Pith. sign in

REVIEW 5 major objections 5 minor 53 references

Restricted Boltzmann machine as a probabilistic Enigma

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper proposes that a restricted Boltzmann machine with permuted visible biases can act as a symmetric cipher with factorial key space, AES-comparable diffusion, and decryption by probabilistic sampling.

desk verdict A novel but internally inconsistent RBM-based cipher: key space and sampling cost scale in incompatible ways, so the core security claim fails. read the letter →

arxiv 2507.17236 v1 pith:VGWZHQEC submitted 2025-07-23 cond-mat.stat-mech physics.app-phphysics.data-an

Bin Chen , Weichao Yu This is my paper
classification cond-mat.stat-mechphysics.app-phphysics.data-an
keywords restrictedBoltzmannmachineprobabilisticcomputingsymmetricencryptionmarginaldistributionpermutationkey#P-completenesspost-quantumcryptographydiffusion
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 proposes a symmetric encryption scheme in which a restricted Boltzmann machine plays the role of a physical 'Enigma': the plaintext is written into the marginal probability distribution over the machine's visible states, and the secret key is a permutation of the visible biases. The authors argue that this construction gives a factorial key space that grows as $n!\times m!$, diffusion that at modest permutation ratios already outperforms AES by their cross-correlation measure, and decryption hardness tied to #P-complete partition-function evaluation (counting problems at least as hard as NP). A legitimate receiver who knows the permutation can restore the original biases and recover the message either by exact evaluation or by sampling the machine's equilibrium distribution, which is naturally fast on emerging probabilistic-computing hardware. The broader claim is that this creates an asymmetry: legitimate decryption scales gently with sample count while an eavesdropper faces exponential cost, and the scheme does not rely on integer factorization, so known quantum factoring attacks do not apply.

What carries the argument

The engine of the scheme is the trained marginal distribution $$P(v_k)=\frac{2^m}{Z}\left(\prod_{i=1}^m \$\cosh$\left(\sum_j W_{ij}v_k^j+\beta_i\right)\right)\exp\left(\sum_j \alpha_j v_k^j\right),$$ combined with the key operation $\alpha'=K\alpha$, where $K$ is an $n\times n$ permutation matrix. The marginal formula is what carries the message: the normalized bit string becomes a probability landscape over the $2^n$ visible configurations, and permuting $\alpha$ reshapes that landscape while leaving $W$ and $\beta$ intact. Decoding is the inverse operation: $K^{-1}$ restores $\alpha$, and the bits are recovered by comparing each recovered probability with half the uniform value. The security wall is the partition function $Z$ (equivalently the normalization of the marginal), whose evaluation is #P-complete, so an attacker who has only the ciphertext parameters $W$, $\alpha'$, and $\beta$ cannot feasibly compute the correct landscape, while a sampler can estimate it with error $1/\sqrt{N}$.

What would settle it

Replicate the protocol at a secure size, for example $n=37$ visible nodes (where $n!$ exceeds $2^{128}$), on probabilistic hardware and measure the time to reach the paper's sampling threshold $N>2^{2(n+1)}$; if that time exceeds the message-expiration window, the central asymmetry between cheap legitimate decryption and expensive adversary decryption is not realized.

Watch

Extended reading notes

Core claim

The central claim is that a restricted Boltzmann machine with permuted visible biases is a usable symmetric cipher, not just a metaphor. The sender converts a plaintext bit string into a target marginal distribution $P_t(v_k)$, trains the RBM weights and biases so that the model's marginal $P_\theta(v_k)\simeq P_t(v_k)$, and ships the weight matrix $W$, the permuted visible biases $\alpha'=K\alpha$, and the hidden biases $\beta$ as ciphertext, with $K$ a secret permutation matrix. The receiver applies $K^{-1}$ and thresholds the recovered marginal at half the uniform value to reconstruct the bits. The paper reports three supporting results: the key space is $n!\times m!$; the cross-correlation between true and wrong-key pseudo-plaintext falls below AES once about 20% of the biases are permuted; and the partition function entering the marginal distribution is #P-complete to evaluate, so direct brute-force decryption costs $O(2^n m n)$ while sampling-based decryption costs $O(N)$ samples with error $1/\sqrt{N}$.

