Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

Information-Time Benchmarks for Comparing Learning Dynamics

Janiszewski, Wojciech

Published Jan 1, 2025• 0 citations
Publisher

Abstract

This paper proposes information-time benchmarks (BIT) as a geometry- and physics-aware way to compare learning dynamics in quantum and classical models. Building on Bures/Fisher information geometry, we define information-time as the length of an observer’s state trajectory in prediction space and show that it is constrained by observer-specific quantum speed limits and entropy continuity bounds, leading to an explicit energy–entropy–information-time trade-off. On the practical side, the paper develops implementable estimators of information-time for standard deep-learning pipelines (gradient-based and Fisher-weighted, including diagonal Fisher approximations), discusses their computational complexity and limitations, and shows how to turn them into BIT scores such as information-time per unit wall-clock time or per Joule. We position BIT relative to path length in parameter space, sharpness/SAM, gradient flow and NTK perspectives, and natural gradient methods, highlighting that BIT uniquely measures trajectory length directly in prediction space and is bounded by device-level energy variance. Finally, we formulate explicit, falsifiable hypotheses linking BIT to transfer learning, sharpness of minima, optimiser choice (e.g. SGD vs natural gradient/K-FAC), and quantum–classical comparisons, and we outline concrete experimental protocols for future validation. The goal is not to replace existing benchmarks, but to add an information-geometric and resource-aware lens on learning dynamics that can be applied across architectures and hardware platforms.

DOI
10.17605/osf.io/tsm8q

Related papers

Information-Time of Quantum Observers: Geometry, Quantum Speed Limits, and Implications for Quantum Computing and AI

Janiszewski, Wojciech

This paper introduces a notion of information-time for quantum observers, defined as the Bures length of the trajectory of an observer’s reduced density matrix in state space. In contrast to approaches where time is an external parameter or a purely global relational variable, information-time is intrinsically observer-dependent and built only from the information accessible to a given physical subsystem (no appeal to consciousness or measurement postulates is made). On this basis the work derives two formal results: (i) an observer-specific Mandelstam–Tamm–type quantum speed limit that bounds the physical time required to generate a given amount of information-time, and (ii) an energy–entropy–information-time inequality that relates the change of an observer’s von Neumann entropy to available energy, physical time and Hilbert-space dimension. Two simple models are used to illustrate the framework: a dephasing spin model from Quantum Darwinism, and a Page–Wootters–type relational clock including a “blind” observer that experiences zero information-time. The paper argues that this “HCI” (Hilbert-space Change of Information) perspective offers a useful geometric language for the experience of time by quantum subsystems and discusses potential implications for quantum computing, quantum machine learning and AI. In particular, it suggests using information-time as a basis for hardware-agnostic benchmarks of learning and computation, complementing standard measures such as wall-clock time and FLOPs.

•Dec 1, 2025•0 citations

Fisher-Informational Time: A Causal-Geometric Framework for Emergent Clock Time Physical Distinguishability

J. Sumaya-Martinez

We develop a Fisher-informational reformulation of physical time in which clock time is not regarded as a fundamental ontological substance, but as an emergent calibration of causally ordered distinguishability among physical states. The operational starting point is that clocks do not measure time itself; rather, they instantiate reproducible physical processes whose distinguishable states are correlated with other events. We introduce a causal-informational parameter, denoted by Lambda_F, defined as an accumulated Fisher-geometric distance along a causally admissible trajectory in state space. In classical statistical systems, this parameter is generated by the Fisher information metric; in quantum systems, the corresponding construction is associated with quantum Fisher information, the Bures metric, and the Fubini-Study geometry of projective Hilbert space. The manuscript distinguishes model-dependent Fisher information from quantum Fisher information, clarifies the reparameterization of Schrodinger dynamics, and gives explicit examples involving a qubit clock, an exponential decay process, and a Fisher characterization of clock quality. The proposal is positioned relative to relational time, the Page-Wootters mechanism, thermal time, quantum speed-limit relations, information geometry, and the problem of time in quantum gravity. We do not claim that relational or emergent time is new. The specific contribution is the use of Fisher distinguishability as an operational precursor from which ordinary clock time can be reconstructed. In this sense, the central statement of the paper is: time is not measured by clocks; clock time is reconstructed from the Fisher distinguishability accumulated along causally ordered physical changes.

quant-pharXiv:2605.03958•May 6, 2026•0 citations

Unified speed limits in classical and quantum dynamics via temporal Fisher information

