Pith. sign in

REVIEW 4 major objections 5 minor 83 references

A local tensor-network operator refines QTT-encoded functions to exponentially finer grids while keeping the added cores' rank bounded by the interpolation degree and the error independent of resolution.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 12:10 UTC pith:NZTFQQUW

load-bearing objection Classical TTI upsampling is solid and useful; the quantum circuit claim is unsupported and likely wrong as stated. the 4 major comments →

arxiv 2601.03885 v2 pith:NZTFQQUW submitted 2026-01-07 math.NA cs.GRcs.NAquant-ph

Efficient upsampling for tensor-network and quantum-state encoded functions

Siddhartha E. Guzman , Egor Tiunov , Leandro Aolita This is my paper
classification math.NA cs.GRcs.NAquant-ph MSC 65D0565D1515A69
keywords tensor trainsQTTinterpolationquasi-interpolationmatrix product operatorssuper-resolutionsynthetic turbulencefractal noise
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper introduces Tensor Train Interpolation (TTI), a way to upsample a function that is already stored as a quantized tensor train (QTT) without ever forming the full fine-grid tensor. Starting from an n-core QTT on a 2^n coarse grid, TTI produces an (n+m)-core QTT on a 2^{n+m} grid whose last m cores have rank bounded by the polynomial degree of the interpolation kernel. The method guarantees an ℓ2 error that does not grow with m, exponential compression at fixed accuracy, and runtime logarithmic in the number of grid points. This matters because standard ways of building fine QTT representations (e.g., TT-cross or TT-SVD) become costly or rank-inflated as resolution grows, while TTI keeps cost and rank essentially constant. The paper demonstrates the approach on functions, derivatives, airfoil masks, image super-resolution, and synthetic noise fields.

Core claim

The central object is the TTI operator, which expresses local kernel interpolation as a superposition over q neighboring coarse samples: shift operators S^(k) (rank-2 MPOs) act on the coarse cores and are tensor-multiplied by polynomial QTT encodings P^(k) whose rank is at most p+1 for degree-p kernels. This MPO-MPS operator, applied to a coarse QTT, produces the fine-scale representation directly. Consequently, the added m cores of the refined QTT have ranks bounded by p+1, independent of m; the interpolation error obeys an ℓ2 bound that does not depend on the number of added scales; and the QTT can be continued to arbitrarily many scales at fixed accuracy and essentially constant cost. In

What carries the argument

The TTI operator (Eq. 4): a sum over shifted coarse-sample selections S^(k) (rank-2 MPO shift operators) tensor-multiplied with QTT encodings of the interpolation polynomials P^(k), assembled into a single MPO-MPS operator whose last m cores are constant-rank vector cores. Its rank bounds come from two known building blocks: shift matrices written as rank-2 MPOs, and degree-p polynomials encoded in QTT with rank p+1 using an Abel-polynomial basis to avoid exponentially small entries. The QTT-Tucker encoding (a core TT plus per-dimension QTT factor matrices) isolates dimensions so that the added-core rank stays p+1 per dimension rather than (p+1)^d.

Load-bearing premise

The entire accuracy of the refined representation is inherited from the coarse QTT — TTI cannot repair a coarse encoding that missed the target function — and, on the quantum side, the claimed poly(n,m) circuit with log(p+1) ancillas is asserted without proof anywhere in the paper.

What would settle it

Compute the effective rank of the last m cores after repeated TTI refinement of a smooth function with a known coarse QTT at several grid spacings h; if any tail rank exceeds p+1, or if the ℓ2 error grows measurably with m for a fixed h, the central claims fail. For the quantum assertion, attempt to compile the TTI-O MPO into a circuit with at most log(p+1) ancillas; if the required ancillas scale with n or m, that claim collapses.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Once a function is in coarse QTT form, its fine-grid QTT at any resolution can be produced with essentially constant runtime and memory, rather than re-running TT-cross or TT-SVD.
  • Derivatives of the refined function are obtained analytically by differentiating the interpolation kernel, so derivative encodings carry the same constant-rank tail and error behaviour.
  • Soft-mask encodings of 2D/3D obstacles can be embedded in CFD-style tensor-network pipelines at exponentially fine grids without rank inflation, provided the mask is smoothed.
  • Synthetic noise fields (Perlin-style, midpoint displacement, fractal cascades) can be generated natively in QTT format with logarithmic complexity, including divergence-free 3D turbulence with a Kolmogorov spectrum.
  • QTT-Tucker is the favoured architecture in three or more dimensions, since it avoids the (p+1)^d rank blow-up of the interleaved encoding.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The quantum claim in the abstract — that the refined state can be prepared by a poly(n,m)-size circuit with log(p+1) ancillas — is not derived, sketched, or referenced anywhere in the body; a sympathetic reader should treat it as a conjecture. If a circuit can be found, TTI would give a concrete route to preparing amplitude-encoded functions at exponential resolution, useful for quantum algorithms
  • The method's usefulness for PDE solvers may go beyond warm starts: since TTI gives an explicit low-rank multiscale interpolant, one could couple it with tensor-network solvers to provide adaptive refinement of boundary conditions or solution fields, something the paper only gestures at.
  • A testable consequence: for a fixed kernel, the refined QTT's effective rank should stay flat as m grows even when the coarse encoding is built from different algorithms (TT-SVD, TT-cross); if rank climbs for a sufficiently smooth function, the p+1 bound is violated.
  • The same construction should extend to non-dyadic refinement factors (e.g., triadic grids) if a matching shift-and-polynomial decomposition exists; the dyadic choice is convenient but not essential to the mechanism.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes Tensor-Train Interpolation (TTI), a method that takes an n-core QTT (or QTT-interleaved/QTT-Tucker) representation of a function on a coarse dyadic grid and produces an (n+m)-core QTT on a fine grid by attaching constant-rank cores that encode a local polynomial kernel. The central claims are that the added m cores have ranks bounded by p+1 (polynomial degree), that the ℓ2 interpolation error is bounded independently of m, that the method runs in logarithmic (stated even as constant) time in the grid size, and that the same operation can be implemented in the quantum setting by a poly(n,m)-size circuit with log(p+1) ancillas. The classical TTI construction is validated on 1D functions and derivatives, 2D masks/images, and 3D synthetic turbulence, with comparisons to TT-Cross. The quantum circuit claim is stated in the abstract but never derived anywhere in the body or appendices.

