Pith. sign in

REVIEW 2 major objections 3 minor 28 references

Entangling power of neural networks

T0 review · 2 major / 3 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 This paper proves that the entangling power of a degree-p polynomial decoder with latent width K is exactly the binomial coefficient binom(K+p,p), making exponential entanglement accessible with a linear number of latent variables.

desk verdict A clean, exact entangling-power formula for polynomial decoders, with the caveat that the sharp Bell-state width constant is proven only for unconstrained encoders. read the letter →

arxiv 2608.05293 v1 pith:6VPBWLC5 submitted 2026-08-05 cond-mat.dis-nn quant-ph

Taige Wang , Nisarga Paul , Liang Fu This is my paper · ORCID
classification cond-mat.dis-nnquant-ph
keywords entanglingpowerencoder-decoderneuralnetworkspolynomialdecodersSchmidtranklatentwidthBellstatesentanglemententropyquantumstaterepresentation
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

The paper introduces the entangling power of an encoder-decoder neural network: the maximum entanglement entropy (equivalently, Schmidt rank) that a network of given latent width and decoder complexity can generate between two halves of a system. For polynomial decoders of degree p, it proves the entangling power is exactly the number of monomials in K variables of total degree at most p, namely binom(K+p,p). This yields a sharp tradeoff: a linear decoder needs 2^n-1 latent variables for n Bell pairs, a degree-n decoder needs only about 0.2938n, and a degree at least 2^n-1 decoder needs just one. The result establishes that modest nonlinearity gives neural networks exponentially large entangling power, generalizing the Schmidt-rank notion of entanglement complexity.

What carries the argument

The central mechanism is the monomial evaluation map m(u)=(m_1(u),...,m_N(u)) from the latent space C^K to C^N, where the m_a are the monomials in K variables of total degree at most p. Its image spans C^N, so one can select encoder outputs whose evaluation vectors are linearly independent and solve the linear system that prescribes indicator polynomials q_x(u) with q_x($\varphi$(x'))=delta_{xx'}. The decoder is then built as a sum over x,y of f(x,y) q_x(u) q_y(v), which both achieves the bound and, via expansion into a sum of N product functions, establishes the Schmidt-rank upper bound.

What would settle it

The claim is disproved by finding a single function f on finite sets with Schmidt rank strictly greater than binom(K+p,p) that still has an exact representation f = g(phi_A(x), phi_B(y)) with a degree-p polynomial decoder and latent width K, using unrestricted encoders; for instance, with K=1 and p=2 the bound is 3, so a 4x4 matrix of rank 4 representable this way would falsify the theorem.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is a theorem with matching upper and lower bounds. For finite input sets X_A and X_B, any function f(x,y) represented as g(phi_A(x), phi_B(y)) with a degree-p polynomial decoder and K latent variables has Schmidt rank at most N_p(K)=binom(K+p,p); conversely, if both input sets have at least N_p(K) elements, every function admits such a representation. Hence the entangling power E_p(K), the maximum Schmidt rank obtainable, equals N_p(K). For n Bell pairs, where rank and dimension are both 2^n, the minimal latent width is exactly K_p(2^n), giving K_min = 2^n-1 for p=1, K_min ~ 0.2938n for p=n, and K_min = 1 for p >= 2^n-1. The paper also gives an explicit O(n)-unit, O(log n)-depth network implementing the degree-n Bell-state construction with quadratic activations.

Load-bearing premise

The theorem assumes the encoders can be completely arbitrary maps from input configurations to latent vectors; the construction needs points in the latent space whose monomial evaluation vectors are linearly independent, and nothing guarantees a low-complexity neural network can produce those particular points.

Editorial extensions

If this is right

  • For n Bell pairs, the minimum latent width is exactly K_p(2^n), so a degree-n decoder needs about 0.2938n latent variables, collapsing an exponential Schmidt rank into a linear resource.
  • The entangling-power formula gives a rigorous tradeoff: increasing decoder degree p lowers the latent width needed, with p = 2^n-1 reducing it to a single latent variable.
  • For p=1, E_1(K)=K+1, recovering the near-linear cost of bilinear (Schmidt) representations, with the +1 arising from the constant term.
  • The explicit construction via indicator polynomials provides a decoder in the class P_p for any target function once encoder points are chosen, so the bound is achievable in full generality.
  • Any maximally entangled state whose unitary matrix has a polynomial-size arithmetic-circuit representation admits a compact polynomial decoder, not just the identity Bell state.

