Pith. sign in

REVIEW 4 major objections 3 minor 28 references

Quantum machine learning models can be represented as stochastic processes, but only through a trade-off: informationally complete measurements give Markovian yet quasi-stochastic dynamics, while projective measurements give positive but hi

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 18:20 UTC pith:K7347RV6

load-bearing objection The core constructions are sound, but the advertised impossibility trade-off is oversold: it holds for the two representation families they analyze, not for the full space of POVMs their own framework admits. the 4 major comments →

arxiv 2607.17327 v2 pith:K7347RV6 submitted 2026-07-19 quant-ph cs.LG

Interpreting Quantum Learning Models via Stochastic Processes

Johannes Fankhauser , Lukas J. Fiderer , Hans J. Briegel This is my paper
classification quant-ph cs.LG MSC 81P6860J20 PACS 03.67.-a
keywords quantum machine learningstochastic processesquasi-stochastic mapsChapman-Kolmogorov divisibilityPOVM representationsProjective Simulationquantum channelsnon-Markovianity
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper sets out to show that any quantum learning model's internal evolution can be viewed as a stochastic process over a configuration space, provided one accepts one of two costs. With an informationally complete measurement (such as a SIC-POVM), the induced transition kernel is first-order Markovian but generally has negative entries, so it is only quasi-stochastic. With a projective measurement basis, the kernel is genuinely positive, but the process generally violates Chapman-Kolmogorov divisibility and must depend on past configurations. The authors argue these are the two complementary ways quantum interference shows up in any probabilistic description, and that generalized Projective Simulation gives a common language for interpreting both. A sympathetic reader would take the central contribution as a structural trade-off theorem: you cannot keep both positivity and first-order divisibility in a stochastic representation of quantum dynamics.

Core claim

Starting from any fixed POVM, a quantum channel induces a linear transition kernel on configuration probabilities via T^j_i(E) = Tr(E_i E(F_j)). For informationally complete POVMs the dynamics closes in first order and obeys the Chapman-Kolmogorov composition law, but the kernel is generally quasi-stochastic with negative entries. For projective POVMs the kernel can be chosen strictly stochastic, but the induced process is generically non-Markovian; divisibility holds if and only if the evolution of diagonal probabilities is independent of off-diagonal coherence. The paper's main claim is that this trade-off is fundamental: quantum dynamics can be represented either by Markovian quasi-stocha

What carries the argument

The central object is the transition kernel induced by a quantum channel on the probability representation defined by a fixed POVM, T^j_i(E) := Tr(E_i E(F_j)), where {F_j} is a dual frame. For IC-POVMs this kernel satisfies the Markovian composition law T(E_2 ∘ E_1) = T(E_2) T(E_1), yielding divisibility but allowing negative entries; for SIC-POVMs it takes the explicit form T^j_i(E) = d(d+1) Tr(E_i E(E_j)) - 1/d. For projective POVMs the kernel is positive but only closes at higher order, with divisibility characterized by Eq. (18), requiring that the diagonal probabilities evolve independently of off-diagonal components.

Load-bearing premise

The paper asserts a global impossibility trade-off but proves it only for two extremal families of representations, informationally complete POVMs and projective POVMs; the load-bearing premise is that no intermediate or enlarged POVM representation can achieve both positivity and first-order divisibility simultaneously.

What would settle it

Construct a POVM that is neither informationally complete nor projective, and a non-classical quantum channel, such that the induced transition kernel from Eq. (5) is entirely non-negative and satisfies Eq. (7) for composition. One concrete route: numerically search all three-outcome qubit POVMs for a channel where the kernel is stochastic and the composition law holds; finding even one such pair would refute the claim that positivity and first-order divisibility cannot coexist in any representation.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Any quantum channel can be represented as a linear, Markovian, yet quasi-stochastic map on the probability simplex if one chooses an informationally complete POVM, and the resulting configuration probabilities never become negative.
  • Any projective representation yields a positive stochastic process, but in general the Markov order must reach back to the initial configuration, reproducing the kind of indivisible stochastic dynamics that has been proposed for quantum systems.
  • In stabilizer quantum mechanics, the projective representation becomes a positive stochastic process of finite Markov order, at most 2N for an N-qudit system, giving a concrete example of bounded memory.
  • Classical Projective Simulation is recovered exactly when the admissible state space is the full simplex, the kernel is stochastic, and the process is Markovian and time-homogeneous.
  • Finite-order positive stochastic approximations of quantum dynamics may exist when the relevant eigenphases are rational or approximately periodic, since the induced probability sequences then satisfy finite recurrence relations.