Significance. If the classical part holds, the paper makes a useful contribution: it provides an explicit low-rank upsampling operator for QTTs with controlled ranks and error, avoids rerunning TT-Cross at every refinement level, and gives a practical scheme to generate fractal/noise fields directly in TT format with logarithmic memory. The explicit core construction (Eqs. (4)-(6), Appendices B and D) and the numerical benchmarks against TT-Cross are valuable. The quantum-circuit half of the abstract, however, is not supported by any derivation or reference, and as stated it is likely incorrect because the TTI map is not norm-preserving. The turbulence application is partly a construction (the Kolmogorov weights are chosen to enforce the spectrum) rather than an emergent validation. These issues do not undermine the classical TT interpolation machinery but do affect the advertised scope of the paper.

major comments (4)
  1. [Abstract, Sec. I, Sec. III] The abstract promises that the refined quantum state is prepared by a poly(n,m)-size circuit using log(p+1) ancillas, but no quantum circuit construction, unitary decomposition, or even a sketch appears in the body or the appendices. This is not a harmless omission: the TTI map K defined in Eq. (4) is not an isometry. For n=1, m=1 and linear interpolation, K^T K has eigenvalues 2 and 1, so ||K|ψ>|| depends on |ψ>. A deterministic unitary circuit on the state plus ancillas would induce a linear norm-preserving map on |ψ>, so the normalized refined state cannot be produced by a fixed unitary for all inputs. Any implementation must involve post-selection or amplitude amplification, and the paper provides no analysis of success probability or dependence on kernel condition number. The quantum claim should either be derived with the necessary non-unitary treatment or removed/qualified from th
  2. [Sec. IV.A (Fig. 2 caption), Sec. III, Sec. V] The statement that TTI runs in 'constant time regardless of the final number of scales' is misleading. The output is an (n+m)-core QTT, so writing out the m new cores requires Ω(m) storage and Ω(m) time if the cores are all distinct. The observed constant time in Fig. 2 is relative to n (the coarse core count) or because the numerical setup fixes m per experiment while varying the refinement level. The complexity should be stated as O(n+m) or, more precisely, O(n) plus O(m) for the tail cores, not O(1) in m. This also affects the 'logarithmic complexity' claims in the abstract; those claims should be made with respect to the appropriate parameters.
  3. [Sec. IV.B.1, Eq. (F1), Algorithm 5, Fig. 4] The Kolmogorov spectrum is built into the construction by choosing the cascade weights ω_m = 2^{-4m/3} in Eq. (F1) 'to enforce the Kolmogorov scaling.' Therefore Fig. 4(a) displaying E(k) ∝ k^{-5/3} is a consistency check of the chosen weights, not a predictive or emergent result. The text should be rephrased to state that the method can synthesize fields with a prescribed spectrum, rather than saying the framework 'reproduces' or 'confirms' the Kolmogorov law. Similarly, the 'intermittency' claim in Fig. 4(b) should be presented as a property of the constructed cascade, with a discussion of what aspect of the construction produces the flatness deviations, rather than as a universal property of TTI.
  4. [Sec. III, QTT-Tucker case] For QTT-T (QTT-Tucker), the paper states that applying TTI is 'trivial' because one can act on each dimension independently. The manuscript does not specify how the refinement interacts with the Tucker core and the factor matrices: refining dimension i requires updating both the core G and the factor matrix U_i, and the bond dimensions and rounding steps are not given. Since QTT-T is used in the main numerical experiments (Figs. 3 and 4), this missing algorithmic detail is load-bearing for reproducibility. A precise contraction/rounding algorithm for the QTT-T case is needed.
minor comments (5)
  1. [Abstract / Sec. II] The notation n, m, p is introduced in the abstract but the quantum part is never connected to qubit count or gate model. If the quantum claim is retained, all symbols should be defined and the circuit model specified.
  2. [References] Several references are duplicated: [21] and [41] are the same Oseledets SIAM paper; [12] and [61] appear to be the same Fournier et al. reference; [22] and [41] overlap. Please consolidate the bibliography.
  3. [Eq. (6)] The tensor-product identity notation in Eq. (6), 1⊗(m−1)_{rk−1} ⊗ G_k(c_{m,k}) ⊗ 1⊗(d−m)_{rk}, is ambiguous about the order of core indices and which ranks are being padded. A small example or caption would help.
  4. [Appendix D3] The text calls the B-spline coefficients c_i = f(x_i) a quasi-interpolant and states the approximation is O(h^2) independently of p. This is correct, but the preceding line in Appendix C says quasi-interpolants generally come from a pre-filtering step; the contrast should be made explicit to avoid confusion.
  5. [General] There are numerous typos and formatting issues (e.g., 'gird points', 'reescales', 'ilustrated', 'Teneva' capitalization, inconsistent math spacing in Eq. (5)). A careful proofreading pass is recommended.

