Beyond Ground States: Learning Many-Body Hamiltonians for p-Bit Invertible Arithmetic

arXiv:2611.14744 · cond-mat.dis-nn, cs.ET, physics.app-ph · Submitted 2026-11-10 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.

Kai: We're looking at the paper: "Beyond Ground States".

Mira: Here's the gist of the paper. Probabilistic-bit (p-bit) invertible logic encodes Boolean relations as an energy function whose ground states correspond to legal assignments, allowing computation in any direction via sampling.

Kai: The method frames Hamiltonian design as inverse statistical mechanics by fitting an energy function with interactions up to order four, ensuring the Gibbs law approaches uniformity on each clamp case's legal states under a gap floor that keeps every legal state below every illegal one (guaranteeing all ground states are legal).

Mira: Constrained maximum likelihood techniques, involving barrier Newton methods, pruning, and refitting candidate terms from a fixed pool, are employed.

Kai: Every delivered structure undergoes exact auditing across all cases.

Paper summary: Kai: So to recap this paper, 'Beyond Ground States: Learning Many-Body Hamiltonians for p-Bit Invertible Arithmetic', the core idea is framing Hamiltonian design as inverse statistical mechanics.

Mira: They are fitting an energy function with interactions up to order four, and they're trying to make that Gibbs law approach uniformity on each clamp case’s legal states under a gap floor.

Lev: This means they're ensuring every legal state has a gap below it compared to every illegal one, which guarantees that all the ground states found are actually legal assignments.

Kai: They use constrained maximum likelihood techniques, like barrier Newton methods and pruning, to fit these candidate terms from a fixed pool of interactions.

Mira: The main claim is that this formulation is convex over a fixed pool and certified by exact audit across all cases.

Lev: It's important because they show that this method works even when dealing with finite-time samplers that might traverse excited states, which traditional constructions don't handle well.

Kai: They also demonstrate equal tuning across multiplier width series and an adder ladder, showing the learned Hamiltonians reduce update counts in underdetermined modes for every sampler family.

Mira: It’s a method for learning Hamiltonians that is robust across different hardware setups, which is significant because it applies to many types of logic circuits.

Lev: The implication here is that we can design these energy functions using machine learning techniques rather than relying solely on hand-written encodings or traditional methods.

Kai: It suggests a way to generate Hamiltonians that are inherently well-behaved with respect to the legal assignments, which should make sampling much more efficient.

Mira: The paper is interested in how these learned Hamiltonians perform when compared to canonical encodings, especially concerning update counts and hitting times.

Conclusion: Kai: So looking at the full picture of 'Beyond Ground States: Learning Many-Body Hamiltonians for p-Bit Invertible Arithmetic', the authors are Tian Feng, Xutang Fan, and Yiyun Chen.

Mira: The title itself points to moving beyond just finding ground states to learning Hamiltonians that respect the structure of probabilistic-bit invertible arithmetic.

Lev: Essentially, they're showing how to build these energy functions by fitting them through a constrained optimization process that enforces strict constraints on the energy landscape.

Kai: The real implication for us is that this method lets us generate Hamiltonians for complex circuits where we don't have an explicit definition of the interactions, just some pool of candidates.

Mira: It means we can build systems with better properties regarding sampling efficiency, because they are guaranteed to operate within a structure that respects the legal assignments.

Lev: So for engineers working on real hardware like superconducting qubits or CMOS p-bits, this gives them a way to generate structures that are empirically shown to require fewer updates than standard methods.

Advanced Cryptography and System Security Key Laboratory of Sichuan Province · The Hong Kong University of Science and Technology (Guangzhou)

cond-mat.dis-nn, cs.ET, physics.app-ph

Submitted: 2026-11-10

Updated: 2026-11-10

Comments: arXiv v1

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 91/100

The gist: Probabilistic-bit (p-bit) invertible logic encodes Boolean relations as an energy function whose ground states correspond to legal assignments, allowing computation in any direction via sampling.

Key concepts

Probabilistic-bit (p-bit) invertible logic
This is a way to encode Boolean relations using an energy function. This allows for computation in any direction by sampling the ground states of the Hamiltonian. The system's legal assignments correspond to the ground states, and this framework enables computation through statistical mechanics principles.
Hamiltonian design as inverse statistical mechanics
The method treats designing a Hamiltonian like fitting an energy function. The goal is to ensure that the Gibbs law approaches uniformity over all legal states within each constraint case. This is achieved by enforcing a gap floor, which guarantees that every legal state has an energy lower than every illegal state.
Constrained maximum likelihood techniques
These are the specific mathematical tools used to find the optimal Hamiltonian. They involve barrier Newton methods, pruning candidate terms from a fixed pool, and refitting those terms. The objective is to minimize a complex function that balances fitting the energy function with maintaining strict gap inequalities across all constraint cases.