Where Pith is reading between the lines

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

  • If the trade-off is genuinely structural, then 'interpretability' of a quantum learning model is not a fixed property but a representational choice: one can buy trajectory-based explanation by sacrificing Markov order, or buy Markov order by sacrificing positive transition weights. Model designers would need to decide which failure mode is less harmful for a given application.
  • The Markov order required by a positive projective representation could be read as a resource measure for quantum computation, providing a concrete bridge to algorithmic complexity: stabilizer circuits would be 'cheap' (bounded memory) while universal circuits would need unbounded history dependence.
  • A natural extension is to test whether intermediate POVMs that are neither informationally complete nor projective can break the trade-off; the paper's own construction leaves that case open, and a numerical scan over low-dimensional POVMs would settle whether the claimed impossibility holds beyond the two extremal families.
  • The finite-order approximation idea suggests a practical recipe: given a quantum circuit, diagonalize its unitary to find rational approximations of eigenphases, then build a hidden-Markov-style stochastic model that approximates the output distribution; this could turn quantum-circuit output statistics into classical stochastic simulations with bounded memory.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 3 minor

Summary. The paper introduces a framework for representing quantum channels as transition kernels over configuration spaces defined by a fixed POVM. For informationally complete (IC) and SIC POVMs, the induced kernel is first-order Markovian but generally quasi-stochastic (Eqs. (4)–(9)). For projective POVMs, the kernel is positive and stochastic but generally non-Markovian, with a divisibility condition given in Eq. (18). The authors interpret this as a trade-off between negativity and memory, apply it to a generalized Projective Simulation framework, and discuss finite-history approximations, including a stabilizer example with finite Markov order.

Significance. If the trade-off were established as a structural result, the paper would provide a useful conceptual bridge between quantum dynamics, quasi-probability representations, and stochastic-process models of learning. The IC-POVM formalism is correct and clearly presented, the distinction between Markovian quasi-stochastic and positive non-Markovian descriptions is valuable, and the stabilizer example (Eq. (25)) is a concrete and interesting positive result. However, the central dichotomy/no-go claim is not proven over the full space of representations, and one of the explicit SIC formulas is incorrect for non-unital channels. The paper's main interpretive claim therefore needs substantial revision before it can be accepted.

major comments (4)
  1. [Sec. VI / Abstract] The paper's central claim—'impossibility of simultaneously maintaining positivity and first-order divisibility' (Sec. VI) and the abstract's 'either ... or ...'—is asserted, not derived. The analysis treats two corners: IC/SIC-POVMs (Markovian, quasi-stochastic) and projective POVMs (positive, non-Markovian). Intermediate POVMs, which the framework itself admits in Sec. III, are not analyzed; their dual frame need not be positive and the one-time probabilities need not close, so a third class (quasi-stochastic and non-Markovian) is possible. Moreover, dephasing channels are positive and first-order divisible in the projective representation (Eq. (18) is satisfied), so the impossibility claim needs a precise generic-scope formulation or a no-go proof.
  2. [Sec. III.B, Eq. (9)] Eq. (9) is incorrect for non-unital channels. With F_j = d(d+1)E_j - 1, T_j^i = Tr(E_i Φ(F_j)) = d(d+1)Tr(E_iΦ(E_j)) - Tr(E_iΦ(1)). The constant term is 1/d only if Φ(1)=1. Since the channel is introduced as arbitrary CPTP in Sec. III, either restrict to unital channels or keep the general term. This affects the SIC example and any conclusions drawn from the explicit form.
  3. [Sec. III.C, Eq. (14)] The proposed higher-order kernel for the projective representation, T_{c_n}^{c_{n-1},...,c_1}=Tr(P_{c_n}E(P_{c_1})), is not a conditional transition kernel: it drops the intermediate outcomes and is not obtained from the conditional probability Pr(c_n | c_{n-1},...,c_1). A stochastic process requires joint path probabilities; forward probabilities Pr(c_n|c_1) alone do not define a path measure. The stabilizer example (Eq. (25)) is a genuine second-order kernel, but it does not support the general claim that arbitrary projective representations are positive stochastic processes of finite/higher Markov order.
  4. [Sec. III, Eqs. (4)–(5)] Eqs. (4)–(5) require ρ_n = Σ_j p_n^j F_j. This reconstruction is valid only when ρ_n∈S = span{E_i}. For non-spanning POVMs (e.g. projective POVMs) an arbitrary state evolving under a channel leaves S, so the one-time probabilities p_n do not in general determine p_{n+1}. The text's claim that 'Non-spanning POVMs ... induce effective non-Markovian stochastic processes' should be stated with this caveat: for states/channels that keep the state in S, the induced evolution is first-order; non-Markovianity arises precisely from the loss of the reconstruction.
minor comments (3)
  1. [Sec. III.C, after Eq. (14)] 'Markov order is maximal—reaching back to the initial configuration. But there is no dependency on intermediate steps' is contradictory: Eq. (13) defines an L-th order kernel depending on the last L outcomes, while Eq. (14) depends only on the first outcome. Please clarify the notion of Markov order used.
  2. [Table I] For projective representations, the Markov order should be 'L>1 or unbounded' rather than just 'L>1'; Sec. IV.A describes an unbounded-order example.
  3. [Sec. IV.A, Eqs. (20)–(22)] The index convention in |U(Δt)_{10}|² is confusing (basis states are 0 and 1); also superscript/subscript order is inconsistent with Eq. (13).