Circularity Check

1 steps flagged

TTI interpolation and error bounds are self-contained; turbulence k^{-5/3} is enforced by construction and the abstract's quantum-circuit claim is unsupported.

specific steps
  1. fitted input called prediction [Sec. IV B 1, App. F (Eq. F1, Algorithm 5), Fig. 4a]
    "The interpolated field’s derivatives are rescaled accordingly so that the spectrum follows Kolmogorov’s law (see Algorithm 5). ... the weights ωm = 2−4m/3 enforce the Kolmogorov scaling. Algorithm 5: ω m ←(2 −4/3)m ▷cascade Kolmogorov weight. Fig. 4a caption: indicating ∝ k−5/3, confirming the expected Kolmogorov inertial-range scaling."

    The k^{-5/3} spectrum is an input to the construction, not an independent prediction. Each scale's amplitude is multiplied by ω_m = 2^{-4m/3} precisely so that the cascade obeys Kolmogorov scaling; reporting that the synthetic field's spectrum matches k^{-5/3} is therefore a consistency check of the chosen weights, not evidence for a physical or numerical claim. The circularity is confined to the turbulence demonstration; the TTI rank-bound and interpolation-error derivations do not rest on this choice.

full rationale

The central derivation is not circular. Eq. (4) defines TTI as a sum of shift-MPOs acting on the coarse QTT tensored with QTT-encoded polynomial kernels; rank bounds q+1 and p+1 and the m-independent ℓ2 error follow from standard local interpolation theory and the known polynomial QTT construction, and the method is benchmarked externally against TT-Cross (Teneva). No fitted constants enter the interpolation itself, and no uniqueness theorem or load-bearing self-citation is invoked. The only construction-driven circularity is the turbulence spectrum: the cascade weights are chosen to enforce k^{-5/3}, so Fig. 4a validates the input weights rather than predicting the law. Separately, the abstract's claim of a poly(n,m)-size quantum circuit with log(p+1) ancillas is asserted without derivation or reference in the body; since the TTI map is not norm-preserving, a unitary implementation would need post-selection or amplitude amplification that is not analyzed. This is a missing-proof/correctness risk, not a circularity of the classical derivation.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The central TT interpolation bound rests on standard rank bounds for shift MPOs and polynomial QTTs, plus the classical interpolation error theorem; these are acceptable. The paper introduces no new physical entities. The main ledger burden is the turbulent-cascade weights ω_m and the quantum circuit claim, both of which are imposed rather than derived.