Tomohiro Nishiyama, Yoshihiko Hasegawa

The importance of Fisher information is increasing in nonequilibrium thermodynamics, as it has played a fundamental role in trade-off relations such as thermodynamic uncertainty relations and speed limits. In this work, we investigate temporal Fisher information, which measures the temporal information content encoded in probability distributions, for both classical and quantum systems. We establish that temporal Fisher information is bounded from above by physical costs, such as entropy production in classical Langevin and Markov processes, and the quantum entropy production and the expectation of the effective interaction Hamiltonian in open quantum systems. Conversely, temporal Fisher information is bounded from below by statistical distances (e.g., the Bhattacharyya arccos distance), leading to classical and quantum speed limits that constrain the minimal time required for state transformations. We perform numerical simulations on two quantum dot models to validate the obtained bounds. Our work provides a unified perspective on speed limits from the point of view of temporal Fisher information in both classical and quantum dynamics.

arXiv:2504.04790•Physical review. E•May 13, 2026•6 citations

Information geometry approach to quantum stochastic thermodynamics

Laetitia P. Bettmann, John Goold

Recent advancements have revealed new links between information geometry and classical stochastic thermodynamics, particularly through the Fisher information (FI) with respect to time. Recognizing the nonuniqueness of the quantum Fisher metric in Hilbert space, we exploit the fact that any quantum Fisher information (QFI) can be decomposed into a metric-independent incoherent part and a metric-dependent coherent contribution. We demonstrate that the incoherent component of any QFI can be directly linked to entropic acceleration, and for GKSL dynamics with local detailed balance, to the rate of change of generalized thermodynamic forces and entropic flow, paralleling the classical results. Furthermore, we tighten a classical uncertainty relation between the geometric uncertainty of a path in state space and the time-averaged rate of information change and demonstrate that it also holds for quantum systems. We generalize a classical geometric bound on the entropy rate for far-from-equilibrium processes by incorporating a nonnegative quantum contribution that arises from the geometric action due to coherent dynamics. Finally, we apply an information-geometric analysis to the recently proposed quantum-thermodynamic Mpemba effect, demonstrating this framework's ability to capture thermodynamic phenomena.

arXiv:2409.06083•Physical review. E•Jan 16, 2025•28 citations

Performance Evaluation of Ising and QUBO Variable Encodings in Boltzmann Machine Learning

Yasushi Hasegawa, Masayuki Ohzeki

We compare Ising ([Formula: see text]) and QUBO ([Formula: see text]) encodings for Boltzmann machine learning under controlled protocols that fix the sampler, optimizer, and learning-rate design within each comparison. Exploiting the identity that the Fisher information matrix (FIM) equals the covariance of sufficient statistics, we visualize empirical moments from model samples and reveal systematic, representation-dependent differences. QUBO induces larger cross terms between first- and second-order statistics, creating more small-eigenvalue directions in the FIM and lowering spectral entropy. This ill-conditioning explains slower convergence under stochastic gradient descent (SGD). In contrast, full-FIM natural gradient descent (NGD), which rescales updates by the FIM metric, achieves similar convergence across encodings, whereas diagonal-FIM approximation can reintroduce representation-dependent differences. Practically, for SGD-based training, the Ising encoding provides more isotropic curvature and faster convergence; for QUBO, centering/scaling or NGD-style preconditioning mitigates curvature pathologies. These results clarify how representation shapes information geometry and finite-time learning dynamics in Boltzmann machines and yield actionable guidelines for variable encoding and preprocessing.

arXiv:2510.13210•Journal of the Physical Society of Japan•Sep 29, 2026•0 citations

Quantum Machine Learning vs Classical Machine Learning: A Comparative Study from Scratch

