

    2023

  10. [18]

    Tensor networks for complex quantum systems,

    R. Or ´us, “Tensor networks for complex quantum systems,” Nature Reviews Physics , vol. 1, no. 9, pp. 538–550, Sep 2019. [Online]. Available: https://doi.org/10.1038/s42254-019-0086-7

    2019 doi

  11. [19]

    Compactifai: Extreme compression of large language models using quantum-inspired tensor networks,

    A. Tomut, S. S. Jahromi et al., “Compactifai: Extreme compression of large language models using quantum-inspired tensor networks,” 2024

    2024

  12. [20]

    Efficient mps representations and quantum circuits from the fourier modes of classical image data,

    B. Jobst, K. Shen et al. , “Efficient mps representations and quantum circuits from the fourier modes of classical image data,” 2023

    2023

  13. [21]

    The density-matrix renormalization group,

    U. Schollw ¨ock, “The density-matrix renormalization group,” Rev. Mod. Phys. , vol. 77, pp. 259–315, Apr 2005. [Online]. Available: https://link.aps.org/doi/10.1103/RevModPhys.77.259

    2005 doi

  14. [22]

    Unsupervised generative modeling using matrix product states,

    Z.-Y . Han, J. Wang et al. , “Unsupervised generative modeling using matrix product states,” Phys. Rev. X , vol. 8, p. 031012, Jul

  15. [23]

    Symmetric tensor networks for generative modeling and constrained combinatorial opti- mization,

    J. Lopez-Piqueres, J. Chen, and A. Perdomo-Ortiz, “Symmetric tensor networks for generative modeling and constrained combinatorial opti- mization,” 2023

    2023

  16. [24]

    Cons-training tensor networks,

    J. Lopez-Piqueres and J. Chen, “Cons-training tensor networks,” 2024. [Online]. Available: https://arxiv.org/abs/2405.09005

    2024 arXiv

  17. [25]

    Perfect sampling with unitary tensor networks,

    A. J. Ferris and G. Vidal, “Perfect sampling with unitary tensor networks,” Phys. Rev. B , vol. 85, p. 165146, Apr 2012. [Online]. Available: https://link.aps.org/doi/10.1103/PhysRevB.85.165146

    2012 doi

  18. [26]

    A practical introduction to tensor networks: Matrix product states and projected entangled pair states,

    R. Or ´us, “A practical introduction to tensor networks: Matrix product states and projected entangled pair states,” Annals of Physics, vol. 349, pp. 117–158, 2014. [Online]. Available: https: //www.sciencedirect.com/science/article/pii/S0003491614001596

    2014

  19. [27]

    Quantum inspired optimization for industrial scale problems,

    W. P. Banner, S. B. Hadiashar et al. , “Quantum inspired optimization for industrial scale problems,” 2023. [Online]. Available: https: //arxiv.org/abs/2305.02179

    2023 arXiv

  20. [28]

    Tensor network states and algorithms in the presence of a global u(1) symmetry,

    S. Singh, R. N. C. Pfeifer, and G. Vidal, “Tensor network states and algorithms in the presence of a global u(1) symmetry,” Phys. Rev. B , vol. 83, p. 115125, Mar 2011. [Online]. Available: https://link.aps.org/doi/10.1103/PhysRevB.83.115125

    2011 doi

  21. [29]

    Renormalization algorithms for quantum- many body systems in two and higher dimensions,

    F. Verstraete and J. I. Cirac, “Renormalization algorithms for quantum- many body systems in two and higher dimensions,” 2004. [Online]. Available: https://arxiv.org/abs/cond-mat/0407066

    2004 arXiv

  22. [30]

    Classical simulation of quantum many-body systems with a tree tensor network,

    Y .-Y . Shi, L.-M. Duan, and G. Vidal, “Classical simulation of quantum many-body systems with a tree tensor network,” Phys. Rev. A , vol. 74, p. 022320, Aug 2006. [Online]. Available: https://link.aps.org/doi/10.1103/PhysRevA.74.022320

    2006 doi

  23. [31]

