Pith. sign in

REVIEW 3 major objections 5 minor 54 references

Theoretical Analysis of Thermodynamic Matrix Inversion: First-order Equivalence to Preconditioned Gradient Descent and Implications for Analog Computing

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read The paper proves that discretized thermodynamic matrix inversion is, to first order in the step size, identical to preconditioned gradient descent, so thermal noise is algorithmically redundant for symmetric positive definite matrices.

desk verdict A clean but low-novelty re-derivation of Richardson iteration as the first-order covariance evolution of a discretized OU process, whose redundancy claim outruns the math. read the letter →

arxiv 2608.09743 v1 pith:6YH3H4AX submitted 2026-08-10 math.NA cs.NA

Gerhard Kirsten , Michael Selby , Janith Petangoda , Olle Halqvist Elias , James Meech , Phillip Stanley-Marbell This is my paper · ORCID
classification math.NAcs.NA MSC 65F1065F3060H1065C30
keywords thermodynamiccomputingmatrixinversionOrnstein-UhlenbeckprocesspreconditionedgradientdescentsymmetricpositivedefinitematricesEuler-MaruyamadiscretizationanalogNewton-Schulziteration
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

Thermodynamic computing solves linear algebra by letting a physical system of coupled oscillators equilibrate under thermal noise and reading the answer from the statistics of its fluctuations. This paper asks what that stochastic process is really computing, and answers: for symmetric positive definite matrices, the same thing as a plain deterministic optimization. Tracking the covariance matrix of the discretized Ornstein-Uhlenbeck process, the authors prove that to first order in the step size the covariance update is exactly preconditioned gradient descent on the Frobenius-norm residual $\|\Sigma A - I\|_F$. The corollary is that thermal fluctuations, though physically essential for driving the analog hardware, are algorithmically redundant: a deterministic update reproduces the stochastic dynamics without any random sampling. If this is right, it converts a physics-inspired hardware program into a recognizable optimization algorithm, giving hardware designers a step-size rule and convergence rate and giving digital implementers a sampling-free route to the same approximation.

What carries the argument

The load-bearing object is the covariance evolution of the discretized Ornstein-Uhlenbeck process, defined by the Euler-Maruyama update $x_{i+1} = (I - \Delta t A)x_i + \sqrt{2\Delta t}\,Z_i$ with independent standard normal increments $Z_i$. Because the process is Gaussian, its full statistical content is carried by the covariance matrix $\Sigma_i$, which satisfies the closed deterministic recursion $\Sigma_{i+1} = (I - \Delta t A)\Sigma_i(I - \Delta t A) + 2\Delta t\,I$ of Lemma 1. The argument then hinges on two algebraic steps: the commutativity of $A$ with every $\Sigma_i$ (Lemma 2, guaranteed when $\Sigma_0 = \alpha I$), which lets the two factors $(I - \Delta t A)\Sigma_i(I - \Delta t A)$ be combined, and a first-order Taylor expansion in $\Delta t$ that discards the $O(\Delta t^2)$ term and leaves the gradient-descent update of Theorem 1. This machinery carries the argument from the SDE world of thermal equilibration to the optimization world of a quadratic loss, a preconditioner, and a learning rate, without ever solving the high-dimensional Fokker-Planck equation.

What would settle it

Iterate the deterministic update $\Sigma_{i+1} = \Sigma_i - 2\Delta t(\Sigma_i A - I)$ with a finite step $\Delta t$ inside the claimed stable range $0 < \Delta t < 1/\lambda_{\max}$, and compare the trajectory at every time point with the exact Ornstein-Uhlenbeck covariance $\Sigma_t = A^{-1}(I - e^{-2At})$ sampled at the same instants, across matrices with widely separated eigenvalues. If the gap between the two trajectories grows with $\Delta t^2$ or with iteration count in the regime where an analog circuit would operate, then the neglected noise terms carry algorithmic weight, and the redundancy claim fails.