Load-bearing premise

The scheme stands or falls with the assumption that a single visible-layer size $n$ can be large enough for the factorial key space to defeat brute force and still small enough for the legitimate receiver to compute or sample all $2^n$ marginal probabilities accurately in feasible time.

Editorial extensions

If this is right

  • If the factorial key-space claim holds, roughly 37 visible nodes already give $37!\times m! > 2^{128}$ key choices, so the scheme can in principle reach conventional security levels with modest visible-layer sizes.
  • If the diffusion result holds across larger systems, a key need only permute a fraction of biases (about 20% in the demonstrated case) to push plaintext-to-pseudo-plaintext correlation below AES, making key application cheap.
  • Because security does not rest on integer factorization or discrete logarithms, the usual quantum factoring attack does not apply; the authors note that a generic quantum search speedup is countered by doubling key length.
  • On probabilistic hardware with sampling rates near $10^{11}$ samples per second, the receiver can approach the paper's threshold $N>2^{2(n+1)}$ quickly, making sampling-based decryption the practical path.

Reading between the lines

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

  • One testable extension is to measure how the diffusion metric scales beyond the demonstrated $n=7$; if the cross-correlation advantage over AES persists for larger $n$ and longer plaintexts, the cipher can be evaluated as a block-cipher candidate rather than a proof of principle.
  • A natural variant would reuse a single trained RBM as a keystream generator, deriving a fresh bias permutation per message block, which would turn the factorial key space into a stream-cipher mode without retraining each block.
  • Because the ciphertext contains the full weight matrix, a known-plaintext attack that tries to recover $K$ from matched plaintext-ciphertext pairs would be a natural next test of the security claim.
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

5 major / 5 minor

Summary. This Letter proposes a symmetric encryption scheme based on Restricted Boltzmann Machines (RBMs). The plaintext is encoded into the marginal distribution of the visible layer, the key is a permutation matrix applied to the visible biases, and the ciphertext consists of the trained weights and partially permuted biases. The authors claim factorial key-space growth, diffusion comparable to AES, efficient decryption on probabilistic hardware, and an adversarial barrier based on #P-completeness. The numerical work demonstrates decryption and cross-correlation metrics for a small n=7 example. I find that the central claim of simultaneous factorial key space and efficient decryption is contradicted by the paper's own sampling threshold and by the actual definition of the key space.

Significance. If the central claims held, the work would be noteworthy: it proposes a concrete physical cipher for emerging p-bit hardware and a post-quantum direction that does not rely on factoring or discrete logarithms. The paper also contains a clear protocol and small-scale numerical demonstrations, which are useful for probing the idea. However, the significance depends entirely on the claimed asymmetric computational barrier. That barrier is not established: the key space is n!, not m!×n!, the sampling criterion N>2^{2(n+1)} makes the secure regime computationally infeasible, and #P-completeness does not prevent an adversary from using the same sampling procedure as the legitimate receiver. As a result, the protocol as stated does not achieve both security and efficiency in any parameter regime.