Circularity Check

0 steps flagged

No significant circularity: the central trade-off is derived from stated POVM/dual-frame definitions; the Sec. VI scope overclaim is a completeness gap, not a circular step.

full rationale

The paper's derivation chain is self-contained linear algebra. The IC-POVM Markovianity follows from informational completeness via the dual-frame expansion (Eq. (8)) and the Chapman–Kolmogorov product (Eq. (7)); the quasi-stochasticity follows from the dual frame not being positive in Eq. (5). The projective representation's positivity follows from Tr(P_i E(P_j)) ≥ 0 for CPTP E and positive P_i, P_j (Eq. (11)), while its generic non-Markovianity follows from non-closure under non-informationally-complete configurations (Eqs. (13)–(18)). No parameter is fitted and no target result is assumed in these derivations. The SIC-POVM formula (Eq. (9)) uses the standard canonical dual, not a smuggled ansatz. Self-citations are confined mainly to the Projective Simulation application section, which is explicitly presented as one possible interpretation rather than a load-bearing premise ('the following construction is not intended as a unique quantum generalization of the PS framework, but should rather be understood as one possible extension'). The stabilizer finite-Markov-order example is derived from the Cayley–Hamilton theorem, not from prior PS results. The strongest legitimate concern is that Sec. VI's phrase 'the impossibility of simultaneously maintaining positivity and first-order divisibility' is asserted more broadly than the two constructed cases prove; non-IC, non-projective POVMs are not analyzed. That is a scope/completeness objection, not a circularity, because no equation in the paper defines its own conclusion or renames a fitted quantity as a prediction.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

No free parameters: the framework contains no fitted or hand-tuned numbers; example parameters (ω, Δt, S ∈ Sp(2,Z_d)) are part of the specified dynamics, not fitted. No invented entities: the paper postulates no new physical objects; the history vector η_n and admissible set K are mathematical devices. Main burdens: the frame/dual-frame reconstruction axiom, IC/SIC existence, the coarse-graining-dependence of Markov order, and the Sec. VI.d approximation leap.

axioms (5)
  • standard math The POVM-frame representation: for a fixed spanning POVM {E_i} with dual frame {F_j}, T^i_j(E) = Tr(E_i E(F_j)) correctly gives the induced linear dynamics on configuration probabilities.
    Eqs. (4)-(6); standard frame/QBism machinery, requires the reconstruction X = Σ_j Tr(X E_j) F_j on the span of the POVM. Cited context: Fuchs-Schack [8], Ferrie [5].
  • domain assumption Informational completeness implies span{E_i} is the full operator space, so the expansion E_1(F_j) = Σ_k T^k_j(E_1) F_k used in the Chapman-Kolmogorov proof (Eq. (7)) is valid; SIC-POVMs exist in the dimensions used.
    Definition of IC-POVM, Sec. III.A. SIC-POVM existence in all dimensions is a famous open conjecture; the paper's explicit examples use d=2 and prime dimensions where SICs are known, but the general SIC formula (Eq. (9)) inherits the assumption.
  • standard math Quantum channels are CPTP and Born probabilities Pr(c=i) = Tr(ρ E_i) define the configuration probabilities; no measurement is actually performed.
    Sec. III opening paragraph; standard quantum mechanics.
  • domain assumption For the projective representation, non-Markovianity is assessed for a fixed PVM and discrete-time coarse-graining; Markov order is taken relative to the chosen discretization, which the paper acknowledges is not an intrinsic property.
    Sec. III.C, including the caveat that 'the Markov order of a stochastic representation ... is not an intrinsic property of the underlying physical process.'
  • ad hoc to paper The spectral observation of Sec. VI.d — probability sequences q_ν(n) = Σ_{r,s} A^{(ν)}_{rs} e^{in(θ_r−θ_s)} satisfy finite recurrences when eigenphase differences are rational — is assumed to indicate the existence of finite-order positive stochastic kernels approximating quantum dynamics.
    Sec. VI.d; no construction, no error bound, and the paragraph immediately after labels the compression question 'an interesting open question.' This is the main speculative leap.