Watch

Extended reading notes

Core claim

The central discovery is an identity, stated as Theorem 1. Under the Euler-Maruyama discretization of the Ornstein-Uhlenbeck process $dX_t = -AX_t\,dt + \sqrt{2}\,dW_t$, the covariance matrix $\Sigma_i = \langle x_i \otimes x_i \rangle$ evolves as $\Sigma_{i+1} = (I - \Delta t A)\Sigma_i(I - \Delta t A) + 2\Delta t\,I$; when $\Sigma_0 = \alpha I$ commutes with $A$, expanding to first order in $\Delta t$ yields $\Sigma_{i+1} = \Sigma_i - 2\Delta t(\Sigma_i A - I) + O(\Delta t^2)$. This is precisely one step of preconditioned gradient descent on the loss $L(\Sigma) = \|\Sigma A - I\|_F^2$ with preconditioner $A$ and learning rate $\Delta t$. The paper concludes that the stochastic thermal machinery is algorithmically redundant for convex problems with a single global minimum: the deterministic update captures the first-order computational content of the analog machine, and simulating it digitally requires no random samples. Convergence of the iteration is governed by the spectral-radius condition $\rho(I - 2\Delta t A) < 1$, equivalent to $0 < \Delta t < 1/\lambda_{\max}(A)$, with optimal step $\Delta t^{*} = 1/(\lambda_{\max} + \lambda_{\min})$ and contraction rate $(\kappa(A) - 1)/(\kappa(A) + 1)$ — the same spectral-radius thread the paper traces back to the 1950 Monte Carlo method for matrix inversion.

Load-bearing premise

The claim stands or falls on the first-order truncation being a faithful picture of the real device: the equivalence is proven only up to terms of order the time step squared, and the paper does not bound how large those neglected terms are against the exact continuous-time covariance evolution that physical analog hardware actually undergoes.

Editorial extensions

If this is right

  • Digital implementations of thermodynamic matrix inversion need no random number generation: the deterministic update reproduces the convergence behavior of the stochastic process without sampling.
  • The thermodynamic time constant of a physical implementation should be matched to $\Delta t^{*} = 1/(\lambda_{\max} + \lambda_{\min})$, and hardware coupling and reservoir parameters map onto the preconditioner $A$, giving designers a convergence-theoretic rule for choosing equilibration times.
  • The derived algorithm sits in the same efficiency class as classical iterative inversion: it matches the Newton-Schulz iteration on well-conditioned matrices and, accelerated by Schur complements, becomes competitive or faster on large block-structured problems.
  • The spectral-radius condition that governed the 1950 Monte Carlo approach to matrix inversion is the same condition governing this iteration, so the mathematics connecting those two historical methods and the thermodynamic hardware forms one continuous thread.
  • The redundancy of thermal noise is specific to convex quadratic potentials with a single global minimum; for non-convex landscapes such as combinatorial optimization or multimodal sampling, the paper expects thermal fluctuations to remain algorithmically essential.

Reading between the lines

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

  • If the first-order identity holds at the step sizes real analog devices use, the honest benchmark for thermodynamic matrix-inversion hardware is the deterministic gradient-descent iteration, not stochastic simulation; any energy or speed advantage would have to be argued against that classical algorithm.
  • The same moment-closure move — tracking covariance instead of the full probability density — should transfer to any analog computation with a quadratic potential, so the derivation is a template for extracting the algorithmic core of other physics-based linear-algebra proposals; a direct test is whether the equivalence survives a non-isotropic diffusion matrix.
  • A falsifiable design prediction follows: a physical implementation whose convergence rate per effective time constant deviates from $(\kappa(A) - 1)/(\kappa(A) + 1)$ is carrying algorithmic content in its finite-$\Delta t$ or device-specific noise that the first-order equivalence misses.
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

3 major / 5 minor

Summary. The paper analyzes the Ornstein-Uhlenbeck process underlying thermodynamic SPD matrix inversion. It derives the exact Euler-Maruyama covariance recurrence, then shows that a first-order expansion in the step size coincides with preconditioned gradient descent on the Frobenius residual ||ΣA−I||²_F. The authors prove convergence conditions for the resulting iteration, identify an optimal step size, compare the deterministic iteration with the Thermox stochastic simulator and with the Newton-Schulz iteration, and propose a Schur-complement acceleration. The stated headline conclusion is that thermal fluctuations are algorithmically redundant for convex problems with a single global minimum.