major comments (5)
  1. [Encryption and decryption / Fig. 2(a)] The key space is n!, not m!×n!. In the encryption step only the visible biases α are permuted by the matrix K; the weights W and hidden biases β are transmitted as part of the ciphertext without permutation. Therefore the number of distinct keys is the number of permutations of the n visible bias entries, n!, and the m! factor used later in the 'Optimization under limited resources' section is not part of the protocol. For the demonstrated n=7 example this gives 7! = 5,040 keys, not 7!×m! keys, and exhaustive search is trivial.
  2. [Physical implementation on probabilistic computers] The sampling threshold stated in this section, N > 2^{2(n+1)}, is incompatible with the claimed security. To obtain a key space of at least 2^{128} from n! one needs n ≥ 37. At n=37 the criterion requires N > 2^{76} samples; at the paper's cited p-bit rate of 10^{11} samples/s this is roughly 24,000 years per block, and the plaintext block is already 2^{37} bits. Conversely, the demonstrated n=7 regime requires only N > 2^{16} samples but has only 5,040 keys. Thus no parameter regime simultaneously provides a secure key space and efficient sampling-based decryption under the paper's own equations.
  3. [Physical implementation on probabilistic computers / Security discussion] The claimed asymmetry between legitimate users and adversaries is not supported by #P-completeness. The legitimate receiver decrypts by sampling from the RBM steady-state distribution, and an adversary holding the ciphertext can run exactly the same sampling procedure for each candidate permutation. #P-hardness of exact partition-function evaluation does not prevent efficient approximate sampling; indeed, the paper relies on that fact for the receiver's efficiency. The only barrier for the adversary is therefore key enumeration over the n! permutations, which reduces the security analysis to the key-space size and does not create the advertised computational asymmetry.
  4. [Encryption and decryption / Eq. (6)] Equation (6) defines the target marginal probability as Pt(v_k) = B_k / sum_i B_i, which assigns probability zero to every configuration corresponding to a zero plaintext bit. However, the RBM marginal distribution in Eq. (4) is strictly positive for every visible configuration, since it contains only exponentials and cosh factors. The Kullback-Leibler divergence in Eq. (5) is therefore infinite whenever any B_k = 0, so the stated training objective is not well-defined for arbitrary binary plaintexts. The paper does not introduce any smoothing or minimum-probability floor to address this.
  5. [Cross-correlation metric / Optimization under limited resources] The diffusion and resource-allocation claims rely on quantities that are not anchored to standard cryptographic criteria. The cross-correlation C in Eq. (8) is a custom metric, and the comparison with AES in Fig. 3 does not establish the claimed 'diffusion comparable to AES' without standard avalanche or statistical-test analyses. Likewise, the Information-Security Balance Index (ISBI) in Eq. (9) is introduced ad hoc; optimizing it produces n=5, m=95, which would imply a key space of at most 5! = 120 and a plaintext block of only 32 bits, directly contradicting the security requirements stated elsewhere in the paper.
minor comments (5)
  1. [Abstract / Physical implementation] 'sharp P-complete' should be '#P-complete' in the abstract and in the Physical implementation section; the '#' symbol is missing or corrupted.
  2. [Throughout] There are several typographical errors, including 'dervied', 'permutaion', 'depcited', and 'approahces' in the text and Supplemental Material URL 'hppt://link.aps.org'.
  3. [Encryption and decryption / Eq. (6)] The normalization in Eq. (6) is undefined for an all-zero plaintext, and the text does not discuss how the sender handles binary strings with very few nonzero bits, which create extremely sparse target distributions.
  4. [Fig. 1 and Fig. 4] Figure 1(f) labels a specific '43rd visible configuration' without explaining the ordering convention, and Fig. 4 does not label the time axes or sample-count axes clearly enough to reproduce the quantitative claims about decryption time.
  5. [Model / Information capacity] The phrase 'exponential information capacity scaling (2 n)' is ambiguous; it should be written as 2^n to avoid confusion with the notation for the visible-layer size n.

Circularity Check

2 steps flagged · score 3.0 of 10

Two self-referential steps (ISBI optimization and wrong-key security argument) appear in the security/design narrative; the core encryption/decryption construction remains self-consistent and the #P-completeness support is external.

  1. self definitional ['Optimization under limited resources' section, Eq. (9) and Fig. 3(d)]
    "We further define the Information-Security Balance Index (ISBI) ISBI = n log10((mn+m+n)/(2^n)), where the prefactor n represents the information entropy of n nodes ... Our analysis reveals that the optimal resource allocation that maximizes the ISBI occurs at an asymmetric distribution of n=5 visible nodes and m=95 hidden nodes."

    The 'optimal' number of visible nodes is the maximizer of a metric the authors introduce ad hoc in the same section. No external cryptographic or information-theoretic criterion fixes the functional form of ISBI; changing the prefactor or the argument of the logarithm changes the maximizer. The conclusion 'n=5 is optimal' therefore restates the definition of ISBI rather than deriving an independent security or capacity result. This is a self-referential optimization, though it does not feed back into the core encryption/decryption equations.