free parameters (5)
  • Kernel degree p and support q = p=3, q=4 (Keys cubic / cubic B-spline); p=1 for linear cases
    Tail rank bound p+1 and error order O(h^{p+δ}) in App. C/D depend on this manual choice; different experiments use different kernels.
  • Kolmogorov cascade weights ω_m = ω_m = 2^{-4m/3}
    Eq. (F1) and Algorithm 5 set these weights explicitly 'to enforce the Kolmogorov scaling'; the k^{-5/3} spectrum in Fig. 4a is a consequence, not an independent prediction.
  • Persistence α for fractal noise = α = 1/2 (terrain/circle), 1/10 (1D landscape)
    App. E4/F define octave amplitude decay; chosen per example in Fig. G6/G7; affects statistics but not the interpolation error bound.
  • TT-rounding tolerance = 10^{-6} in Fig. G7; unspecified elsewhere
    Controls accuracy-vs-rank tradeoff after TTI contraction; not derived from theory.
  • Initial coarse cores n (grid spacing h) = n = 18 (1D), n ≈ 10–11 (masks), etc.
    The final error scales as O(h^{p+δ}); user must choose n so the coarse QTT is already accurate, since TTI cannot improve coarse accuracy.
axioms (6)
  • domain assumption Kernel interpolation error is O(γ(f,φ) h^{p+δ}) for functions of class C^{p+δ}
    App. C states this order; the scale-independent ℓ2 error claim rests on it.
  • standard math Shift matrices S, R, L have rank-2 MPO representations
    App. B2 restates [49]; used to build the TTI-O cores.
  • standard math Degree-p polynomials admit QTT cores of rank ≤ p+1 (binomial/Abel basis)
    App. B1 from [22]; ensures added cores have constant rank.
  • ad hoc to paper Non-periodic boundaries are handled by rank-1 delta-function boundary terms
    Sec. III states this without proof; used for masks/images where periodicity fails.
  • ad hoc to paper A poly(n,m)-size circuit with log(p+1) ancillas implements the quantum TTI
    Abstract asserts this; no construction, gate decomposition, or error analysis appears in the manuscript.
  • domain assumption Random QTT fields with unit variance, cubic B-spline smoothing, and weights ω_m yield the Kolmogorov spectrum
    App. F argues interpolation acts as low-pass filter and high-m terms are suppressed, but no rigorous spectral proof is given.

pith-pipeline@v1.3.0-alltime-deepseek · 25755 in / 16886 out tokens · 156158 ms · 2026-08-03T12:10:23.796510+00:00 · methodology

0 comments
read the original abstract

Both tensor trains (TTs) and quantum states provide compressed representations of grid-structured data with potentially exponential compression power. We present a unified framework for upsampling data encoded in vector amplitudes, with efficient realizations in both classical TT and quantum settings. Starting from an \(n\)-core TT or an \(n\)-qubit state on a coarse grid with \(2^n\) points, the construction produces an \((n+m)\)-core TT or \((n+m)\)-qubit state on a finer grid with \(2^{n+m}\) points. In the TT setting, it supports interpolation, quasi-interpolation, augmentation, and synthesis through efficient low-rank contractions, with the added \(m\) cores retaining constant rank. For function-value encodings, the resulting interpolation satisfies an \(\ell^2\)-error bound independent of the number of added grid points, achieves exponential compression at fixed accuracy, and has a logarithmic complexity in the number of grid points. In the quantum setting, the refined state is prepared by a \(\mathrm{poly}(n,m)\)-size circuit using \(\log(p+1)\) ancillas, where \(p\) controls the smoothness of the quasi-interpolant; the corresponding error scales quadratically with the initial grid spacing. We validate our framework for tensor networks in one-, two-, and three-dimensional examples, including functions, derivatives, airfoil masks, and synthetic random fields such as three-dimensional turbulence. In particular, fractal fields can be generated directly in TT format with logarithmic memory and runtime. These results open a practical route to multiscale solvers, generative models, and geometry-aware algorithms on tensor-network and quantum platforms, with potential applications in scientific simulation, imaging, and real-time graphics.

Figures

Figures reproduced from arXiv: 2601.03885 by Egor Tiunov, Leandro Aolita, Siddhartha E. Guzman.

Figure 1
Figure 1. Figure 1: FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 1
Figure 1. Figure 1: This two-level structure combines the sta [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

83 extracted references · 1 canonical work pages

  1. [1]

    Since the function is highly oscillatory, accurate interpola- tion requires at least two samples per period

    Functions and Derivatives We begin with a complex 1D function of classC 2 that exhibits varying oscillatory behavior across its domain; see Eq.(G1) for a formal definition. Since the function is highly oscillatory, accurate interpola- tion requires at least two samples per period. There- fore, we create a QTT of 18 cores usingTT-SVD and use TTI with a C 1...

  2. [2]

    3D masks The idea of encoding indicator functions as QTTs was first introduced in [27], where they incorpo- rated 2D objects into a completequantum-inspired pipeline for CFD. In [27] they show that TTcross performs better on softened indicator functions, al- lowing for the correct encoding (no artifacts) of non- slip boundary conditions around the boundar...

    2000

  3. [3]

    Although Per- lin’s noise is widely used in the literature (see App

    3D Noise: Turbulence As the main application of synthetic coherent noise, we developed a TN framework for creat- ing three-dimensional synthetic turbulence snap- shots for our two main architectures,QTT inter- leavedandQTT Tuckerencodings. Although Per- lin’s noise is widely used in the literature (see App. E3 and Fig. E2), we developed amultiscale additi...