Juan Pablo Barón Sandoval

This thesis presents a comparative study of quantum machine learning and classical machine learning developed from first principles. It begins by establishing the foundations of the quantum circuit model, observables, and projective measurement, and then revisits the classical learning framework with an emphasis on risk minimisation, task-aligned metrics, and rigorous evaluation protocols. Building on this shared perspective, the thesis examines two central paradigms in quantum machine learning: quantum kernels and variational quantum circuits. A unifying theme is that quantum models must be assessed fairly against strong classical baselines and under realistic hardware constraints. The work therefore explicitly accounts for the limitations of noisy intermediate-scale quantum (NISQ) devices, including noise, restricted connectivity, compilation overhead, and finite-shot estimation. The empirical study instantiates both quantum pipelines on a controlled supervised benchmark adapted from IBM Quantum: classifying small synthetic images containing either a vertical or a horizontal line under bounded noise. Using the same dataset, identical splitting strategies, and matched budgets for shots and optimisation, the thesis compares quantum and classical models under a consistent evaluation protocol. The results show that both quantum kernel and variational approaches achieve meaningful accuracy in this regime, but do not demonstrate a decisive quantum advantage on the selected task. The thesis concludes by outlining limitations and directions for future work, including scaling to richer benchmarks, incorporating error mitigation, and strengthening assumption-explicit methodology for credible claims of quantum-enhanced learning.