  2. self definitional ['Encryption and decryption' section, paragraph after Eq. (7) (Fig. 2(c) discussion)]
    "Since KK′ ̸= I (for K′ ̸= K−1), it follows that α′′ ̸= α. The pseudo-plaintext dervied from this operation results in corrupted information that bears no meaningful resemblance to the original message. This shows that without the correct key, an eavesdropper cannot successfully recover the original plaintext."

    The security conclusion ('cannot successfully recover the original plaintext') is asserted as the meaning of 'corrupted information that bears no meaningful resemblance.' The only fact actually derived is α'' ≠ α, which holds for any permutation matrix and does not distinguish a secure cipher from a trivial keyed permutation. No adversary model, distinguishing experiment, or complexity argument is supplied in this step; the claim reduces by construction to the definition of a permutation key.

full rationale

The encryption/decryption loop is self-consistent by construction: the plaintext is encoded into Pt(vk) by Eq. (6), the RBM is trained to approximate that distribution, and Eq. (7) inverts it; this is the design of a cipher, not a fitted quantity relabeled as a prediction. The two genuinely circular steps are the ISBI optimization and the wrong-key security claim: the first presents the maximizer of an ad hoc index as an 'optimal' design result, and the second proves only that a wrong permutation gives a different output and then equates that with an adversary's inability to recover the plaintext. Neither step is needed for the basic mechanics, but both are used in the paper's security and design narrative. The #P-completeness argument (citations [40-42]) and the p-bit sampling rate ([49]) are external support and do not reduce to the paper's own definitions. The more serious issue in the manuscript is an internal inconsistency, not circularity: if the key is restricted to n×n visible-bias permutations, the key space is n! (the m! factor in the claimed key space is unsupported because hidden biases β are transmitted unpermuted), so n ≥ 37 is needed for n! ≥ 2^128; at that size the paper's own sampling criterion N > 2^{2(n+1)} demands more than 2^76 samples, contradicting the claimed efficient legitimate decryption. That, together with the weak wrong-key argument, is a correctness and feasibility concern rather than a circular derivation of the core construction.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

The central claim depends on three free parameters or ad hoc choices, five background assumptions, and one invented metric. The most consequential is the implicit assumption that large n is simultaneously feasible and secure; the paper's own sampling bound contradicts it.

free parameters (3)
  • KL divergence threshold = 0.005
    Training stops when D_KL falls below 0.005; chosen by hand, no analysis of how training error affects decoding accuracy.
  • Decoding threshold = (2 * 2^n)^(-1) = 2^(-(n+1))
    Eq. (7) classifies a bit as 1 if P(v_k) exceeds half of 2^(-n); this threshold is arbitrary and drives the required sample count N > 2^(2(n+1)).
  • ISBI weighting = n=5 for n+m=100
    The Information-Security Balance Index (Eq. 9) is introduced ad hoc and then maximized to select the optimal allocation n=5, m=95; no independent security or information-theoretic justification.
assumptions (5)
  • standard math The RBM marginal distribution formula Eq. (4) is correct as a mathematical identity.
    It is a standard derivation summing over hidden units and is not in question.
  • standard math Partition function evaluation for RBMs is #P-complete.
    Cited from Refs. [40-42]; used to argue an adversary's exact-decryption cost is exponential. This is a known result, but it does not account for the small n regime used in the paper.
  • domain assumption Training with contrastive divergence can drive D_KL below 0.005 for the target distributions used.
    The paper states the optimization ends when D_KL < 0.005 but provides no convergence guarantee or representational analysis; the numerical evidence is only for one plaintext.
  • domain assumption Probabilistic computers can sample at 10^11 samples per second, making sampling-based decryption fast.
    Quoted from Ref. [49]; even if true, the required sample count N > 2^(2(n+1)) grows exponentially with n, so the speed advantage does not fix the scaling problem.
  • ad hoc to paper An adversary with the ciphertext {W, alpha', beta} cannot recover plaintext without trying permutations of alpha.
    The security argument assumes brute-force over n! keys is the best attack and that no structural attack uses the known W, beta, and the fact that alpha' is a permutation of alpha. No proof is given.
invented entities (1)
  • Information-Security Balance Index (ISBI)
    purpose: A metric to select the optimal number of visible nodes n given a fixed total node count; used to claim n=5 is optimal.
    The metric is defined in Eq. (9) by hand, with no derivation from cryptographic or information-theoretic first principles, and its maximization is presented as a result.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Restricted Boltzmann machine as a probabilistic Enigma." pith.science (2026). https://pith.science/paper/VGWZHQEC