  4. [4]

    A. A. S. a. B. D. Bojanov, H. A. Hakopian,Spline functions and multivariate interpolations, 1st ed., Mathematics and Its Applications : Main Series №248 (Springer, 1993)

    1993

  5. [5]

    Schafer and L

    R. Schafer and L. Rabiner, Proceedings of the IEEE 61, 692–702 (1973)

    1973

  6. [6]

    Wang, in2022 2nd International Conference on Computer Graphics, Image and Virtualization (IC- CGIV)(IEEE, 2022) p

    X. Wang, in2022 2nd International Conference on Computer Graphics, Image and Virtualization (IC- CGIV)(IEEE, 2022) p. 73–77

    2022

  7. [7]

    Lagae, S

    A. Lagae, S. Lefebvre, R. Cook, T. DeRose, G. Dret- takis, D. Ebert, J. Lewis, K. Perlin, and M. Zwicker, Computer Graphics Forum29, 2579–2600 (2010)

    2010

  8. [8]

    Perlin, ACM SIGGRAPH Computer Graphics 19, 287–296 (1985)

    K. Perlin, ACM SIGGRAPH Computer Graphics 19, 287–296 (1985)

    1985

  9. [9]

    G. E. Fasshauer, Advances in Computational Math- ematics11, 139–159 (1999)

    1999

  10. [10]

    Hagstrom and D

    T. Hagstrom and D. Appelö, Solving pdes with her- mite interpolation, inSpectral and High Order Meth- ods for Partial Differential Equations ICOSAHOM 2014(Springer International Publishing, 2015) p. 31–49

    2014

  11. [11]

    J. A. Parker, R. V. Kenyon, and D. E. Troxel, IEEE Transactions on Medical Imaging2, 31–39 (1983)

    1983

  12. [12]

    Keys, IEEE Transactions on Acoustics, Speech, and Signal Processing29, 1153–1160 (1981)

    R. Keys, IEEE Transactions on Acoustics, Speech, and Signal Processing29, 1153–1160 (1981)

    1981

  13. [14]

    Saupe, Algorithms for random fractals, inThe Science of Fractal Images(Springer New York,

    D. Saupe, Algorithms for random fractals, inThe Science of Fractal Images(Springer New York,

  14. [15]

    Fournier, D

    A. Fournier, D. Fussell, and L. Carpenter, Commu- nications of the ACM25, 371–384 (1982)

    1982

  15. [16]

    Perlin, ACM Transactions on Graphics21, 681–682 (2002)

    K. Perlin, ACM Transactions on Graphics21, 681–682 (2002)

    2002

  16. [17]

    R. L. Cook and T. DeRose, ACM Transactions on Graphics24, 803–811 (2005)

    2005

  17. [18]

    Leland, M

    W. Leland, M. Taqqu, W. Willinger, and D. Wilson, IEEE/ACM Transactions on Networking2, 1–15 (1994)

    1994

  18. [19]

    F. J. Molz, H. H. Liu, and J. Szulga, Water Re- sources Research33, 2273–2286 (1997)

    1997

  19. [20]

    B. B. Mandelbrot,The fractal geometry of nature (W.H. Freeman, New York, NY, 1982)

    1982

  20. [21]

    Jilesen, J

    J. Jilesen, J. Kuo, and F.-S. Lien, Computers & Geosciences46, 164–173 (2012)

    2012

  21. [22]

    J. M. Lilly, A. M. Sykulski, J. J. Early, and S. C. Olhede, Nonlinear Processes in Geophysics24, 481–514 (2017)

    2017

  22. [23]

    T. Kim, N. Thürey, D. James, and M. Gross, ACM Transactions on Graphics27, 1–6 (2008)

    2008

  23. [24]

    I. V. Oseledets, SIAM Journal on Scientific Com- puting33, 2295 (2011)

    2011

  24. [25]

    I. V. Oseledets, Constructive Approximation37, 1 (2013)

    2013

  25. [26]

    Ryzhakov and I

    G. Ryzhakov and I. Oseledets, Constructive tt- representation of the tensors given as index inter- action functions with applications (2022)

    2022

  26. [28]

    B. N. Khoromskij, Tensor numerical methods for high-dimensional pdes: Basic theory and initial ap- plications (2014), arXiv:1408.4053

    Pith/arXiv arXiv 2014

  27. [29]

    Kazeev and C

    V. Kazeev and C. Schwab, Numerische Mathematik 138, 133–190 (2017)

    2017

  28. [30]

    R. D. Peddinti, S. Pisoni, A. Marini, P. Lott, H. Ar- gentieri, E. Tiunov, and L. Aolita, Communications Physics7, 10.1038/s42005-024-01623-8 (2024)

    doi:10.1038/s42005-024-01623-8 2024

  29. [31]

    Markeeva, I

    L. Markeeva, I. Tsybulin, and I. Oseledets, QTT- isogeometric solver in two dimensions (2018), arXiv:1802.02839 [math.NA]

    Pith/arXiv arXiv 2018

  30. [32]

    M. Ali, A. Cortines, S. Morales, S. Mugel, M. Olave, R. Orus, S. Palmer, and H. Usabiaga, Quantum- inspired solver for simulating material deformations (2025), arXiv:2501.12151 [quant-ph]

    Pith/arXiv arXiv 2025

  31. [33]