pith-pipeline@v1.3.0-alltime-deepseek · 16024 in / 28300 out tokens · 271248 ms · 2026-08-01T18:20:34.072743+00:00 · methodology

0 comments
read the original abstract

Quantum machine learning models define probabilistic input--output maps through coherent quantum evolution and measurement. While such models can exhibit computational advantages, their internal functioning and decision making generally resists interpretation in terms of stochastic trajectories through intermediate configurations. In contrast to classical (Markovian) stochastic processes, quantum dynamics generically violates the Chapman--Kolmogorov divisibility condition, preventing a decomposition into probabilistically meaningful intermediate transitions. We develop a probabilistic framework for representing quantum learning models as stochastic processes over configuration spaces where the dynamics are modeled as linear maps on probability distributions. Starting from a fixed POVM, arbitrary quantum channels induce transition kernels on the associated probability representation. For informationally complete POVMs, and in particular SIC-POVMs, these kernels are Markovian but generally quasi-stochastic, with non-classicality appearing as negativity. By contrast, projective spaces admit positive stochastic kernels but generally require non-Markovian dynamics due to the failure of Chapman--Kolmogorov divisibility. This yields a trade-off between negativity and dependence on past configurations, i.e. quantum dynamics can be represented either by Markovian quasi-stochastic maps or by positive stochastic processes with higher Markov order. We discuss how such representations of quantum dynamics can be interpreted as stochastic walks through a memory space in the spirit of Projective Simulation, a model of learning and agency in which decisions arise from random walks over an episodic memory network. We further outline how finite-order stochastic kernels can approximate such quantum deliberation processes and show in what regimes the classical machine learning model is recovered.

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

28 extracted references · 6 linked inside Pith

  1. [1]

    Biamonte, P

    J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, and S. Lloyd, Quantum machine learning, Na- ture549, 195 (2017)

    2017

  2. [2]

    Schuld, A

    M. Schuld, A. Bocharov, K. M. Svore, and N. Wiebe, Circuit-centric quantum classifiers, Phys. Rev. A101, 032308 (2020)

    2020

  3. [3]

    Havl ´ ıˇ cek, A

    V. Havl ´ ıˇ cek, A. D. C´ orcoles, K. Temme, A. W. Harrow, A. Kandala, J. M. Chow, and J. M. Gambetta, Super- vised learning with quantum-enhanced feature spaces, Nature567, 209 (2019)

    2019

  4. [4]

    C. W. Gardiner,Stochastic Methods(Springer, 2009)

    2009

  5. [5]

    C. Ferrie, Quasi-probability representations of quantum theory with applications to quantum information sci- ence, Reports on Progress in Physics74, 116001 (2011), arXiv:1101.5323 [quant-ph]

    Pith/arXiv arXiv 2011

  6. [6]

    R. L. Hudson, When is the wigner quasi-probability den- sity non-negative?, Reports on Mathematical Physics6, 249 (1974)

    1974

  7. [7]

    S. D. Bartlett, T. Rudolph, and R. W. Spekkens, Recon- struction of gaussian quantum mechanics from liouville mechanics with an epistemic restriction, Physical Review A86, 012103 (2012), arXiv:1111.5057 [quant-ph]

    Pith/arXiv arXiv 2012

  8. [8]

    C. A. Fuchs and R. Schack, Quantum-bayesian coher- ence, Rev. Mod. Phys.85, 1693 (2013)

    2013

  9. [9]

    C. A. Fuchs, Qbism: The perimeter of quantum bayesian- ism, arXiv preprint arXiv:1003.5209 (2010)

    Pith/arXiv arXiv 2010

  10. [10]

    J. A. Barandes, Quantum systems as indivisible stochas- tic processes (2025), arXiv:2507.21192 [quant-ph]

    Pith/arXiv arXiv 2025

  11. [11]