@misc{pith2026250717236,
  author       = {Pith},
  title        = {Pith review of: Restricted Boltzmann machine as a probabilistic Enigma},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VGWZHQEC}},
  note         = {Machine review of arXiv:2507.17236}
}
read the original abstract

We theoretically propose a symmetric encryption scheme based on Restricted Boltzmann Machines that functions as a probabilistic Enigma device, encoding information in the marginal distributions of visible states while utilizing bias permutations as cryptographic keys. Theoretical analysis reveals significant advantages including factorial key space growth through permutation matrices, excellent diffusion properties, and computational complexity rooted in sharp P-complete problems that resist quantum attacks. Compatible with emerging probabilistic computing hardware, the scheme establishes an asymmetric computational barrier where legitimate users decrypt efficiently while adversaries face exponential costs. This framework unlocks probabilistic computers' potential for cryptographic systems, offering an emerging encryption paradigm between classical and quantum regimes for post-quantum security.

Figures

Figures reproduced from arXiv: 2507.17236 by the authors.

Figure 1
Figure 1. FIG. 1. (a) Schematic representation of an RBM architecture [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. (a) Encryption: The sender encodes plaintext into binary using ASCII and embeds it into the visible layer’s marginal [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. The dependence of cross correlation [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: FIG. 4. (a) Decryption accuracy (characterized by cross [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 48 canonical work pages

  1. [1]

    M. M. Alani, in Proceedings of the 3rd International Con- ference on cryptography, security and privacy(2019) pp. 23–27

    work page 2019

  2. [2]

    J¨ aschke and F

    A. J¨ aschke and F. Armknecht, in International confer- ence on selected areas in cryptography(Springer, 2018) pp. 453–478

    work page 2018

  3. [3]

    Andonov, J

    S. Andonov, J. Dobreva, L. Lumburovska, S. Pavlov, V. Dimitrova, and A. Popovska-Mitrovikj, in Proc. ICT Innov. (2020) pp. 1–11

    work page 2020

  4. [4]

    M. M. Alani, in International Conference on Neural In- formation Processing(Springer, 2012) pp. 637–646

    work page 2012

  5. [5]

    Hospodar, B

    G. Hospodar, B. Gierlichs, E. De Mulder, I. Ver- bauwhede, and J. Vandewalle, Journal of Cryptographic Engineering 1, 293 (2011)

    work page 2011

  6. [6]

    C. E. Shannon, The Bell system technical journal 28, 656 (1949)

    work page 1949

  7. [7]

    Katz and Y

    J. Katz and Y. Lindell, Introduction to modern cryptogra- phy: principles and protocols(Chapman and hall/CRC, 2007)

    work page 2007

  8. [8]

    R. L. Rivest, A. Shamir, and L. Adleman, Communica- tions of the ACM 21, 120 (1978)

    work page 1978

Show all 53 references
  1. [9]

    P. Kaye, R. Laflamme, and M. Mosca, An introduction to quantum computing(OUP Oxford, 2006)

    2006

  2. [10]

    P. W. Shor, in Proceedings 35th annual symposium on foundations of computer science (Ieee, 1994) pp. 124– 134

    1994

  3. [11]

    D. J. Bernstein and T. Lange, Nature 549, 188 (2017)

    2017

  4. [12]

    Rijmen and J

    V. Rijmen and J. Daemen, Proceedings of federal infor- mation processing standards publications, national insti- tute of standards and technology 19, 1 (2001)