Significance. If made fully rigorous, the connection is a useful theoretical bridge: the continuous-time covariance of a linear Ornstein-Uhlenbeck process evolves deterministically, and the paper's identification of this evolution with preconditioned gradient descent gives a clean optimization-theoretic interpretation of thermodynamic matrix inversion. The proofs of Lemmas 1–2 and Theorems 2–3 are elementary and correct, and the proposal of a concrete optimal time constant for hardware is a falsifiable design prediction. The historical link to the von Neumann–Ulam method is a nice contextual addition. However, the discrete-time formulation suffers from an unquantified O(Δt²) truncation that changes the fixed point of the exact Euler-Maruyama recurrence, and the numerical validation relies on an oracle learning rate and a loose tolerance. The paper's central insight is salvageable, but the current statement and validation do not support the strongest claims in the abstract.

major comments (3)
  1. [Numerical Experiments] Theorem 1 and Eq. (8): the O(Δt²) remainder is not benign over the iteration horizon. Starting from Lemma 1, Eq. (2), on an eigenmode of A with eigenvalue λ the exact Euler-Maruyama covariance recurrence has fixed point σ_∞ = 1/(λ − Δtλ²/2), whose relative error with respect to 1/λ is Δtλ/(2 − Δtλ). At the paper's optimal step Δt* = 1/(λ_max + λ_min), this equals κ/(κ+2) for λ = λ_max, i.e., about 96% error at κ=50. Hence the discretized stochastic process whose covariance is tracked in Lemma 1 does not converge to A⁻¹; only the truncated update (8) does. The statement that the discretized Ornstein-Uhlenbeck process produces iterative approximations to A⁻¹ via preconditioned gradient descent is therefore not supported without a uniform-in-time estimate of the accumulated O(Δt²) remainder. The exact continuous-time OU covariance ODE from Proposition 1 would give dΣ/dt = −2(ΣA − I) without any discretization bias; the paper would be better served by proving that exact equivalence and treating (8) as its forward-Euler discretization.
  2. [Numerical Experiments] Table 3 and the surrounding text: the validation against Thermox does not test the first-order truncation claimed in Theorem 1. The deterministic algorithm is run with the oracle step Δt* to a loose tolerance ε = 10⁻², while Thermox is run with a fixed sample count of 10⁶ and no reported integration time T; the resulting Thermox errors (for example, 0.154 for n=100, κ=2) are much larger than the GD tolerance. The speedup of 10⁵–10⁶ is therefore not evidence that the deterministic update faithfully reproduces the thermodynamic covariance evolution; it compares a five-iteration deterministic moment computation with a heavily sampled stochastic estimator. The sentence in the paper stating that this 'confirms the theoretical prediction of Theorem 1' is not logically supported by the experimental design.
  3. [Abstract] The headline claim that thermal fluctuations are algorithmically redundant is stated without the qualifications needed by the analysis. The redundancy is shown, at most, for the first-order truncation of one Euler-Maruyama discretization; the exact discrete recurrence (2) has a biased fixed point, while the exact continuous-time hardware follows the deterministic ODE only in the moment sense. The authors should either prove a quantitative bound on the O(Δt²) error over the full iteration count, or reformulate the claim as the exact continuous-time moment ODE being equivalent to gradient flow, with (8) as its forward-Euler discretization. As written, the abstract overstates what Theorem 1 establishes.
minor comments (5)
  1. [Abstract] The citation 'Aiferet al.' is missing a space in the abstract and in the introduction; it should read 'Aifer et al.'.
  2. [Numerical Experiments] The text says 'See Figure 5 for an algorithmic illustration,' but Figure 5 shows the validation against Thermox, not an algorithmic illustration; the reference should likely be to Figure 4.
  3. [Table 3] The column header 'Iterations/Samples' and the narrative do not specify the integration horizon T used for Thermox, even though Thermox's accuracy depends on both the number of samples and the final time; please clarify the protocol.
  4. [Abstract] The abstract's phrase 'remaining competitive with the Newton-Schulz iteration' is not supported by Table 4 for ill-conditioned problems: for κ=50, 100, and 500, Newton-Schulz requires far fewer iterations and is several times faster. The claim should be restricted to the well-conditioned, small-dimension regime.
  5. [Theorem 1] Theorem 1 assumes x0 = 0 and hence Σ0 = 0, while the pseudocode in Figure 4 initializes Σ0 = αI with α > 0; the notation should be unified or the difference explicitly justified.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the OU-to-gradient-descent equivalence is derived, not assumed, and external validation anchors it.