    J. A. Barandes, The stochastic-quantum correspondence, Philosophy of Physics3, 10.31389/pop.186 (2025)

    doi:10.31389/pop.186 2025

  12. [12]

    H. J. Briegel and G. D. las Cuevas, Projective simulation for artificial intelligence, Scientific Reports2, 400 (2012)

    2012

  13. [13]

    H. J. Briegel and T. M¨ uller,Projective Simulation in Ac- tion: Quantum-Mechanical Perspectives on the Problem of Agency(Springer Nature, 2025)

    2025

  14. [14]

    Mautner, A

    J. Mautner, A. Makmal, D. Manzano, M. Tiersch, and H. J. Briegel, Projective Simulation for Classical Learn- ing Agents: A Comprehensive Investigation, New Gener- ation Computing33, 69 (2015)

    2015

  15. [15]

    Pazem, M

    J. Pazem, M. Krumm, A. Q. Vining, L. J. Fiderer, and H. J. Briegel, Free Energy Projective Simulation (FEPS): Active inference with interpretability, PLoS One 20, e0331047 (2025)

    2025

  16. [16]

    B. Eva, K. Ried, T. Mueller, and H. J. Briegel, How a Minimal Learning Agent can Infer the Existence of Unob- served Variables in a Complex Environment, Minds and Machines33, 185 (2023)

    2023

  17. [17]

    Tiersch, E

    M. Tiersch, E. J. Ganahl, and H. J. Briegel, Adaptive quantum computation in changing environments using projective simulation, Scientific reports5, 12874 (2015)

    2015

  18. [18]

    Walln¨ ofer, A

    J. Walln¨ ofer, A. A. Melnikov, W. D¨ ur, and H. J. Briegel, Machine Learning for Long-Distance Quantum Commu- nication, PRX quantum1, 010301 (2020)

    2020

  19. [19]

    A. A. Melnikov, H. Poulsen Nautrup, M. Krenn, V. Dun- jko, M. Tiersch, A. Zeilinger, and H. J. Briegel, Active learning machine learns to create new quantum experi- ments, Proceedings of the National Academy of Sciences 115, 1221 (2018)

    2018

  20. [20]

    S. W. Al-Safi and A. J. Short, Simulating all Nonsignaling Correlations via Classical or Quantum Theory with Neg- ative Probabilities, Physical Review Letters111, 170403 (2013)

    2013

  21. [21]

    Kaszlikowski and P

    D. Kaszlikowski and P. Kurzy´ nski, A Little Bit of Classi- cal Magic to Achieve (Super-) Quantum Speedup, Foun- dations of Physics51, 55 (2021)

    2021

  22. [22]

    Morris, L

    B. Morris, L. J. Fiderer, B. Lang, and D. Goldwater, Witnessing Bell violations through probabilistic negativ- ity, Physical Review A105, 032202 (2022)

    2022

  23. [23]

    Onggadinata, A

    K. Onggadinata, A. Tanggara, M. Gu, and D. Kasz- likowski, Ideal stochastic process modeling with post- quantum quasiprobabilistic theories, Quantum10, 2005 (2026)

    2005

  24. [24]

    Chiribella and R

    G. Chiribella and R. W. Spekkens, eds.,Quantum The- ory: Informational Foundations and Foils, Fundamental Theories of Physics, Vol. 181 (Springer, Dordrecht, 2016)

    2016

  25. [25]

    Hardy, Quantum theory from five reasonable axioms, arXiv preprint (2001), arXiv:quant-ph/0101012

    L. Hardy, Quantum theory from five reasonable axioms, arXiv preprint (2001), arXiv:quant-ph/0101012

    Pith/arXiv arXiv 2001

  26. [26]

    Barrett, Information processing in generalized proba- bilistic theories, Physical Review A75, 032304 (2007), arXiv:quant-ph/0508211

    J. Barrett, Information processing in generalized proba- bilistic theories, Physical Review A75, 032304 (2007), arXiv:quant-ph/0508211

    Pith/arXiv arXiv 2007

  27. [27]

    Fankhauser and W

    J. Fankhauser and W. C. Myrvold, Indivisible quantum theory and relativity (2026), manuscript in preparation

    2026

  28. [28]

    Doukas, On the emergence of quantum mechanics from stochastic processes (2026), arXiv:2602.22095 [quant-ph]

    J. Doukas, On the emergence of quantum mechanics from stochastic processes (2026), arXiv:2602.22095 [quant-ph]

    arXiv 2026