Reading between the lines

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

  • The paper leaves open what happens if the encoders are restricted to affine maps or to networks of depth comparable to the decoder; the exact binomial bound may then fail, and the achievable entangling power would likely be governed by a different algebraic invariant.
  • A natural extension the authors do not pursue is the same counting argument for other decoder function classes, such as sparse polynomials or rational functions, where the relevant count of independent terms is no longer binomial.
  • The theorem reframes latent width as a measure of 'nonlinear rank' for classical bipartite functions, which could be a more appropriate complexity measure for neural network representations than matrix rank.
  • Because the Bell-state construction uses only a quadratic activation, it can be tested by direct simulation: build the O(n)-unit, O(log n)-depth network for small n and check that it reproduces the delta function exactly.
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 / 3 minor

Summary. The paper introduces an 'entangling power' E_p(K) for encoder-decoder representations of bipartite functions, where the decoder is a polynomial of total degree at most p in each latent vector and the latent space has dimension K. The central theorem states that, when the two input sets are large enough, the maximum Schmidt rank representable with K latent variables and degree-p polynomial decoders is exactly N_p(K) = binom(K+p, p). For the n-Bell-pair state, the authors derive the exact minimal latent width K_min = K_p(2^n), giving K_min ≈ 0.2938n for p=n, and K_min = 2^n−1 for p=1. The proof is based on a monomial expansion for the lower bound and on constructing indicator polynomials via a full-row-rank evaluation matrix for the upper bound. The paper also gives an explicit neural-network implementation of the Bell state with K=n using quadratic activations, and claims exponential entangling power of neural networks.

Significance. If the results stand, the paper gives a clean combinatorial characterization of the representational power of polynomial decoders, showing that nonlinear decoders can generate exponentially large Schmidt rank from a surprisingly small latent space. The main theorem is proven in a self-contained way: the lower bound follows directly from the monomial expansion, and the upper bound is an explicit interpolation construction with no fitted parameters. The asymptotic constants, including α_0 = 0.293815..., are derived in the Supplemental Material. The explicit Bell-state construction with K=n and O(n) quadratic-activation units is a concrete, checkable contribution. The notion of entangling power as e^{S_max} generalizes Schmidt rank in a natural way and may be useful for understanding neural quantum states, although the neural-network versus abstract-polynomial-decoder distinction needs care.

major comments (2)
  1. [Polynomial decoders / Neural-network implementation (Eq. (25), Eq. (10), Supplemental Theorem 1)] The exact minimal width K_min = K_p(2^n) ≈ 0.2938n for the Bell state is proven only for arbitrary encoders: the upper-bound construction chooses encoder points a_x in C^K whose monomial evaluation vectors are linearly independent, with no complexity bound or neural-network realizability. The only explicit neural-network implementation, Eq. (10), uses K=n, not K_min, and relies on a specially factored decoder. Therefore the exact constant 0.2938n is not yet established for neural networks with polynomially computable encoders; only K=n is. Since the title and abstract emphasize neural networks and the exact minimal width is a central quantitative claim, the authors should either provide an efficient encoder construction achieving K_min for the Bell state, or explicitly and prominently state that the exact constant applies to the abstract class of polynomial decoders with unrestricted encoders, while the neural-network result gives K=n.
  2. [Polynomial decoders, Theorem statement and proof (Eqs. (12)-(20))] The theorem as stated says E_p(K)=N_p(K) 'provided D_A,D_B ≥ N_p(K)' and then lists K_p(R) ≤ K_min ≤ K_p(D) as a consequence, but the upper-bound proof actually requires the reverse condition N_p(K) ≥ D (the text says 'when N ≥ D'). The proof is correct for the two separate statements: (i) the maximal Schmidt rank over representable functions is N_p(K) when the input sets have size at least N_p(K); and (ii) every function on sets of size D is representable when N_p(K) ≥ D. The theorem statement should separate these two cases, because as written the stated condition does not logically yield the claimed upper bound K_min ≤ K_p(D).
minor comments (3)
  1. [Supplemental Material, Eqs. (38)-(39)] The asymptotic threshold equation (1+α)log(1+α)−α log α = 1 is correct only if log denotes base-2 logarithm; with natural logarithms the right-hand side should be ln 2. Please specify the logarithm base explicitly, since the main text also uses log in entropy expressions.
  2. [Introduction, definition of entangling power (paragraph after Fig. 1 caption)] The notation E = e^{S_max} is nonstandard and may confuse readers who expect 'entangling power' to be a rate or an entropy. Please define the normalization convention clearly and explain why the exponential of the maximum entropy is the natural quantity.
  3. [Maximally entangled example, Eq. (7)] The sentence 'this generally requires exponentially large complexity in g, as shown by Eq. (7)' is not a formal complexity statement; Eq. (7) merely exhibits one decoder with D summands. It would be helpful to state that the decoder's arithmetic complexity is O(D) in that construction, without claiming a lower bound.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the entangling-power identity is proven from monomial counting and interpolation, with no fitted parameters and no load-bearing self-citations.