    Gourianov, M

    N. Gourianov, M. Lubasch, S. Dolgov, Q. Y. van den Berg, H. Babaee, P. Givi, M. Kiffner, and D. Jaksch, Nature Computational Science2, 30–37 (2022)

    2022

  32. [34]

    Pisoni, R

    S. Pisoni, R. D. Peddinti, E. Tiunov, S. E. Guzman, and L. Aolita, Compression, simulation, and syn- thesis of turbulent flows with tensor trains (2025), arXiv:2506.05477

    arXiv 2025

  33. [35]

    J. J. García-Ripoll, Quantum5, 431 (2021)

    2021

  34. [36]

    Lindsey, Multiscale interpolative construction of quantized tensor trains (2024), arXiv:2311.12554 [math]

    M. Lindsey, Multiscale interpolative construction of quantized tensor trains (2024), arXiv:2311.12554 [math]

    Pith/arXiv arXiv 2024

  35. [37]

    J. J. Rodríguez-Aldavero, P. García-Molina, L. Tagliacozzo, and J. J. García-Ripoll, Chebyshev approximation and composition of functions in ma- trix product states for quantum-inspired numerical analysis (2024)

    2024

  36. [38]

    Lubasch, P

    M. Lubasch, P. Moinier, and D. Jaksch, Journal of Computational Physics372, 587 (2018)

    2018

  37. [39]

    Ali and M

    M. Ali and M. Kabel, Piecewise Polynomial Ten- sor Network Quantum Feature Encoding (2024), arXiv:2402.07671 [quant-ph] version: 4

    Pith/arXiv arXiv 2024

  38. [40]

    Oseledets and E

    I. Oseledets and E. Tyrtyshnikov, Linear Algebra and its Applications432, 70 (2010)

    2010

  39. [41]

    Savostyanov and I

    D. Savostyanov and I. Oseledets, inThe 2011 Inter- national Workshop on Multidimensional (nD) Sys- tems(IEEE, 2011) p. 1–8

    2011

  40. [42]

    Núñez Fernández, M

    Y. Núñez Fernández, M. K. Ritter, M. Jeannin, J.- W. Li, T. Kloss, T. Louvet, S. Terasaki, O. Par- collet, J. von Delft, H. Shinaoka, and X. Waintal, 9 SciPost Physics18, 10.21468/scipostphys.18.3.104 (2025)

    doi:10.21468/scipostphys.18.3.104 2025

  41. [43]

    Tindall, M

    J. Tindall, M. Stoudenmire, and R. Levy, Compress- ing multivariate functions with tree tensor networks (2024)

    2024

  42. [44]

    I. V. Oseledets, SIAM Journal on Scientific Com- puting33, 2295–2317 (2011)

    2011

  43. [45]

    B. N. Khoromskij, Constructive Approximation34, 257–280 (2011)

    2011

  44. [46]

    I. V. Oseledets, SIAM Journal on Matrix Analysis and Applications31, 2130–2145 (2010)

    2010

  45. [47]

    Ye and N

    E. Ye and N. F. Loureiro, Journal of Plasma Physics 90, 10.1017/s0022377824000503 (2024)

    doi:10.1017/s0022377824000503 2024

  46. [48]

    L. R. Tucker, Psychometrika31, 279–311 (1966)

    1966

  47. [49]

    De Lathauwer, B

    L. De Lathauwer, B. De Moor, and J. Vandewalle, SIAM Journal on Matrix Analysis and Applications 21, 1324–1342 (2000)

    2000

  48. [50]

    Thevenaz, T

    P. Thevenaz, T. Blu, and M. Unser, IEEE Transac- tions on Medical Imaging19, 739–758 (2000)

    2000

  49. [51]

    Stoer, R

    J. Stoer, R. Bartels, W. Gautschi, R. Bulirsch, and C. Witzgall,Introduction to Numerical Analysis, Texts in Applied Mathematics (Springer New York, 2002)

    2002

  50. [52]

    V. A. Kazeev, B. N. Khoromskij, and E. E. Tyr- tyshnikov, SIAM Journal on Scientific Computing 35, A1511–A1536 (2013)

    2013

  51. [53]

    L. B. Markeeva and I. V. Tsybulin, Computa- tional Mathematics and Mathematical Physics60, 2108–2115 (2020)

    2020

  52. [54]

    A. J. Ferris and G. Vidal, Phys. Rev. B Condens. Matter Mater. Phys.85(2012)

    2012

  53. [55]

    Chertkov, G

    A. Chertkov, G. Ryzhakov, and I. Oseledets, Teneva, https://github.com/AndreiChertkov/ teneva(2024)

    2024

  54. [56]

    Ion, torchtt, https://github.com/ion-g-ion/ torchTT(2026)

    G. Ion, torchtt, https://github.com/ion-g-ion/ torchTT(2026)

    2026

  55. [57]

    A. I. Boyko, M. P. Matrosov, I. V. Oseledets, D. Tsetserukou, and G. Ferrer, in2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)(IEEE, 2020) p. 10116–10121

