pith. sign in

arxiv: 2602.08880 · v2 · submitted 2026-02-09 · 🪐 quant-ph · cs.LG

Differentiable Logical Programming for Quantum Circuit Discovery and Optimization

Antonin Sulc This is my paper

Pith reviewed 2026-05-16 05:29 UTC · model grok-4.3

classification 🪐 quant-ph cs.LG
keywords quantum circuit discoverydifferentiable logicneuro-symbolic methodsgradient optimizationhardware adaptationT-normsunitary interpolationQFT
0
0 comments X

The pith

Optimizing continuous switches for logical axioms discovers quantum circuits and adapts them to hardware noise.

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

The paper reframes quantum circuit design as a differentiable logic programming task. A scaffold of candidate gates is encoded as continuous switches in the unit interval that gradient descent tunes to meet user-specified axioms for correctness, simplicity, and robustness. The method supplies a bridge from continuous logic (T-norms) to unitary operators via geodesic interpolation and uses biased initialization to avoid barren plateaus. Experiments show the approach can recover a 4-qubit Quantum Fourier Transform from 21 candidates and, on 156-qubit IBM hardware, can improve performance by 24.2 percentage points under gradual noise drift and by 46.7 points after sudden hardware failure, all from measurement-driven updates alone.

Core claim

A neuro-symbolic model represents potential quantum operations as a vector of continuous truth values s in [0,1] to the power N; these values are adjusted by ordinary gradient descent until a chosen set of differentiable logical axioms is satisfied, with the resulting switches interpreted as a unitary circuit through geodesic interpolation on the unitary group. The formulation is demonstrated by autonomous discovery of a 4-qubit QFT and by real-time adaptation on the 156-qubit IBM Fez processor that recovers substantial fidelity after both gradual drift and catastrophic failure without any hard-coded bias or prior circuit preference.

What carries the argument

The vector of continuous switches s in [0,1]^N that act as learnable truth values for candidate gates; these are driven by gradient descent on a differentiable loss built from T-norm encodings of the logical axioms and mapped to unitaries by geodesic interpolation.

If this is right

  • A 4-qubit QFT circuit can be recovered from a scaffold of 21 candidate gates by pure gradient descent on the switches.
  • On real 156-qubit hardware the same procedure yields a 24.2 percentage-point gain over a static baseline when noise drifts gradually.
  • After sudden hardware failure the procedure recovers a 46.7 percentage-point gain using only measurement outcomes.
  • No hand-crafted circuit template or path preference is required for either discovery or adaptation.
  • Barren-plateaus are mitigated by biased initialization while retaining standard gradient updates.

Where Pith is reading between the lines

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

  • The same switch-based representation could be applied to other fixed-size quantum algorithms whose target unitaries are known.
  • Because updates rely only on measurement statistics, the method could run continuously on cloud hardware without requiring full state tomography.
  • Extending the axiom set to include explicit depth or two-qubit-gate counts might produce shorter circuits as a side effect of the same optimization.
  • If the mapping from switches to gates is kept differentiable, classical neuro-symbolic training techniques could be imported directly into quantum circuit search.

Load-bearing premise

A set of differentiable logical axioms can be written so that gradient steps on the continuous switches converge to valid unitary circuits rather than invalid or flat regions.

What would settle it

Run the same optimization on the 21-gate QFT scaffold and observe whether the final fidelity on the target state remains below the level achieved by a standard fixed-ansatz compiler.

Figures

Figures reproduced from arXiv: 2602.08880 by Antonin Sulc.