    Deep generative modelling: A com- parative review of vaes, gans, normalizing flows, energy-based and autoregressive models,

    S. Bond-Taylor, A. Leach et al. , “Deep generative modelling: A com- parative review of vaes, gans, normalizing flows, energy-based and autoregressive models,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 11, pp. 7327–7347, 2022

    2022

  24. [32]

    Applications of negative dimensional tensors,

    R. Penrose, “Applications of negative dimensional tensors,” Combinato- rial Mathematics and its Applications , pp. 221–244, 1971

    1971

  25. [33]

    The density-matrix renormalization group in the age of matrix product states,

    U. Schollw ¨ock, “The density-matrix renormalization group in the age of matrix product states,” Annals of Physics , vol. 326, no. 1, pp. 96–192, 2011, january 2011 Special Issue. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0003491610001752

    2011

  26. [34]

    Tensor-train decomposition,

    I. V . Oseledets, “Tensor-train decomposition,” SIAM Journal on Scientific Computing , vol. 33, no. 5, pp. 2295–2317, 2011. [Online]. Available: https://doi.org/10.1137/090752286

    2011 doi

  27. [35]

    Supervised learning with tensor networks,

    E. Stoudenmire and D. J. Schwab, “Supervised learning with tensor networks,” Advances in neural information processing systems , vol. 29, 2016

    2016

  28. [36]

    Density matrix renormalization group algorithms with a single center site,

    S. R. White, “Density matrix renormalization group algorithms with a single center site,” Phys. Rev. B, vol. 72, p. 180403, Nov 2005. [Online]. Available: https://link.aps.org/doi/10.1103/PhysRevB.72.180403

    2005 doi

  29. [37]

    Density-matrix algorithms for quantum renormalization groups,

    S. R. White, “Density-matrix algorithms for quantum renormalization groups,” Phys. Rev. B , vol. 48, pp. 10 345–10 356, Oct 1993. [Online]. Available: https://link.aps.org/doi/10.1103/PhysRevB.48.10345

    1993 doi

  30. [38]

    Facts, conjectures, and improve- ments for simulated annealing,

    P. Salamon, R. Frost, and P. Sibani, “Facts, conjectures, and improve- ments for simulated annealing,” ser. Society for Industrial and Applied Mathematics, 2002., 1987

    2002

  31. [39]

    Tensor network estimation of distribution algorithms,

    J. Gardiner and J. Lopez-Piqueres, “Tensor network estimation of distribution algorithms,” 2024. [Online]. Available: https://arxiv.org/abs/ 2412.19780

    2024 arXiv

  32. [40]

    Gurobi Optimization, LLC, Gurobi Optimizer Reference Manual ,

  33. [41]

    Available: https://www.gurobi.com/documentation/ current/refman/index.html

    [Online]. Available: https://www.gurobi.com/documentation/ current/refman/index.html

  34. [42]

    Comb tensor networks,

    N. Chepiga and S. R. White, “Comb tensor networks,” Physical Review B , vol. 99, no. 23, Jun. 2019. [Online]. Available: http: //dx.doi.org/10.1103/PhysRevB.99.235426

    2019 doi

  35. [44]

    Supervised learning with quantum-inspired tensor networks,

    E. M. Stoudenmire and D. J. Schwab, “Supervised learning with quantum-inspired tensor networks,” 2017. [Online]. Available: https://arxiv.org/abs/1605.05775

    2017 arXiv

  36. [45]

    Generative learning of continuous data by tensor networks,

    A. Meiburg, J. Chen et al., “Generative learning of continuous data by tensor networks,” 2024. [Online]. Available: https://arxiv.org/abs/2310. 20498

    2024

  37. [2018]

    Available: https://link.aps.org/doi/10.1103/PhysRevX.8

    [Online]. Available: https://link.aps.org/doi/10.1103/PhysRevX.8. 031012

    doi

  38. [2024]

    Available: https://www.science.org/doi/abs/10.1126/ sciadv.adm6761

    [Online]. Available: https://www.science.org/doi/abs/10.1126/ sciadv.adm6761

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.