    2020

  56. [58]

    Loeschcke, D

    S. Loeschcke, D. Wang, C. Leth-Espensen, S. Be- longie, M. J. Kastoryano, and S. Benaim, Coarse-to- fine tensor trains for compact visual representations (2024), arXiv:2406.04332

    Pith/arXiv arXiv 2024

  57. [59]

    Shi and A

    T. Shi and A. Townsend, On the compressibility of tensors (2018)

    2018

  58. [60]

    Dolgov and B

    S. Dolgov and B. Khoromskij, SIAM Journal on Ma- trix Analysis and Applications34, 593–623 (2013)

    2013

  59. [61]

    Chepiga and S

    N. Chepiga and S. R. White, Physical Review B99, 10.1103/physrevb.99.235426 (2019)

    work page doi:10.1103/physrevb.99.235426 2019

  60. [62]

    I. J. Schoenberg, Contributions to the problem of approximation of equidistant data by analytic functions, inI. J. Schoenberg Selected Papers (Birkhäuser Boston, 1988) p. 3–57

    1988

  61. [63]

    D. P. Mitchell and A. N. Netravali, ACM SIG- GRAPH Computer Graphics22, 221–228 (1988)

    1988

  62. [64]

    Fournier, D

    A. Fournier, D. Fussell, and L. Carpenter, Commu- nications of the ACM25, 371 (1982). 10 Appendix A: Tensor Train Representation In this Appendix, we introduce the Tensor Train (TT) format as a compact way to store and manipu- late high-dimensional arrays, and then explain how an uni-variate function/data sampled on a dyadic grid can be encoded into a Qua...

    1982

  63. [65]

    In TT form, we introduce core tensors Gk ∈ Rrk−1×nk×rk , k = 1 ,

    Tensor Train Representation Let A ∈R n1×n2×···×nd be a d-way tensor. In TT form, we introduce core tensors Gk ∈ Rrk−1×nk×rk , k = 1 , . . . , d,with r0 = rd = 1. Each entry ofA is then written as a product of ma- trix slices: Ai1...id =G 1(i1)G 2(i2)· · ·Gd(id),(A1) where Gk(ik) ∈R rk−1×rk is the matrix slice of thek- th core at indexik. The integersr1, ....

  64. [66]

    Let the points be uniform distributed xi = i 2N , i = 0, 1,

    Function Encoding in QTT For simplicity, let f : [0 , 1] →R be a function sampled on a uniform dyadic grid of size2 N, al- though the extension to any domain[a, b]is straight- forward. Let the points be uniform distributed xi = i 2N , i = 0, 1, . . . ,2N − 1, and let fi = f (xi). Each integer i admits a binary expansion i = a1 2 N−1 + a2 2 N−2 + · · ·+ aN...

  65. [67]

    Multivariate Encoding There are several ways to encode ad-dimensional 2N × · · · ×2N tensor into a QTT-like format: sequen- tial (all bits of each coordinate grouped), interleaved (bit-interleaving across dimensions), or more general tree-tensor networks (see [24] for different examples), like QTT-Tucker (see Fig. 1). While our scheme ap- plies to any, we...

  66. [68]

    Let A ∈ Rn1×n2×···×nd be a d-way tensor

    QTT-Tucker encoding Tucker decomposition [45, 46] provides a multi- linear generalization of the singular value decom- position (SVD) to higher-order tensors. Let A ∈ Rn1×n2×···×nd be a d-way tensor. The Tucker model represents A as a product of a smaller core tensor G and a collection of factor matrices Uk ∈R nk×rk , k = 1 , . . . , d,where each rk is th...

  67. [69]

    Reshaping {M (xi)}2N −1 i=0 into the N-way tensor Ma1···aN yields Ma1···aN = G1(a1) G2(a2) · · ·GN (aN )

    Polynomials Let M (x) = Pp k=0 ck xk be a degree-p polynomial, sampled at xi = i/2N with i = PN m=1 am2N−m, am ∈ {0, 1} and tm = am2−m. Reshaping {M (xi)}2N −1 i=0 into the N-way tensor Ma1···aN yields Ma1···aN = G1(a1) G2(a2) · · ·GN (aN ). The cores are G1(a1) = ϕ0(t1), . . . , ϕp(t1) , [Gk(ak)]i,j =    i i−j t i−j k , i≥j, 0, i < j, GN (aN ) = (1, t...

  68. [70]

    For k∈ 12 {0,

    Shift Matrices We now restate the results of [49], writing the explicit TT cores of the shift operators. For k∈ 12 {0, . . . ,2N − 1}, we define the shift matricesS(k) aa′, R(k) aa′,L (k) aa′ witha, a′ ∈ {0, . . . ,2N −1}, by S(k) aa′ =δ a,(a′−k) mod 2N , R(k) aa′ =1 {a=a ′+k<2N }, L(k) aa′ =1 {a ′=a+k<2N }. (B3) Where,1 {·} denotes the indicator function...