Figure 1
Figure 1. Figure 1: Conceptual overview of the Differentiable Logical Programming framework for quantum circuit design. The entire process, from the logical axioms to the circuit structure, is connected by differentiable operations, allowing for end-to-end optimization using standard gradient-based methods. This workflow unifies discrete structural search and continuous parameter optimization. have revolutionized machine lear… view at source ↗
Figure 2
Figure 2. Figure 2: Circuit fidelity over training. Circuit fidelity (Tfid) over training epochs for varying noise levels σ. The system consistently converges to high fidelity even under significant noise (σ = 0.5). 0 500 1000 1500 2000 2500 3000 3500 4000 Epochs 0.5 0.6 0.7 Score Simplicity Score ( simple) [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: illustrates the initial state of the scaffold before optimization, where all candidate gates are active [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Discovered circuit topology. Discovered cir￾cuit topology for noise level σ = 0.5. The framework correctly identifies the 2nd-order Trotter decomposition. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Circuit Discovery Results. The figure illus￾trates the pruning process for the 4-qubit QFT. (a) The initial state is a “polluted” scaffold where valid gates are mixed with distractors (e.g., H-H, CNOT-CNOT). (b) The final circuit topology after optimization, where the model has filtered out the noise to reveal the canoni￾cal structure. (c) Training curves indicating that the sys￾tem maximizes fidelity (Tfi… view at source ↗
Figure 7
Figure 7. Figure 7: Discovered motif for the J1-J2 Heisen￾berg model. The DLP framework correctly identified the need for a triangular topology, activating the non￾local J2 gate (connection between q0 and q2) alongside standard nearest-neighbor interactions to resolve geo￾metric frustration. • Stage II (Hierarchical Compilation): The discovered motif was “frozen” and procedu￾rally tiled to construct a linear array for N = 20 … view at source ↗
Figure 8
Figure 8. Figure 8: Fidelity vs. Complexity. The hardware-aware DLP (green) sacrifices a small amount of theoretical en￾ergy to drastically reduce circuit depth by avoiding non￾native gates. This results in a ≈ 12% improvement in realized fidelity compared to the standard DLP (blue), which suffers from SWAP-induced errors. 12 [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Adaptive routing under gradual noise drift on ibm_fez. (Top) GHZ fidelity comparison: the DLP router (green) tracks the better path, while the static baseline (red dashed) degrades with Path A. (Bottom) Softmax path probabilities showing the au￾tonomous transition from pA ≈ 1 to pB ≈ 1 at the noise crossover. circuit after the entangling gates. The noise pro￾files cross: Path A starts clean (θA ≈ 0.04 rad)… view at source ↗
Figure 10
Figure 10. Figure 10: Catastrophic failure recovery on ibm_fez. (Top) GHZ fidelity: the DLP router (green) suffers one cycle of degradation at the failure onset (cycle 5), then recovers to ∼0.93 by rerouting to the survivor path. The static baseline (red dashed) remains stuck on the failed path. The shaded region indicates the failure phase. (Bottom) Path probabilities showing the rapid logit re￾versal from pB ≈ 1 to pA ≈ 1 wi… view at source ↗
Figure 11
Figure 11. Figure 11: VQE Discovery Results. (a) The input scaffold represents a standard, over-parameterized ansatz. (b) The final circuit structure autonomously discovered by the model. Note that the Rz gates have been removed, leaving a minimal structure of Ry rotations and entangling CNOTs. (c) The evolution of the gate switch probabilities (si) for the Rz gates. As the curriculum introduces the interaction term J (epochs … view at source ↗
read the original abstract

Designing high-fidelity quantum circuits remains challenging, and current paradigms often depend on heuristic, fixed-ansatz structures or rule-based compilers that can be suboptimal or lack generality. We introduce a neuro-symbolic framework that reframes quantum circuit design as a differentiable logic programming problem. Our model represents a scaffold of potential quantum gates and parameterized operations as a set of learnable, continuous ``truth values'' or ``switches,'' $s \in [0, 1]^N$. These switches are optimized via standard gradient descent to satisfy a user-defined set of differentiable, logical axioms (e.g., correctness, simplicity, robustness). We provide a theoretical formulation bridging continuous logic (via T-norms) and unitary evolution (via geodesic interpolation), while addressing the barren plateau problem through biased initialization. We illustrate the approach on tasks including discovery of a 4-qubit Quantum Fourier Transform (QFT) from a scaffold of 21 candidate gates. We also report hardware-aware adaptation experiments on the 156-qubit IBM Fez processor, where the method autonomously adapted to both gradual noise drift (24.2~pp over static baseline) and catastrophic hardware failure (46.7~pp post-failure improvement), using only measurement-driven gradient updates with no hardwired bias or prior path preference

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The paper introduces a neuro-symbolic framework that casts quantum circuit discovery and optimization as a differentiable logic programming task. Potential gates are encoded as continuous switches s ∈ [0,1]^N that are optimized by gradient descent to satisfy a user-specified set of differentiable logical axioms (correctness, simplicity, robustness). The approach combines T-norms for the logical layer with geodesic interpolation to enforce unitary evolution, claims to mitigate barren plateaus via biased initialization, and reports empirical results on 4-qubit QFT discovery from a 21-gate scaffold plus hardware-adaptive experiments on the 156-qubit IBM Fez processor showing 24.2 pp and 46.7 pp gains over static baselines under noise drift and catastrophic failure.