full rationale

The paper's central derivation is self-contained and non-circular. Theorem 1 does not assume its conclusion: Lemma 1 derives the exact covariance recurrence Sigma_{i+1} = (I - Delta t A) Sigma_i (I - Delta t A) + 2 Delta t I directly from the Euler-Maruyama update (16) by expanding the outer product and using independence of x_i and Z_i. Theorem 1 then expands this recurrence to first order in Delta t and algebraically identifies the resulting update Sigma_i - 2 Delta t (Sigma_i A - I) with the gradient-descent update for L(Sigma) = ||Sigma A - I||_F^2 with preconditioner P = A: grad L = 2(Sigma A - I)A, so P^{-1} grad L = 2(Sigma A - I). This identification is a rewriting of the derived expansion, not an input. The optimal step size Delta t* is derived from spectral-radius analysis (Theorem 3) rather than fitted to the target data, and the paper explicitly labels it a controlled experimental condition rather than a practical recommendation. The numerical validation compares against Thermox, an external exact-OU simulator, and against Newton-Schulz, an established classical iteration, so the algorithmic claim is not self-confirming. The self-citations in the paper (TTR references [47,48], the ADC bottleneck reference [30], and a general unconventional-computing reference [29]) supply background methods or motivation and are not load-bearing in the proof of Theorem 1. The main caveat that the equivalence is only first-order in Delta t and no uniform bound on the discarded O(Delta t^2) term over the iteration horizon is provided is an error-analysis/correctness limitation, not a circularity, because the paper repeatedly and explicitly states the 'to first-order' scope rather than smuggling the conclusion in as a premise.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central derivation uses no fitted parameters: the covariance recursion is derived analytically and the convergence conditions follow from spectral analysis. The only hand-tuned inputs are the learning rate (set to the oracle optimal value using eigenvalues) and the Schur block size. The paper relies on standard OU theory, the Euler-Maruyama discretization, the SPD assumption, and the Thermox simulator as ground truth.

free parameters (2)
  • learning rate Δt = 1/(λ_max + λ_min) in experiments
    Chosen from the theoretical optimum using eigenvalues of A; in practice this requires full eigendecomposition, an O(n^3) oracle. The theory leaves Δt as a free scalar to be tuned.
  • Schur recursion cutoff τ = 4, 8, 16, 32
    Minimum block size for the recursive Schur complement algorithm; selected a priori by the authors in the acceleration experiments.
assumptions (4)
  • standard math Standard properties of the Ornstein-Uhlenbeck process: explicit solution, mean and covariance evolution (Proposition 1).
    Used in Theorem 4 to show stationary covariance equals A^{-1}. Standard results in stochastic calculus, cited to [45, 51].
  • domain assumption Euler-Maruyama discretization faithfully represents the continuous-time OU process for the chosen Δt.
    Used in Definition 3 and Lemma 1. The paper's equivalence is only first-order in Δt, and the truncation error is not bounded.
  • domain assumption A is symmetric positive definite.
    The problem class is fixed to SPD matrices, as stated throughout; the method does not extend to general matrices.
  • domain assumption Thermox is a faithful digital proxy for ideal thermodynamic hardware.
    Used in the validation experiments. The paper takes the simulator's OU sampling as ground truth for the physical dynamics, but does not independently verify it against physical hardware.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Theoretical Analysis of Thermodynamic Matrix Inversion: First-order Equivalence to Preconditioned Gradient Descent and Implications for Analog Computing." pith.science (2026). https://pith.science/paper/6YH3H4AX

@misc{pith2026260809743,
  author       = {Pith},
  title        = {Pith review of: Theoretical Analysis of Thermodynamic Matrix Inversion: First-order Equivalence to Preconditioned Gradient Descent and Implications for Analog Computing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6YH3H4AX}},
  note         = {Machine review of arXiv:2608.09743}
}
abstract