  69. [71]

    The first method has a convergence error rate ofO(h3), whereas the second has a rate ofO(h2), wherehis the grid size

    , see Appendix D), while figure (b) shows aC 2 B- spline quasi interpolant [59]. The first method has a convergence error rate ofO(h3), whereas the second has a rate ofO(h2), wherehis the grid size. generality we write the interpolation functionFas: F(x) = NX ℓ=1 cℓ φℓ(x),(C1) where {φℓ(x)} are basis (or kernel) functions and {cℓ}are coefficients chosen b...

  70. [72]

    , xi+⌈m/2⌉} and forming the unique degree-m polynomial that passes through those values

    Local Lagrange Interpolants A local Lagrange interpolant of degreem is con- structed by selecting, for each interval[xi, xi+1], the m + 1nearest data points {x i−⌊m/2⌋, . . . , xi+⌈m/2⌉} and forming the unique degree-m polynomial that passes through those values. Concretely, if{xk}i+m k=i are the chosen nodes with values {f (xk)}, then ℓ(m) k (x) = Q i+m ...

  71. [73]

    cardinal

    Natural Cubic Splines Natural cubic splines on an equispaced grid (hk = h = xk+1 −x k ∀k) satisfy F (xk) = f (xk), F ′′(x1) = F ′′(xN ) = 0, and the second-derivatives mk = F ′′(xk)solve the simplified tridiagonal system mk−1 + 4m k +m k+1 = 6 h2 f(x k+1)−2f(x k)+f(x k−1) , (D2) with k = 2, . . . , N−1, m 1 = mN = 0. On each interval[x k, xk+1]one then ha...

  72. [74]

    B–Spline Quasi Interpolants For smooth approximation one uses the normal- ized degree-p B-spline basis {B(p) k } on a uniform grid with spacing h. There are two equivalent construc- tions: Cox–de Boor recursion.Define the zeroth-degree basis by the box function χ(x), and then build higher-degree splines recursively: B(0) k (x) =χ(x−x k), B(p) k (x) = x−x ...

  73. [75]

    Cubic Convolution (Keys) Interpolation Cubic convolution as introduced by Keys provides a C 1 interpolant without solving a linear system for the coefficients as natural splines, see App. D2. We start with a convolution kernel interpolation F (x) = P k f (xk) φ x−xk h , where the Keys ker- nel is (see Fig. D1) φ(r) =    3 2 |r|3 − 5 2 |r|2 + 1,0≤ |r|...

  74. [76]

    Midpoint Displacement Algorithm The midpoint displacement algorithm is a recur- sive, one-dimensional method for generating fractal- like terrain profiles. Starting with two endpoints at positions x0 and xN with heights h0 and hN, the al- gorithm repeatedly inserts midpoints, setting each new height to the average of its two neighboring heights plus a ran...

  75. [77]

    For each segment endpoint pair at indicesi and i + d, compute the midpoint indexm = i + d/2

    At level ℓ, the segment length isd = N/2ℓ−1. For each segment endpoint pair at indicesi and i + d, compute the midpoint indexm = i + d/2

  76. [78]

    Set H[m] = 1 2 H[i]+ H[i+d] +rand [−R, R]

  77. [79]

    The final profile{H[0], H[1],

    After processing all segments at this level, up- dateR←R αand proceed to the next level. The final profile{H[0], H[1], . . . , H[N ]} exhibits sta- tistical self-similarity and natural roughness. Algorithm 11D Midpoint Displacement 1: functionMidpointDisplace- ment(N, h0, hN , R, α) 2: Allocate array H[0 . . . N], set H[0] ←h 0, H[N]←h N 3:forℓ= 1tokwhere...

  78. [80]

    For a real coordinate x, set i0 = ⌊x⌋ and t = x−i 0

    Cubic Noise Cubic noise generates a smooth random signal by assigning pseudo-random values at integer lattice points and interpolating between them via a general cubic kernelϕ(x), see Appendix D. For a real coordinate x, set i0 = ⌊x⌋ and t = x−i 0. Let V (i)be the pseudo-random value at integer lattice positionsi (e.g. obtained via a hash function). The n...

  79. [81]

    color- ing

    Perlin’s Noise Perlin’s noise is a procedural texture technique that generates smooth, pseudo-random patterns by interpolating gradient values on a regular lattice. Classical Perlin uses unit-length gradients, so that each gi (org i for multiple dimenisons) has|g| = 1, yielding Var(g) = 1and a flat gradient spectrum Sg(ω) = 1. Consequently, the noise spec...

  80. [82]

    Fractal Noise Fractal NoiseorFractal Brownian Motiontakes a smooth noise functionn(x)(see Appendices E3 and E2) as its base and generates a hierarchical su- perposition of rescaled and attenuated copies, called octaves. The resulting signal exhibits statistical self- similarity across scales and with the proper attenua- tion it approximates the power-law ...

Showing first 80 references.