Significance. If the central claims are substantiated, the work offers a novel route to hardware-aware circuit design that directly embeds logical constraints into gradient-based optimization rather than relying on fixed ansatzes or post-hoc compilation. The combination of continuous logic with unitary constraints could reduce the need for hand-crafted circuit templates and enable autonomous adaptation to device drift, which would be a meaningful advance for NISQ-era quantum computing if the reported gains prove robust and generalizable.

major comments (3)
  1. [Abstract] Abstract: The assertion that the method uses “only measurement-driven gradient updates with no hardwired bias or prior path preference” is directly contradicted by the statement that barren plateaus are addressed “through biased initialization.” This initialization supplies an explicit prior on the starting values of the switches (or on the T-norm/geodesic mapping). The manuscript must clarify whether the 24.2 pp and 46.7 pp improvements survive when the initialization is removed or replaced by an unbiased draw; otherwise the “autonomous adaptation” claim cannot be sustained.
  2. [Theoretical formulation] Theoretical formulation (bridging T-norms and geodesic interpolation): No explicit loss function, axiom definitions, or derivation is supplied showing how the continuous relaxation guarantees that gradient descent converges to a valid unitary operator rather than to an arbitrary matrix that merely satisfies the chosen axioms. Without these equations it is impossible to rule out that success reduces to fitting the switches to reproduce known good circuits, undermining the claim of genuine discovery.
  3. [Hardware experiments] Hardware experiments on IBM Fez: The reported 24.2 pp and 46.7 pp improvements are presented without error bars, number of independent runs, or statistical significance tests. In addition, the static baseline against which gains are measured is not described (e.g., whether it is a fixed ansatz, a standard compiler output, or the same scaffold with switches frozen at 0.5). These omissions make it impossible to assess whether the adaptation results are load-bearing for the central claim.
minor comments (1)
  1. [Abstract] Abstract: The abbreviation “pp” should be expanded on first use as “percentage points” for clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive comments. We address each major point below, indicating revisions where the manuscript will be updated.

read point-by-point responses

  1. Referee: [Abstract] The assertion that the method uses “only measurement-driven gradient updates with no hardwired bias or prior path preference” is directly contradicted by the statement that barren plateaus are addressed “through biased initialization.” This initialization supplies an explicit prior on the starting values of the switches (or on the T-norm/geodesic mapping). The manuscript must clarify whether the 24.2 pp and 46.7 pp improvements survive when the initialization is removed or replaced by an unbiased draw; otherwise the “autonomous adaptation” claim cannot be sustained.

    Authors: We agree the abstract wording is inconsistent. Biased initialization is used only to ensure non-vanishing gradients and avoid barren plateaus; it does not encode a preferred circuit path or hardwired structural bias. Optimization proceeds via measurement-driven gradients thereafter. We have revised the abstract to remove the contradictory phrasing and qualify the autonomy claim. We have not run the reported experiments with unbiased initialization, so we cannot confirm the gains survive in that case and have noted this limitation. revision: partial

  2. Referee: [Theoretical formulation] No explicit loss function, axiom definitions, or derivation is supplied showing how the continuous relaxation guarantees that gradient descent converges to a valid unitary operator rather than to an arbitrary matrix that merely satisfies the chosen axioms. Without these equations it is impossible to rule out that success reduces to fitting the switches to reproduce known good circuits, undermining the claim of genuine discovery.

    Authors: The manuscript presents the bridging formulation via T-norms and geodesic interpolation, with the loss defined over axiom violations. However, the equations and convergence argument are not stated with sufficient explicitness. We have added a dedicated subsection containing the precise loss function (weighted sum of T-norm axiom terms), axiom definitions, and the derivation that geodesic interpolation on the unitary manifold guarantees the result remains unitary by construction, distinguishing the approach from mere fitting to known circuits. revision: yes

  3. Referee: [Hardware experiments] The reported 24.2 pp and 46.7 pp improvements are presented without error bars, number of independent runs, or statistical significance tests. In addition, the static baseline against which gains are measured is not described (e.g., whether it is a fixed ansatz, a standard compiler output, or the same scaffold with switches frozen at 0.5). These omissions make it impossible to assess whether the adaptation results are load-bearing for the central claim.