Recent research has demonstrated the possibility of exploiting the thermodynamics of coupled electrical oscillators to implement computational tasks such as matrix inversion. While physical implementations rely on thermal noise to drive equilibration, we show that the underlying dynamics reduce to a deterministic iterative algorithm. Building on the framework of Aifer et al., we analyze the moment evolution of the Ornstein-Uhlenbeck process governing thermodynamic symmetric positive definite (SPD) matrix inversion. We prove that to a first-order approximation, the covariance dynamics are mathematically identical to preconditioned gradient descent on the Frobenius norm of the residual $\tilde{A}^{-1}A-I$. This equivalence demonstrates that thermal fluctuations, while essential for physical thermodynamic hardware, are algorithmically redundant for convex problems with a single global minimum. We validate the resulting algorithm against Thermox (Duffield et al.), a stochastic thermodynamic simulator, achieving speedups exceeding 100,000-fold while remaining competitive with the Newton-Schulz iteration. We also demonstrate acceleration through Schur complement techniques. These results establish a rigorous link between analog thermodynamic computing, statistical physics, and deterministic optimization methods.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 52 canonical work pages

  1. [1]

    & Hsu, J

    Leonard, T. & Hsu, J. S. J. Bayesian inference for a covariance matrix. The Annals of Statistics20, 1669–1696 (1992)

    work page 1992

  2. [2]

    Matrix inversion

    Franklin, J. Matrix inversion. InComputational Methods for Physics, 220–245 (Cambridge University Press, 2013)

    work page 2013

  3. [3]

    & Blau, L

    Cahill, P. & Blau, L. M. The preliminary application of a matrix inversion method for radionuclide imaging. Journal of Nuclear Medicine11, 613–615 (1970)

    work page 1970

  4. [4]

    Algebraic Methods and Computational Strategies for Pseudoinverse-Based MR Image Reconstruction (Pinv-Recon)

    Yeung, K.et al. Pinv-Recon: Generalized MR image reconstruction via pseudoinversion of the encoding matrix.arXiv preprint arXiv:2410.06129 (2024).2410.06129

    work page Pith review arXiv 2024

  5. [5]

    Portfolio selection.The Journal of Finance7, 77–91 (1952)

    Markowitz, H. Portfolio selection.The Journal of Finance7, 77–91 (1952)

    work page 1952

  6. [6]

    Daley, R.Atmospheric Data Analysis (Cambridge University Press, Cambridge, UK, 1991)

    work page 1991

  7. [7]

    Bannister, R. N. A review of operational methods of variational and ensemble-variational data assimilation. Quarterly Journal of the Royal Meteorological Society143, 607–633 (2017)

    work page 2017

  8. [8]

    Iterative Methods for Sparse Linear Systems (SIAM, Philadelphia, 2003), 2nd edn

    Saad, Y. Iterative Methods for Sparse Linear Systems (SIAM, Philadelphia, 2003), 2nd edn

    work page 2003

Show all 54 references
  1. [9]

    Golub, G. H. & Van Loan, C. F.Matrix Computations (Johns Hopkins University Press, Baltimore, 2013), 4th edn

    2013

  2. [10]

    Trefethen, L. N. & Bau, D.Numerical Linear Algebra (SIAM, Philadelphia, 1997)

    1997

  3. [11]

    H., Teukolsky, S

    Press, W. H., Teukolsky, S. A., Vetterling, W. T. & Flannery, B. P. Numerical Recipes: The Art of Scientific Computing (Cambridge University Press, 2007), 3rd edn

    2007

  4. [12]

    Watkins, D. S. Fundamentals of Matrix Computations (John Wiley & Sons, 2010), 3rd edn

    2010

  5. [13]

    Mahalanobis, P. C. On the generalised distance in statistics. Proceedings of the National Institute of Sciences of India2, 49–55 (1936)

    1936

  6. [14]

    & Wolf, M

    Ledoit, O. & Wolf, M. A well-conditioned estimator for large-dimensional covariance ma- trices. Journal of Multivariate Analysis88, 365–411 (2004)

    2004

  7. [15]

    M., Dance, S

    Tabeart, J. M., Dance, S. L., Lawless, A. S., Nichols, N. K. & Waller, J. A. Improving the condition number of estimated covariance matrices. Tellus A: Dynamic Meteorology and Oceanography71, 1–19 (2019)

    2019

  8. [16]

    W.Applied Numerical Linear Algebra (SIAM, Philadelphia, 1997)

    Demmel, J. W.Applied Numerical Linear Algebra (SIAM, Philadelphia, 1997)

    1997

  9. [17]

    Higham, N. J. Accuracy and Stability of Numerical Algorithms (SIAM, Philadelphia, 2002), 2nd edn. 27

    2002

  10. [18]