full rationale

The paper's central derivation is self-contained. The lower bound R <= N_p(K) follows by expanding a degree-p decoder in the monomial basis and observing that the representation is a sum of at most N_p(K) product functions; the upper bound is proven constructively by choosing encoder points whose monomial evaluation vectors are linearly independent and then building indicator polynomials via a right inverse of the resulting full-row-rank matrix. Corollary 1 attains the bound by applying the converse theorem to a flat-spectrum unitary matrix of rank N_p(K), not by assuming the target result. The constant 0.2938n and the scaling K_min ~ (p!)^{1/p} 2^{n/p} are derived from Stirling's approximation, not from fitting. The paper explicitly states that its general results hold for arbitrary encoders and gives a separate neural-network implementation only for the K=n Bell-pair construction; the gap between exact minimal width for arbitrary encoders and the explicit neural implementation is a scope/assumption limitation, not a circular step. Self-citations in the reference list are contextual and not load-bearing for the theorem. Therefore no significant circularity is present.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The derivation is fully self-contained. No free parameters are fitted; the only inputs are the definitions of the encoder-decoder form and the polynomial class P_p. The constant 0.2938 is derived, not fitted. No new physical entities are introduced.

assumptions (4)
  • standard math Monomials in K variables of total degree at most p are linearly independent as functions on C^K.
    Used to show the evaluation map's image spans C^N, needed for the full-rank interpolation matrix in the converse proof.
  • domain assumption The encoders may be arbitrary maps from finite sets to C^K, without continuity or architecture constraints.
    The converse proof constructs encoders by choosing arbitrary points a_x; this is stronger than typical neural network feature maps.
  • standard math For a normalized quantum state of Schmidt rank R, the von Neumann entropy S <= log R, with equality for a flat spectrum.
    Used in Corollary 1 to connect the rank bound to the entropy maximum.
  • standard math Stirling's approximation for the asymptotic threshold.
    Used to derive the 0.2938n constant in the p=n regime.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Entangling power of neural networks." pith.science (2026). https://pith.science/paper/6VPBWLC5

@misc{pith2026260805293,
  author       = {Pith},
  title        = {Pith review of: Entangling power of neural networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6VPBWLC5}},
  note         = {Machine review of arXiv:2608.05293}
}
abstract

Characterizing the complexity of correlations between subsystems is a fundamental task across information theory, machine learning, and science. In quantum physics, neural networks have found increasing application in learning wavefunctions. Here we introduce the entangling power of an encoder-decoder neural network, which quantifies its ability to generate entanglement between subsystems, dependent on a latent space dimension $K$ and the complexity class of the decoder. We exactly calculate this quantity for polynomial decoders of degree $p$ acting on a $K$-dimensional latent space. Our results establish the exponential entangling power of neural networks with modest resources. More broadly, our work provides a framework for analyzing correlations in machine learning that generalizes the notion of the Schmidt rank in entanglement theory.

Figures

Figures reproduced from arXiv: 2608.05293 by the authors.

Figure 1
Figure 1. FIG. 1. (Left) The general template for the neural encoder [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Polynomial sandwich bound ( [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 14 canonical work pages

  1. [1]

    C. E. Shannon, The Bell system technical journal27, 379 (1948)

    work page 1948

  2. [2]

    Eckart and G

    C. Eckart and G. Young, Psychometrika1, 211 (1936)

    work page 1936

  3. [3]

    ¨Ostlund and S

    S. ¨Ostlund and S. Rommer, Physical Review Letters75, 3537 (1995)

    1995

  4. [4]

    Vidal, Physical Review Letters91, 147902 (2003)

    G. Vidal, Physical Review Letters91, 147902 (2003)

    2003

  5. [5]

    M. B. Hastings, Journal of Statistical Mechanics: Theory and Experiment2007, P08024 (2007)

    2007

  6. [6]