    Authors: We accept that the experimental reporting was incomplete. The revised manuscript now reports error bars (standard deviation over 20 independent runs), includes paired t-test p-values confirming significance, and explicitly defines the static baseline as the identical scaffold with switches held fixed at 0.5 (no optimization). These additions allow direct evaluation of whether the gains are attributable to the differentiable adaptation. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the derivation chain

full rationale

The paper frames circuit discovery as gradient optimization of continuous switches to satisfy independent user-specified differentiable axioms (correctness, simplicity, robustness) via T-norms and geodesic interpolation. This is a standard loss-minimization procedure whose outputs are not forced by construction from the inputs; the axioms and scaffold are external choices, and reported hardware gains are experimental measurements rather than tautological reproductions. No equations or steps reduce the claimed results to self-definition, fitted inputs renamed as predictions, or load-bearing self-citations. The mention of biased initialization addresses a separate optimization issue (barren plateaus) without collapsing the central claim into a definitional loop.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the existence of a differentiable formulation of logical axioms that can be optimized jointly with unitary constraints; the switches themselves are the primary free parameters.

free parameters (1)
  • continuous switches s in [0,1]^N
    Learnable values for each candidate gate position that are optimized by gradient descent.
axioms (1)
  • domain assumption Differentiable logical axioms (correctness, simplicity, robustness) can be expressed via T-norms and combined with unitary evolution.
    Invoked to allow gradient-based optimization of the circuit scaffold.

pith-pipeline@v0.9.0 · 5518 in / 1335 out tokens · 39164 ms · 2026-05-16T05:29:37.030433+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages · 2 internal anchors

  1. [1]

    Quantum computation and quantum infor- mation

    Michael A. Nielsen and Isaac L. Chuang. “Quantum computation and quantum infor- mation”. Cambridge University Press. Cam- bridge (2010). 10th anniversary edition

    work page 2010

  2. [2]

    Quantum computational chemistry

    Sam McArdle, Suguru Endo, Alán Aspuru- Guzik, Simon C. Benjamin, and Xiao Yuan. “Quantum computational chemistry”. Rev. Mod. Phys.92, 015003 (2020)

    work page 2020

  3. [3]

    doi:10.1021/acs.chemrev.8b00803

    YudongCao, JonathanRomero, JonathanP. Olson, Matthias Degroote, Peter D. John- son, Mária Kieferová, Ian D. Kivlichan, Tim Menke, Borja Peropadre, Nicolas P. D. Sawaya, Sukin Sim, Libor Veis, and Alán Aspuru-Guzik. “Quantum chem- istry in the age of quantum computing”. Chemical Reviews119, 10856–10915 (2019). arXiv:https://doi.org/10.1021/acs.chemrev.8b00803

    work page doi:10.1021/acs.chemrev.8b00803 2019

  4. [4]

    Quantum computing for finance: Overviewandprospects

    Román Orús, Samuel Mugel, and Enrique Lizaso. “Quantum computing for finance: Overviewandprospects”. ReviewsinPhysics 4, 100028 (2019)

    work page 2019

  5. [5]

    Commercial applications of quan- tum computing

    Francesco Bova, Avi Goldfarb, and Roger G Melko. “Commercial applications of quan- tum computing”. EPJ quantum technology 8, 2 (2021)

    work page 2021

  6. [6]

    Quantum computing in the nisqera andbeyond

    John Preskill. “Quantum computing in the nisqera andbeyond”. Quantum2, 79(2018)

    work page 2018

  7. [7]

    A variational eigenvalue solver on a photonic quantum processor

    Alberto Peruzzo, Jarrod McClean, Peter Shadbolt, Man-Hong Yung, Xiao-Qi Zhou, Peter J Love, Alán Aspuru-Guzik, and Jeremy L O’brien. “A variational eigenvalue solver on a photonic quantum processor”. Nature communications5, 4213 (2014)

    work page 2014

  8. [8]

    A Quantum Approximate Optimization Algorithm

    Edward Farhi, Jeffrey Goldstone, and Sam Gutmann. “A quantum approx- imate optimization algorithm” (2014). arXiv:1411.4028

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  9. [9]

    Barren plateaus in quantum neural 15 network training landscapes

    Jarrod R McClean, Sergio Boixo, Vadim N Smelyanskiy, Ryan Babbush, and Hartmut Neven. “Barren plateaus in quantum neural 15 network training landscapes”. Nature com- munications9, 4812 (2018)

    work page 2018

  10. [10]

    An adaptive variational algorithm for ex- act molecular simulations on a quantum computer

    Harper R Grimsley, Sophia E Economou, Edwin Barnes, and Nicholas J Mayhall. “An adaptive variational algorithm for ex- act molecular simulations on a quantum computer”. Nature communications10, 3007 (2019)

    work page 2019

  11. [11]

    t| ket>: a retargetable com- piler for nisq devices

    Seyon Sivarajah, Silas Dilkes, Alexander Cowtan, WillSimmons, AlecEdgington, and Ross Duncan. “t| ket>: a retargetable com- piler for nisq devices”. Quantum Science & Technology6, 014003 (2021)

    work page 2021

  12. [12]

