Pith. sign in

REVIEW 3 major objections 7 minor 37 references

One classical bit exactly simulates continuous isotropic qubit measurements and closely approximates finite symmetric ones.

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 →

T0 review · grok-4.5

2026-07-30 16:45 UTC pith:VS63FUKJ

load-bearing objection Clean exact 1-bit result for the continuous isotropic qubit POVM, plus a usable analytical protocol; finite “extremely accurate” claims rest on uncertified numerics. the 3 major comments →

arxiv 2607.23645 v1 pith:VS63FUKJ submitted 2026-07-26 quant-ph cs.AIcs.NAmath.NA

Neural Network Learning of One-Bit Protocols for Qubit Measurement Simulation

classification quant-ph cs.AIcs.NAmath.NA
keywords qubit prepare-and-measurecommunication complexityone-bit simulationPOVMsymmetric measurementsneural network protocol discoveryBloch sphereclassical simulation
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.

Two classical bits are necessary and sufficient to simulate every qubit prepare-and-measure experiment exactly, but that worst-case cost need not apply to every restricted family of measurements. This paper uses neural networks as a search tool and finds that equal-weight, highly symmetric measurements—especially regular polyhedra on the Bloch sphere—can be reproduced to high average accuracy with only one bit of communication plus shared randomness. From the patterns the networks learn, the authors extract an explicit analytical protocol: Alice sends a single sign bit about a shared random direction, Bob flips that direction accordingly, and samples outcomes by positive overlap. The protocol is extremely accurate for finite informationally complete symmetric configurations, with error shrinking as the measurement becomes more isotropic, and becomes exact for the continuous isotropic measurement. The result shows that symmetry can sharply reduce the classical resources needed to reproduce quantum statistics.

Core claim

A single bit of classical communication plus shared randomness exactly reproduces the statistics of the continuous isotropic (covariant) qubit POVM, and yields highly accurate approximations for finite equal-trace rank-one informationally complete symmetric POVMs such as regular polyhedra, with the discrepancy decreasing as the configuration becomes more isotropic.

What carries the argument

The analytical one-bit protocol: Alice sends c = H(λ · r); Bob forms the flipped shared vector λ' = (2c − 1)λ and reports outcome i with probability proportional to (y_i · λ') H(y_i · λ'). Distilled from neural-network response maps, this rule is the object that carries both the numerical accuracy and the exact continuous-limit proof.

Load-bearing premise

For finite polyhedral measurements, the claim of extreme accuracy rests mainly on average errors over random states and a numerical worst-case search, not a tight analytical bound that every input state is close.

What would settle it

Maximize the L1 distance between the one-bit protocol probabilities and the Born-rule probabilities over all pure states for a fixed regular polyhedron POVM; if that worst-case distance stays large (order 0.1) instead of falling toward zero as the number of outcomes grows, the accuracy claim is false.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • The continuous isotropic qubit measurement is exactly one-bit simulable in the prepare-and-measure scenario.
  • Equal-weight three-dimensional symmetric POVMs require substantially less classical communication than arbitrary qubit measurements.
  • Neural-network searches can surface analytical classical protocols that can then be proved by hand.
  • The two-bit worst-case cost for qubits is not representative of all measurement families.
  • For spherical t-designs the L1 error of the protocol is analytically bounded by O(1/√t).

Where Pith is reading between the lines

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

  • The same neural-to-analytical pipeline could test whether other ‘most quantum’ candidates, such as SIC-POVMs, also admit cheap classical simulations.
  • Isotropy of the measurement frame appears to be the structural feature that collapses communication from two bits to one; asymmetric or planar configurations should remain harder.
  • Device-certification or dimension-witness tasks that already use symmetric POVMs could lower their classical simulation overhead by adopting this protocol.
  • Higher-dimensional analogues may exist once neural searches are run on highly symmetric measurements beyond qubits.

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

3 major / 7 minor

Summary. The authors study whether qubit prepare-and-measure statistics for restricted measurement families can be simulated with one bit of classical communication (plus shared randomness), below the 2-bit worst-case barrier of Renner et al. [10]. A neural network trained per-POVM to reproduce Born-rule statistics under the 1-bit constraint reveals that equal-trace rank-one POVMs (e-POVMs) with isotropic Bloch-vector configurations perform best. From the ensemble-averaged network response maps, the authors distill an explicit analytical protocol: Alice sends c = H(λ·r), Bob flips λ when c = 0 and outputs i with probability proportional to (y_i·λ′)H(y_i·λ′) (Eq. 17). They show this protocol is exact in the continuous isotropic limit (Sec. V.C), exact for regular polyhedral eIC-POVMs when the state aligns with a measurement direction (App. C), and obeys an O(1/√t) L1 upper bound for spherical t-designs (App. D). For finite polyhedral configurations they report Monte Carlo average KLD/MAE (Figs. 7–8) and a numerically obtained worst-case L1 search (Table 1) showing errors at the 10^{-3} level.