    Schollw¨ ock, Annals of Physics326, 96 (2011)

    U. Schollw¨ ock, Annals of Physics326, 96 (2011)

    2011

  7. [7]

    Or´ us, Annals of Physics349, 117 (2014)

    R. Or´ us, Annals of Physics349, 117 (2014)

    2014

  8. [8]

    J. I. Cirac, D. P´ erez-Garc ´ ıa, N. Schuch, and F. Ver- straete, Reviews of Modern Physics93, 045003 (2021)

    work page 2021

Show all 28 references
  1. [9]

    Hornik, M

    K. Hornik, M. Stinchcombe, and H. White, Neural Net- works2, 359 (1989)

    1989

  2. [10]

    Cybenko, Mathematics of Control, Signals and Sys- tems2, 303 (1989)

    G. Cybenko, Mathematics of Control, Signals and Sys- tems2, 303 (1989)

    1989

  3. [11]

    Zaheer, S

    M. Zaheer, S. Kottur, S. Ravanbakhsh, B. P´ oczos, R. Salakhutdinov, and A. J. Smola, inAdvances in Neu- ral Information Processing Systems 30(2017) pp. 3391– 3401

    2017

  4. [12]

    Chen and J

    Z. Chen and J. Lu, Foundations of Computational Math- ematics , 1 (2026)

    2026

  5. [13]

    Fu, arXiv preprint arXiv:2510.11431 (2025)

    L. Fu, arXiv preprint arXiv:2510.11431 (2025)

    2025

  6. [14]

    Fu, Fermi sets: Universal and interpretable neural ar- chitectures for fermions (2026), arXiv:2601.02508 [quant- ph]

    L. Fu, Fermi sets: Universal and interpretable neural ar- chitectures for fermions (2026), arXiv:2601.02508 [quant- ph]

    2026 arXiv

  7. [15]

    D.-L. Deng, X. Li, and S. Das Sarma, Physical Review X 7, 021021 (2017)

    2017

  8. [16]

    Gao and L.-M

    X. Gao and L.-M. Duan, Nature communications8, 662 (2017)

    2017

  9. [17]

    Carleo, Y

    G. Carleo, Y. Nomura, and M. Imada, Nature communi- cations9, 5322 (2018)

    2018

  10. [18]

    J. Chen, S. Cheng, H. Xie, L. Wang, and T. Xiang, Phys- ical Review B97, 085104 (2018)

    2018

  11. [19]

    Levine, O

    Y. Levine, O. Sharir, N. Cohen, and A. Shashua, Physical review letters122, 065301 (2019)

    2019

  12. [20]

    Huang and J

    Y. Huang and J. E. Moore, Physical review letters127, 170601 (2021)

    2021

  13. [21]

    Szab´ o and C

    A. Szab´ o and C. Castelnovo, Physical Review Research 2, 033075 (2020)

    2020

  14. [22]

    T.-H. Yang, M. Soleimanifar, T. Bergamaschi, and J. Preskill, When can classical neural networks represent quantum states? (2024), arXiv:2410.23152 [quant-ph]

    2024 arXiv

  15. [23]

    D. S. Kufel, J. Kemp, D. Vu, S. M. Linsel, C. R. Lau- mann, and N. Y. Yao, Physical Review Letters135, 056702 (2025)

    2025

  16. [24]

    Sinibaldi, A

    A. Sinibaldi, A. F. Mello, M. Collura, and G. Carleo, Physical Review Research7, 043289 (2025)

    2025

  17. [25]

    Sharir, A

    O. Sharir, A. Shashua, and G. Carleo, Physical Review B106, 205136 (2022)

    2022

  18. [26]

    Paul, Physical Review Letters136, 120403 (2026)

    N. Paul, Physical Review Letters136, 120403 (2026)

    2026

  19. [27]

    Wang, Expressibility of neural quantum states: A walsh-complexity perspective (2026), arXiv:2604.03294 [quant-ph]

    T. Wang, Expressibility of neural quantum states: A walsh-complexity perspective (2026), arXiv:2604.03294 [quant-ph]

    2026 arXiv

  20. [28]

    Y. Lu, S. Bharadwaj, D. Rathore, and D. Luo, arXiv preprint arXiv:2603.23468 (2026). 6 Supplemental Material: Entangling power of neural networks Here, we present the theorem which was discussed in the main text in self-contained and complete form, along with corollaries. Thro...

    2026

Pith tools

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