Terminology

Summary

Probabilistic-bit (p-bit) invertible logic encodes Boolean relations as an energy function whose ground states correspond to legal assignments, allowing computation in any direction via sampling. p. 0 The method frames Hamiltonian design as inverse statistical mechanics by fitting an energy function with interactions up to order four, ensuring the Gibbs law approaches uniformity on each clamp case's legal states under a gap floor that keeps every legal state below every illegal one (guaranteeing all ground states are legal). Constrained maximum likelihood techniques, involving barrier Newton methods, pruning, and refitting candidate terms from a fixed pool, are employed. p. 0 Every delivered structure undergoes exact auditing across all cases. p. 4 The objective function minimized is J (theta) = sum m omega m C m sum c in C m h Z c (theta) + beta fit E p* c [E theta] - H(p* c) i, enforcing per-case gap inequalities: E theta (y) - E theta (x) = theta phi(x) - phi(y) at least g for all c, x in L c, y in X c L c. p. 3

The study presents three primary contributions: (i) a constrained Gibbs-likelihood formulation for many-body invertible-logic Hamiltonians that is convex over a fixed pool and certified by exact ground-state audits; (ii) an equal-tuning evaluation across multiplier width series and an adder ladder demonstrating that learned Hamiltonians reduce update counts in underdetermined modes across every sampler family; and (iii) an ablation study separating coefficient values from hypergraph support at 4 times 4 alongside end-to-end hardware costs. p. 1

The empirical results indicate that learned Hamiltonians require fewer updates than canonical encodings across all tested sampler families, ranging from 4 times 4 through 6 times 6. p. 5 For the multiplier series, the best-solver ratio rho* in updates is 0. p. 655 [0.45, 0.65] for a 7 times 7 multiplier and all widths (Table II). p. 8 For the adder ladder, learned adders reduce best-solver updates to 0. p. 771 [0.63, 0.80] at 8 bits and reach about a quarter of this at 24 and 32 bits (Table II). p. 0 Hardware measurements show that color-sweep hard-threshold SSA fits digital FPGA fabric most directly, with real coefficients quantized to W-bit integers under a single scale per structure to maximize the normalized minimum case gap. p. 0 These structures are certified across all cases and modes at their tested word size and all wider ladder words. p. 0 Learned structures preserve canonical color counts while requiring more routed lookup tables except at 16 bits due to evaluating more incident entries. p. 6

The conclusion is that update reductions translate to other underdetermined clamp modes, with work ratios governed by incident fan-in. p. 2 Learned Hamiltonians reduce update counts on every panel of both ladders (Supplementary Fig. p. 1 S4 and Table S7) while matching canonical chromatic numbers in the inverse modes run on the board, achieving 0. p. 9020 of canonical clock cycles at 6 times 6 in exchange for larger lookup-table area. p. 5 The method is applicable to any gate-level circuit whose clamp cases can be listed and audited. p. 1

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Constrained Gibbs-likelihood formulation enables learning many-body Hamiltonians by minimizing J (Eq. 5), which is convex on a fixed pool and certified by exact ground-state audits (Eq. 9). This allows AI systems to find legal assignments for Boolean relations, such as in invertible logic circuits, while ensuring every legal state lies at least g below every illegal state of its case.

  2. The method utilizes constrained maximum likelihood fitting (barrier Newton, pruning, refitting) to fit candidate terms from a fixed pool. This process is verified exactly over all cases and scored by time to a verified solution under restarts using its best of nine tuned samplers, providing a robust metric for performance evaluation.

  3. Learned multipliers and adders can be optimized across width series (4x4 to 7x7) using a sampled-likelihood engine, achieving significant speedups over hand-written encodings in underdetermined modes. For instance, learned multipliers take 0.41 of the hand-written encoding’s cycles at 4x4 and drop to 0.057 at 7x7 with up to 2.4 times its lookup tables.

  4. The system employs equal tuning across different sampler families (SSA, HB, PT, etc.) for both canonical and learned Hamiltonians, demonstrating that learned Hamiltonians cut update counts in underdetermined modes across every sampler family. This equal tuning accelerates hand-written encodings 2.8- to 6.5-fold over standard SSA.

  5. The hardware deployment leverages color-sweep hard-threshold SSA on FPGA fabric, which fits digital logic most directly and reduces each spin update to saturating integer addition, sign readout, and a single random bit. This architecture is applied to the pool P3 multiplier series alongside the 8-to-32-bit adder ladder.

  6. The system can achieve near exact correctness on hardware by quantizing coefficients to W-bit integers under a single scale per structure, maximizing the normalized minimum case gap and certified at the narrowest tested ladder word keeping every legal state below every illegal one across all cases and modes at that word.

  7. The system provides an analytical bridge relating static divergence Dc to finite-time success probability via Proposition 7, which states that "99% verified success requires at most Nrest >= log(0.01) / log(1 - bc) restarts," quantifying the required effort for finding a solution in finite time.

Sources

Related papers