Significance. If it holds, the result is a genuinely interesting addition to the qubit communication-complexity literature: it identifies the continuous covariant qubit POVM as an exactly 1-bit-simulable resource, a clean structural statement below the 2-bit worst-case barrier, and it provides a concrete, parameter-free analytical protocol rather than a black-box fit — the NN is used only as a discovery tool, and the extracted rule (17) contains no fitted parameters. The paper ships several rigorous, checkable components: the continuous-limit exactness proof (Sec. V.C, Eqs. 20–22), an exactness theorem for aligned states on vertex-transitive configurations (App. C), and a fully explicit asymptotic L1 bound for t-designs (App. D), plus falsifiable quantitative predictions in Table 1 that others can reproduce. The NN-to-analytics workflow is also a useful methodological example. I verified the main analytical steps independently: the denominator integral in (21) equals 1/4; the implicit constancy of ∫H(r·λ)|m·λ|dλ in m (needed for the two-parameter affine form in Sec. V.C) is correct by the identity ∫H(a·λ)|b·λ|dΩ = π; and the symmetry arguments in App. C (vertex transitivity for I1, stabilizer ro

major comments (3)
  1. [Sec. V.C, Table 1] Table 1 (Sec. V.C): the quantitative content of the headline claim that the protocol is 'extremely accurate' for finite eIC-POVMs rests entirely on a numerical worst-case L1 maximization over the Bloch sphere whose methodology is nowhere described — no grid resolution, optimizer, number of restarts, convergence criteria, or certification strategy. A missed maximizer would understate the worst case and directly inflate the central claim. Relatedly, the reported values are not monotonic in m: the octahedron (m=6) shows L1=0.0159, worse than the tetrahedron (m=4, 0.0093) and nearly 6× the dodecahedron (m=20, 0.0029). Either this is a genuine feature (in which case the statements in Sec. V.B, Sec. V.C, and the abstract that the error 'decreases as the number of outcomes increases' / 'decreasing behavior ... with increasing number of outcomes' need to be weakened to an asymptotic or trend sta
  2. [Sec. V.C, Appendix D] Appendix D vs. the configurations actually studied: the O(1/√t) bound (D28) is presented in Sec. V.C as one of 'two ways' of bridging the finite and continuous results, but it is vacuous for every configuration in Table 1. The bound is non-vacuous only when the geometric-series condition behind (D19) holds, which under the authors' own estimate (D25) requires roughly floor(t/2) > 16/π, i.e., t ≳ 12; even using the actual coefficient sums rather than the bound, one needs t ≥ 6. The tetrahedron (3-design), octahedron and cube (3-designs), icosahedron, dodecahedron, and icosidodecahedron (5-designs) all fall short. The asymptotic statement t→∞ is correct as mathematics, but as written Sec. V.C implies the bound supports the finite polyhedral claims, which it does not. The text should state explicitly that Appendix D applies only to sufficiently strong designs and does not cover any configur
  3. [Abstract; Sec. V.B; Sec. VI] Wording of the finite-configuration claim (abstract, Sec. V.B, Sec. VI): 'extremely accurate for finite informationally complete symmetric configurations' and 'the approximation improves as the number of outcomes increases' are stronger than what is shown. The rigorous results cover (a) the continuous limit, (b) aligned states r = y_i (App. C), and (c) t-designs with large t; the general finite case is supported by average-case Monte Carlo (Figs. 7–8, M = 20–50 states) and the uncertified worst-case search of Table 1. Given the known O(10^{-3}) counterexample already acknowledged in Sec. V.C (octahedron, diagonal state in the x-z plane), the claims should be recalibrated to state precisely what is proven versus numerically evidenced. This is fixable with wording changes once points 1–2 are addressed.
minor comments (7)
  1. [Appendix D] Appendix D, Eq. (D15): the strict inequality as printed, |(2/m)Σ|y_j·λ′| − 1| < 1, appears to be a typo — the quantity being bounded is (4/m)|R_t|, i.e., |(2/m)Σ|y_j·λ′| − 1/2 · 2|? Please recheck: the correct statement used later is (4/m)|R_t(λ′)| < 1, i.e., |(2/m)Σ_j|y_j·λ′| − 1| < 1. The printed form should be verified for consistency with (D7) and (D13). Also in (D5) and (D7) the summand is written with index i while the sum runs over j.
  2. [Appendix D, Eq. (D24)] Eq. (D24): the separate bound |c_2| ≤ 5/(4√π) is looser than the true value |c_2| = 5/8 from (D10); harmless for the asymptotic claim, but worth correcting for correctness of the displayed chain.
  3. [Sec. V.C] Sec. V.C, after Eq. (22): the step that ∫H(r·λ)|m·λ|dλ is a number independent of m deserves one line of justification (it follows from the known identity ∫H(a·λ)|b·λ|dΩ = π, or by rotational invariance), since the affine form P = γ1 + γ2 r·m with γ1, γ2 fixed by two boundary points is the crux of the exactness proof.
  4. [Sec. V.C] Sec. V.C, text after Eq. (22): 'Note that MAE is the average L1 distance over batch of inputs' is imprecise — MAE as defined in (12) is L1/(m) averaged over states. Please align the terminology.
  5. [Throughout] Numerous typos/grammar: 'scehnario' (Sec. II), 'spceific' (Sec. II), 'ensample'/'ensamble' (Sec. V.A), 'mesacecto'/'share randomness' (Sec. II, protocol description), 'and increase with the positive overlap' (Sec. V.A, subject-verb), 'state aligns with one the measurement outcomes' (App. C heading), Table 1 caption comma splice ('The order of magnitude of error, decreases by increasing'). A full proofread pass is warranted.
  6. [Figs. 3, 5, 10] Fig. 3 caption/text: 'MAE obtained in 100 experiments over 100 different randomly generated POVMs' is redundant — clarify whether it is one training run per POVM. Fig. 10 axis labels render as '/uni00000031' etc., indicating a font/encoding problem in the figure file. In Fig. 5, the sentence ending 'The set of regular polygons and polyhedra inscribed in the Bloch sphere' is truncated before the footnote.
  7. [Appendix B] Appendix B: the random-POVM weight search states that L-BFGS-B enforces that 'no weight became too small', but the tolerance/threshold is not given; since the weight-uniformity trend in Fig. 4 is a key piece of evidence, the sampling acceptance criterion should be stated.