•Jan 16, 2026•0 citations
This paper proposes information-time benchmarks (BIT) as a geometry- and physics-aware way to compare learning dynamics in quantum and classical models. Building on Bures/Fisher information geometry, we define information-time as the length of an observer’s state trajectory in prediction space and show that it is constrained by observer-specific quantum speed limits and entropy continuity bounds, leading to an explicit energy–entropy–information-time trade-off. On the practical side, the paper develops implementable estimators of information-time for standard deep-learning pipelines (gradient-based and Fisher-weighted, including diagonal Fisher approximations), discusses their computational complexity and limitations, and shows how to turn them into BIT scores such as information-time per unit wall-clock time or per Joule. We position BIT relative to path length in parameter space, sharpness/SAM, gradient flow and NTK perspectives, and natural gradient methods, highlighting that BIT uniquely measures trajectory length directly in prediction space and is bounded by device-level energy variance. Finally, we formulate explicit, falsifiable hypotheses linking BIT to transfer learning, sharpness of minima, optimiser choice (e.g. SGD vs natural gradient/K-FAC), and quantum–classical comparisons, and we outline concrete experimental protocols for future validation. The goal is not to replace existing benchmarks, but to add an information-geometric and resource-aware lens on learning dynamics that can be applied across architectures and hardware platforms.
This paper introduces a notion of information-time for quantum observers, defined as the Bures length of the trajectory of an observer’s reduced density matrix in state space. In contrast to approaches where time is an external parameter or a purely global relational variable, information-time is intrinsically observer-dependent and built only from the information accessible to a given physical subsystem (no appeal to consciousness or measurement postulates is made). On this basis the work derives two formal results: (i) an observer-specific Mandelstam–Tamm–type quantum speed limit that bounds the physical time required to generate a given amount of information-time, and (ii) an energy–entropy–information-time inequality that relates the change of an observer’s von Neumann entropy to available energy, physical time and Hilbert-space dimension. Two simple models are used to illustrate the framework: a dephasing spin model from Quantum Darwinism, and a Page–Wootters–type relational clock including a “blind” observer that experiences zero information-time. The paper argues that this “HCI” (Hilbert-space Change of Information) perspective offers a useful geometric language for the experience of time by quantum subsystems and discusses potential implications for quantum computing, quantum machine learning and AI. In particular, it suggests using information-time as a basis for hardware-agnostic benchmarks of learning and computation, complementing standard measures such as wall-clock time and FLOPs.
We develop a Fisher-informational reformulation of physical time in which clock time is not regarded as a fundamental ontological substance, but as an emergent calibration of causally ordered distinguishability among physical states. The operational starting point is that clocks do not measure time itself; rather, they instantiate reproducible physical processes whose distinguishable states are correlated with other events. We introduce a causal-informational parameter, denoted by Lambda_F, defined as an accumulated Fisher-geometric distance along a causally admissible trajectory in state space. In classical statistical systems, this parameter is generated by the Fisher information metric; in quantum systems, the corresponding construction is associated with quantum Fisher information, the Bures metric, and the Fubini-Study geometry of projective Hilbert space. The manuscript distinguishes model-dependent Fisher information from quantum Fisher information, clarifies the reparameterization of Schrodinger dynamics, and gives explicit examples involving a qubit clock, an exponential decay process, and a Fisher characterization of clock quality. The proposal is positioned relative to relational time, the Page-Wootters mechanism, thermal time, quantum speed-limit relations, information geometry, and the problem of time in quantum gravity. We do not claim that relational or emergent time is new. The specific contribution is the use of Fisher distinguishability as an operational precursor from which ordinary clock time can be reconstructed. In this sense, the central statement of the paper is: time is not measured by clocks; clock time is reconstructed from the Fisher distinguishability accumulated along causally ordered physical changes.
The importance of Fisher information is increasing in nonequilibrium thermodynamics, as it has played a fundamental role in trade-off relations such as thermodynamic uncertainty relations and speed limits. In this work, we investigate temporal Fisher information, which measures the temporal information content encoded in probability distributions, for both classical and quantum systems. We establish that temporal Fisher information is bounded from above by physical costs, such as entropy production in classical Langevin and Markov processes, and the quantum entropy production and the expectation of the effective interaction Hamiltonian in open quantum systems. Conversely, temporal Fisher information is bounded from below by statistical distances (e.g., the Bhattacharyya arccos distance), leading to classical and quantum speed limits that constrain the minimal time required for state transformations. We perform numerical simulations on two quantum dot models to validate the obtained bounds. Our work provides a unified perspective on speed limits from the point of view of temporal Fisher information in both classical and quantum dynamics.
Recent advancements have revealed new links between information geometry and classical stochastic thermodynamics, particularly through the Fisher information (FI) with respect to time. Recognizing the nonuniqueness of the quantum Fisher metric in Hilbert space, we exploit the fact that any quantum Fisher information (QFI) can be decomposed into a metric-independent incoherent part and a metric-dependent coherent contribution. We demonstrate that the incoherent component of any QFI can be directly linked to entropic acceleration, and for GKSL dynamics with local detailed balance, to the rate of change of generalized thermodynamic forces and entropic flow, paralleling the classical results. Furthermore, we tighten a classical uncertainty relation between the geometric uncertainty of a path in state space and the time-averaged rate of information change and demonstrate that it also holds for quantum systems. We generalize a classical geometric bound on the entropy rate for far-from-equilibrium processes by incorporating a nonnegative quantum contribution that arises from the geometric action due to coherent dynamics. Finally, we apply an information-geometric analysis to the recently proposed quantum-thermodynamic Mpemba effect, demonstrating this framework's ability to capture thermodynamic phenomena.
We compare Ising ([Formula: see text]) and QUBO ([Formula: see text]) encodings for Boltzmann machine learning under controlled protocols that fix the sampler, optimizer, and learning-rate design within each comparison. Exploiting the identity that the Fisher information matrix (FIM) equals the covariance of sufficient statistics, we visualize empirical moments from model samples and reveal systematic, representation-dependent differences. QUBO induces larger cross terms between first- and second-order statistics, creating more small-eigenvalue directions in the FIM and lowering spectral entropy. This ill-conditioning explains slower convergence under stochastic gradient descent (SGD). In contrast, full-FIM natural gradient descent (NGD), which rescales updates by the FIM metric, achieves similar convergence across encodings, whereas diagonal-FIM approximation can reintroduce representation-dependent differences. Practically, for SGD-based training, the Ising encoding provides more isotropic curvature and faster convergence; for QUBO, centering/scaling or NGD-style preconditioning mitigates curvature pathologies. These results clarify how representation shapes information geometry and finite-time learning dynamics in Boltzmann machines and yield actionable guidelines for variable encoding and preprocessing.
This thesis presents a comparative study of quantum machine learning and classical machine learning developed from first principles. It begins by establishing the foundations of the quantum circuit model, observables, and projective measurement, and then revisits the classical learning framework with an emphasis on risk minimisation, task-aligned metrics, and rigorous evaluation protocols. Building on this shared perspective, the thesis examines two central paradigms in quantum machine learning: quantum kernels and variational quantum circuits. A unifying theme is that quantum models must be assessed fairly against strong classical baselines and under realistic hardware constraints. The work therefore explicitly accounts for the limitations of noisy intermediate-scale quantum (NISQ) devices, including noise, restricted connectivity, compilation overhead, and finite-shot estimation. The empirical study instantiates both quantum pipelines on a controlled supervised benchmark adapted from IBM Quantum: classifying small synthetic images containing either a vertical or a horizontal line under bounded noise. Using the same dataset, identical splitting strategies, and matched budgets for shots and optimisation, the thesis compares quantum and classical models under a consistent evaluation protocol. The results show that both quantum kernel and variational approaches achieve meaningful accuracy in this regime, but do not demonstrate a decisive quantum advantage on the selected task. The thesis concludes by outlining limitations and directions for future work, including scaling to richer benchmarks, incorporating error mitigation, and strengthening assumption-explicit methodology for credible claims of quantum-enhanced learning.