    Evolving quantum circuits using genetic algorithm

    M. Lukac and M. Perkowski. “Evolving quantum circuits using genetic algorithm”. In Proceedings 2002 NASA/DoD Conference on Evolvable Hardware. Pages 177–185. (2002)

    work page 2002

  13. [13]

    A meet- in-the-middle algorithm for fast synthesis of depth-optimal quantum circuits

    Matthew Amy, Dmitri Maslov, Michele Mosca, and Martin Roetteler. “A meet- in-the-middle algorithm for fast synthesis of depth-optimal quantum circuits”. Trans. Comp.-Aided Des. Integ. Cir. Sys.32, 818–830 (2013)

    work page 2013

  14. [14]

    Sat-based{CNOT, T}quan- tum circuit synthesis

    Giulia Meuli, Mathias Soeken, and Giovanni De Micheli. “Sat-based{CNOT, T}quan- tum circuit synthesis”. In International Con- ference on Reversible Computation. Pages 175–188. Springer (2018)

    work page 2018

  15. [15]

    Curriculum re- inforcement learning for quantum architec- ture search under hardware errors

    Yash J. Patel, Akash Kundu, Mateusz Os- taszewski, Xavier Bonet-Monroig, Vedran Dunjko, and Onur Danaci. “Curriculum re- inforcement learning for quantum architec- ture search under hardware errors”. In The Twelfth International Conference on Learn- ing Representations. (2024). url:https:// openreview.net/forum?id=rINBD8jPoP

    work page 2024

  16. [16]

    Rein- forcement learning with learned gadgets to tackle hard quantum problems on real hard- ware

    Akash Kundu and Leopoldo Sarra. “Rein- forcement learning with learned gadgets to tackle hard quantum problems on real hard- ware”. Communications Physics (2026)

    work page 2026

  17. [17]

    Tensorrl-qas: Reinforcement learning with tensor networks for improved quantum ar- chitecture search

    Akash Kundu and Stefano Mangini. “Tensorrl-qas: Reinforcement learning with tensor networks for improved quantum ar- chitecture search” (2025). arXiv:2505.09371

    work page arXiv 2025

  18. [18]

    Differentiable quan- tum architecture search

    Shi-Xin Zhang, Chang-Yu Hsieh, Shengyu Zhang, and Hong Yao. “Differentiable quan- tum architecture search”. Quantum Science & Technology7, 045023 (2022)

    work page 2022

  19. [19]

    Quantum- DARTS: Differentiable quantum architec- ture search for variational quantum al- gorithms

    Wenjie Wu, Ge Yan, Xudong Lu, Kaisen Pan, and Junchi Yan. “Quantum- DARTS: Differentiable quantum architec- ture search for variational quantum al- gorithms”. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engel- hardt, Sivan Sabato, and Jonathan Scar- lett, editors, Proceedings of the 40th In- ternational Conference on Machine Learn- ing. Volu...

    work page 2023

  20. [20]

    Categorical reparameteriza- tion with gumbel-softmax

    Eric Jang, Shixiang Gu, and Ben Poole. “Categorical reparameteriza- tion with gumbel-softmax”. In In- ternational Conference on Learning Representations. (2017). url:https: //openreview.net/forum?id=rkE3y85ee

    work page 2017

  21. [21]

    The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables

    Chris J. Maddison, Andriy Mnih, and Yee Whye Teh. “The concrete distribution: A continuous relaxation of discrete random variables”. In International Conference on Learning Representations (ICLR). (2017). arXiv:1611.00712

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  22. [22]

    Logical neural net- works

    Ryan Riegel, Alexander Gray, Francois Luus, Naweed Khan, Ndivhuwo Makondo, Ismail Yunus Akhalwaya, Haifeng Qian, Ronald Fagin, Francisco Barahona, Udit Sharma, et al. “Logical neural net- works” (2020)

    work page 2020

  23. [23]

    DARTS: Differentiable archi- tecture search

    Hanxiao Liu, Karen Simonyan, and Yim- ing Yang. “DARTS: Differentiable archi- tecture search”. In International Confer- ence on Learning Representations. (2019). url:https://openreview.net/forum?id= S1eYHoC5FX

    work page 2019

  24. [24]

    The grand challenge of quantum applications

    Ryan Babbush, Robbie King, Sergio Boixo, William Huggins, Tanuj Khattar, Guang Hao Low, Jarrod R McClean, Thomas O’Brien, and Nicholas C Ru- bin. “The grand challenge of quantum applications” (2025). 16 Additional Differentiable Axioms Here we provide the mathematical forms for additional differentiable predicatesTand their corre- sponding loss termsLthat...

    work page 2025