    2001

  5. [13]

    Horowitz and E

    M. Horowitz and E. Grumbling, Quantum computing: progress and prospects(National Academies Press, Wash- ington, D.C., 2019)

    2019

  6. [14]

    D. H. Ackley, G. E. Hinton, and T. J. Sejnowski, Cogni- tive science 9, 147 (1985)

    1985

  7. [15]

    Kirkpatrick, C

    S. Kirkpatrick, C. D. Gelatt Jr, and M. P. Vecchi, science 220, 671 (1983)

    1983

  8. [16]

    Fischer and C

    A. Fischer and C. Igel, in Progress in Pattern Recog- nition, Image Analysis, Computer Vision, and Appli- cations: 17th Iberoamerican Congress, CIARP 2012, Buenos Aires, Argentina, September 3-6, 2012. Proceed- ings 17 (Springer, 2012) pp. 14–36

    2012

  9. [17]

    G. E. Hinton and R. R. Salakhutdinov, science 313, 504 (2006)

    2006

  10. [18]

    X. Yan, J. Ma, T. Wu, A. Zhang, J. Wu, M. Chin, Z. Zhang, M. Dubey, W. Wu, M. S.-W. Chen, et al., Nature Communications 12, 5710 (2021)

    2021

  11. [19]

    K. S. Woo, J. Kim, J. Han, W. Kim, Y. H. Jang, and C. S. Hwang, Nature Communications 13, 5762 (2022)

    2022

  12. [20]

    Niazi, S

    S. Niazi, S. Chowdhury, N. A. Aadit, M. Mohseni, Y. Qin, and K. Y. Camsari, Nature Electronics , 1 (2024)

    2024

  13. [21]

    Patel, P

    S. Patel, P. Canoza, and S. Salahuddin, Nature Electron- ics 5, 92 (2022)

    2022

  14. [22]

    W. A. Borders, A. Z. Pervaiz, S. Fukami, K. Y. Camsari, H. Ohno, and S. Datta, Nature 573, 390 (2019)

    2019

  15. [23]

    N. S. Singh, K. Kobayashi, Q. Cao, K. Selcuk, T. Hu, S. Niazi, N. A. Aadit, S. Kanai, H. Ohno, S. Fukami, et al., Nature Communications 15, 2685 (2024)

    2024

  16. [24]

    Y. Wang, B. Chen, W. Gao, B. Ye, C. Niu, W. Wang, Y. Zhu, W. Yu, H. Guo, and J. Shen, National Science Review , nwae338 (2024)

    2024

  17. [25]

    X. Li, C. Wan, R. Zhang, M. Zhao, S. Xiong, D. Kong, X. Luo, B. He, S. Liu, J. Xia, et al., Nano Letters 24, 5420 (2024)

    2024

  18. [26]

    B¨ ohm, D

    F. B¨ ohm, D. Alonso-Urquijo, G. Verschaffelt, and G. Van der Sande, Nature Communications 13, 5847 (2022)

    2022

  19. [27]

    R. P. Feynman, International Journal of Theoretical Physics 21, 467 (1982)

    1982

  20. [28]

    G. E. Hinton, in Neural Networks: Tricks of the Trade: Second Edition (Springer, 2012) pp. 599–619

    2012

  21. [29]

    M. H. Amin, E. Andriyash, J. Rolfe, B. Kulchytskyy, and R. Melko, Physical Review X 8, 021050 (2018)

    2018

  22. [30]

    See Supplemental Material at hppt://link.aps.org/ for derivation of the marginal distribution for visible layer viarables in RBM, discussion and comparison on differ- ent encoding approahces, derivation of the updating rules for the parameters of RBM, and additional details ab...