    Schulz, G. Über die Bereitstellung von Iterationsverfahren höherer Ordnung für die Auflö- sung linearer Gleichungssysteme.ZAMM-Journal of Applied Mathematics and Mechanics/ Zeitschrift für Angewandte Mathematik und Mechanik13, 57–59 (1933)

    1933

  11. [19]

    Some new methods in matrix calculation

    Hotelling, H. Some new methods in matrix calculation. The Annals of Mathematical Statistics14, 1–34 (1943)

    1943

  12. [20]

    On a fast iterative method for approximate inverse of matrices

    Soleymani, F. On a fast iterative method for approximate inverse of matrices. Communications in Nonlinear Science and Numerical Simulation17, 2947–2956 (2012)

    2012

  13. [21]

    Regarding the accuracy of explicit inverse matrix methods

    Soleymani, F. Regarding the accuracy of explicit inverse matrix methods. Applied Mathematics and Computation219, 4142–4152 (2013)

    2013

  14. [22]

    Gower, R. M. & Richtárik, P. Randomized quasi-Newton updates are linearly convergent matrixinversionalgorithms. SIAM Journal on Matrix Analysis and Applications38, 1380– 1409 (2017)

    2017

  15. [23]

    & Takáč, M

    Richtárik, P. & Takáč, M. Stochastic reformulations of linear systems: algorithms and con- vergence theory.SIAM Journal on Matrix Analysis and Applications41, 487–524 (2020)

    2020

  16. [24]

    & Hennig, P

    Wenger, J. & Hennig, P. Probabilistic linear solvers for machine learning. In Advances in Neural Information Processing Systems, vol. 33, 6731–6742 (2020)

    2020

  17. [25]

    Aifer, M. et al. Thermodynamic linear algebra. npj Unconventional Computing1, 13 (2024)

    2024

  18. [26]

    & Sabo, F

    Todri-Sanial, A., Delacour, C., Abernot, M. & Sabo, F. Computing with oscillators from theoretical underpinnings to applications and demonstrators. npj Unconventional Computing1, 14 (2024). URLhttps://doi.org/10.1038/ s44335-024-00015-z

    2024

  19. [27]

    & Zhou, H

    Liu, S., Li, W., Zha, H. & Zhou, H. Neural parametric fokker–planck equation. SIAM Journal on Scientific Computing44, A2472–A2497 (2022)

    2022

  20. [28]

    Numerical solution of high dimensional stationary fokker–planck equations

    Sun, Y.et al. Numerical solution of high dimensional stationary fokker–planck equations. Applied Numerical Mathematics93, 1–21 (2015)

    2015

  21. [29]

    J., Hofmann, S

    Tye, N. J., Hofmann, S. & Stanley-Marbell, P. Materials and devices as solutions to computational problems in machine learning.Nature Electronics6, 479–490 (2023)

    2023

  22. [30]

    & Stanley-Marbell, P

    Meech, J., Tsoutsouras, V. & Stanley-Marbell, P. The data conversion bottleneck in analog computing accelerators. InMachine Learning with New Compute Paradigms (2023). URL https://openreview.net/forum?id=rabZPXiHGI

    2023

  23. [31]

    Forsythe, G. E. & Leibler, R. A. Matrix inversion by a Monte Carlo method. Mathematical Tables and Other Aids to Computation4, 127–129 (1950)

    1950

  24. [32]

    & Melanson, D

    Duffield, S., Donatella, K. & Melanson, D. thermox: Exact ou processes with jax.https: //github.com/normal-computing/thermox(2024)

    2024

  25. [33]

    Probability, Random Variables and Stochastic Processes (McGraw-Hill, 1991)

    Papoulis, A. Probability, Random Variables and Stochastic Processes (McGraw-Hill, 1991)

    1991

  26. [34]

    & Vandenberghe, L

    Boyd, S. & Vandenberghe, L. Convex Optimization (Cambridge University Press, Cam- bridge, 2004)

    2004

  27. [35]

    & Wright, S

    Nocedal, J. & Wright, S. J.Numerical Optimization (Springer, New York, 2006), 2nd edn. 28

    2006

  28. [36]

    Horn, R. A. & Johnson, C. R.Matrix Analysis (Cambridge University Press, 2012), 2nd edn

    2012

  29. [37]