  23. [31]

    Le Roux and Y

    N. Le Roux and Y. Bengio, Neural computation 20, 1631 (2008)

    2008

  24. [32]

    Fernandez-de Cossio-Diaz, S

    J. Fernandez-de Cossio-Diaz, S. Cocco, and R. Monasson, Physical Review X 13, 021003 (2023)

    2023

  25. [33]

    Tubiana and R

    J. Tubiana and R. Monasson, Physical review letters118, 138301 (2017)

    2017

  26. [34]

    A. J. Menezes, P. C. Van Oorschot, and S. A. Vanstone, Handbook of applied cryptography(CRC press, 2018)

    2018

  27. [35]

    Kullback, Information theory and statistics (Courier Corporation, 1997)

    S. Kullback, Information theory and statistics (Courier Corporation, 1997)

    1997

  28. [36]

    J. R. Hershey and P. A. Olsen, in 2007 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing-ICASSP’07, Vol. 4 (IEEE, 2007) pp. IV–317

    2007

  29. [37]

    Little, Communications Society 10, 7 (1973)

    J. Little, Communications Society 10, 7 (1973)

    1973

  30. [38]

    Chang, K

    J. Chang, K. Lu, Y. Guo, Y. Wang, Z. Zhao, L. Huang, H. Zhou, Y. Wang, F. Lei, and B. Zhang, CCF Transac- tions on High Performance Computing 6, 243 (2024)

    2024

  31. [39]

    R. M. Gray, Entropy and information theory (Springer Science & Business Media, 2011)

    2011

  32. [40]

    Regts, Combinatorica 38, 987 (2018)

    G. Regts, Combinatorica 38, 987 (2018)

    2018

  33. [41]

    L. G. Valiant, siam Journal on Computing 8, 410 (1979)

    1979

  34. [42]

    J.-G. Liu, L. Wang, and P. Zhang, Physical Review Let- ters 126, 090506 (2021)

    2021

  35. [43]

    Parks, M

    B. Parks, M. Bapna, J. Igbokwe, H. Almasi, W. Wang, and S. A. Majetich, AIP Advances 8 (2018)

    2018

  36. [44]

    Kaiser, A

    J. Kaiser, A. Rustagi, K. Y. Camsari, J. Z. Sun, S. Datta, and P. Upadhyaya, Physical Review Applied 12, 054056 (2019)

    2019

  37. [45]

    K. Y. Camsari, S. Chowdhury, and S. Datta, Physical Review Applied 12, 034061 (2019)

    2019

  38. [46]

    K. Y. Camsari, B. M. Sutton, and S. Datta, Applied Physics Reviews 6 (2019)

    2019

  39. [47]

    K. Y. Camsari, R. Faria, B. M. Sutton, and S. Datta, Physical Review X 7, 031014 (2017). 7

    2017

  40. [48]

    Grimaldi, L

    A. Grimaldi, L. S´ anchez-Tejerina, N. Anjum Aadit, S. Chiappini, M. Carpentieri, K. Camsari, and G. Finoc- chio, Physical Review Applied 17, 024052 (2022)

    2022

  41. [49]

    Chowdhury, A

    S. Chowdhury, A. Grimaldi, N. A. Aadit, S. Niazi, M. Mohseni, S. Kanai, H. Ohno, S. Fukami, L. Theoga- rajan, G. Finocchio, et al., IEEE Journal on Exploratory Solid-State Computational Devices and Circuits 9, 1 (2023)

    2023

  42. [50]

    R. Zand, K. Y. Camsari, S. D. Pyle, I. Ahmed, C. H. Kim, and R. F. DeMara, in Proceedings of the 2018 on Great Lakes Symposium on VLSI(2018) pp. 15–20

    2018

  43. [51]

    Vodenicarevic, N

    D. Vodenicarevic, N. Locatelli, A. Mizrahi, J. S. Fried- man, A. F. Vincent, M. Romera, A. Fukushima, K. Yakushiji, H. Kubota, S. Yuasa, et al., Physical Re- view Applied 8, 054045 (2017)

    2017

  44. [52]

    Kaplan, G

    M. Kaplan, G. Leurent, A. Leverrier, and M. Naya- Plasencia, arXiv preprint arXiv:1510.05836 (2015)

    2015 arXiv

  45. [53]

    E. T. Campbell, B. M. Terhal, and C. Vuillot, Nature 549, 172 (2017)

    2017

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.