    S.Matrix Iterative Analysis (Springer, Berlin, 2000), 2nd edn

    Varga, R. S.Matrix Iterative Analysis (Springer, Berlin, 2000), 2nd edn

    2000

  30. [38]

    T.Introduction to Optimization (Optimization Software, New York, 1987)

    Polyak, B. T.Introduction to Optimization (Optimization Software, New York, 1987)

    1987

  31. [39]

    Introductory Lectures on Convex Optimization: A Basic Course (Springer, New York, 2004)

    Nesterov, Y. Introductory Lectures on Convex Optimization: A Basic Course (Springer, New York, 2004)

    2004

  32. [40]

    Hansen, L. P. Large sample properties of generalized method of moments estimators. Econometrica50, 1029–1054 (1982)

    1982

  33. [41]

    Harris, C. R.et al. Array programming with NumPy.Nature585, 357–362 (2020). URL https://doi.org/10.1038/s41586-020-2649-2

    2020 doi

  34. [42]

    SciPy 1.0: Fundamental Algorithms for Scientific Computing in Python

    Virtanen, P.et al. SciPy 1.0: Fundamental Algorithms for Scientific Computing in Python. Nature Methods17, 261–272 (2020)

    2020

  35. [43]

    Horn, R. A. & Johnson, C. R.Matrix Analysis (Cambridge University Press, Cambridge, 2013), 2nd edn

    2013

  36. [44]

    Vannevar bush and the differential analyzer: the text and context of an early computer

    Owens, L. Vannevar bush and the differential analyzer: the text and context of an early computer. Technology and culture27, 63–95 (1986)

    1986

  37. [45]

    Uhlenbeck, G. E. & Ornstein, L. S. On the theory of the Brownian motion.Physical Review 36, 823 (1930)

    1930

  38. [46]

    Kloeden, P. E. & Platen, E. Numerical Solution of Stochastic Differential Equations (Springer-Verlag, Berlin, 1992)

    1992

  39. [47]

    The Laplace Microarchitecture for Tracking Data Uncertainty and Its Implementation in a RISC-V Processor.IEEE Micro42, 78–86 (2022)

    Tsoutsouras, V.et al. The Laplace Microarchitecture for Tracking Data Uncertainty and Its Implementation in a RISC-V Processor.IEEE Micro42, 78–86 (2022)

    2022

  40. [48]

    A., Elias, O

    Bilgin, B. A., Elias, O. H., Selby, M. & Stanley-Marbell, P. Quantization of Probability Distributions via Divide-and-Conquer: Convergence and Error Propagation under Distri- butional Arithmetic Operations.arXiv preprint arXiv:2505.15283 (2025).2505.15283

    2025

  41. [49]

    Coles, P. J. et al. Thermodynamic AI and the fluctuation frontier. In 2023 IEEE International Conference on Rebooting Computing (ICRC), 1–10 (IEEE, San Diego, CA, USA, 2023)

    2023

  42. [50]

    W., Hassidim, A

    Harrow, A. W., Hassidim, A. & Lloyd, S. Quantum algorithm for linear systems of equa- tions. Phys. Rev. Lett.103, 150502 (2009). URLhttps://link.aps.org/doi/10.1103/ PhysRevLett.103.150502

    2009

  43. [51]

    W.Handbook of Stochastic Methods for Physics, Chemistry and the Natural Sciences

    Gardiner, C. W.Handbook of Stochastic Methods for Physics, Chemistry and the Natural Sciences. Springer Series in Synergetics (Springer-Verlag, Berlin, 1985)

    1985

  44. [52]

    Thermodynamic algorithms for quadratic programming (2024)

    Bartosik, P.-L.et al. Thermodynamic algorithms for quadratic programming (2024). URL https://arxiv.org/abs/2411.14224.2411.14224

    2024 arXiv

  45. [53]

    Melanson, D. et al. Thermodynamic computing system for ai applications. Nature Communications16, 3757 (2025). URLhttps://doi.org/10.1038/ s41467-025-59011-x

    2025

  46. [54]

    Oseledets, I. V. Tensor-train decomposition. SIAM Journal on Scientific Computing33, 2295–2317 (2011). 29 Figure Legends Figure 1|A visual representation of the TTR discretization. Figure 2|A visual representation of how Langevin dynamics simulation is applied to approximate t...

    2011

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.