Circularity Check

0 steps flagged

No significant circularity: NN only suggests form; analytical protocol and continuous-limit exactness are independently derived and checked against Born rule.

full rationale

The paper’s load-bearing claims do not reduce to their inputs by construction. The neural network is used only as a discovery probe (Sec. III–IV): after inspecting ensemble-averaged response maps, the authors write down an explicit 1-bit rule (15)–(17) with no free parameters fitted to target probabilities. That rule is then validated externally against Born-rule probabilities via MAE/KLD and sampling baselines (Figs. 7–8, Table 1). Continuous-limit exactness (Sec. V.C) is a self-contained integral argument: the denominator collapses by isotropy to 1/4, the response is shown affine in r·m, and the two constants are fixed by evaluating at m=±r, recovering PQ=1+r·m. Appendix C proves exactness when r equals a vertex by symmetry (vertex-transitive I1=1/m and stabilizer argument for I2), again without fitting. Appendix D’s O(1/√t) L1 bound is an independent (if sometimes loose) analytic estimate. Background citations to the 2-bit Renner protocol and related simulation work supply context, not a uniqueness theorem or ansatz that forces the 1-bit result. There is no self-definitional loop, no fitted-input-called-prediction, and no load-bearing self-citation chain. Soft spots (numerical worst-case search methodology; App. D vacuous for small t) are correctness/certification issues, not circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 1 invented entities

Load-bearing content is standard qubit PM formalism plus the Renner et al. two-bit baseline; the new protocol is a construction, not a fitted law. NN hyperparameters affect only the discovery phase. No new physical entities are postulated. Main extra assumptions are equal-weight rank-one (e-POVM) structure for the analytical rule and Haar-shared randomness with a Heaviside one-bit message—the same communication model as prior simulation work.

free parameters (2)
  • NN hidden widths and training hyperparameters (10/16 units, Adam lr=0.001, 100 epochs, M=2000, N=4000) = 10 and 16 hidden units; lr=0.001; M=2000; N=4000
    Chosen for training stability and cost; they influence discovered response maps but not the final analytical protocol once extracted.
  • Monte Carlo sample sizes for reported MAE/KLD/Table 1 = Typically N up to 1e5–1e7 for KLD curves; M=20–50 states for means
    Finite N/M set the numerical error bars and plateau estimates; analytical claims do not depend on specific fitted constants.
axioms (5)
  • domain assumption Born-rule qubit probabilities P(i)=p_i(1+y_i·r) with POVM completeness Σ M_i = I
    Standard quantum measurement theory used throughout Sec. II and all comparisons.
  • domain assumption Worst-case exact classical simulation of arbitrary qubit PM statistics requires two bits; two bits suffice (Renner et al. 2023)
    Motivates the restricted-family question; recalled in Sec. II and used as baseline in Appendix A Fig. 10.
  • domain assumption Shared randomness is a uniform (Haar) unit vector on S^2; Alice’s message is the Heaviside bit c=H(λ·r)
    Communication model fixed in Sec. III and in the analytical protocol (15)–(17); standard in this literature.
  • ad hoc to paper For the analytical 1-bit rule it is enough to take equal weights p_i=1/m and rank-one effects (e-POVMs / unit-norm tight frames)
    Motivated by NN performance vs weight Std (Fig. 4) and then hard-wired into protocol step 1 and Eq. (17); general unequal weights are not covered by the closed form.
  • standard math Spherical t-design moment-matching and Legendre expansion of |x| for the L1 scaling proof
    Appendix D uses standard design and special-function facts to bound the remainder R_t.
invented entities (1)
  • Analytical 1-bit PM response rule P_1-bit(i|λ',Y)=(y_i·λ')H(y_i·λ')/Σ_j(y_j·λ')H(y_j·λ') with λ'=(2c−1)λ independent evidence
    purpose: Explicit classical simulation strategy distilled from NN response maps for e-POVMs
    Construction rather than a new physical object; independent checks are the continuous-limit proof and finite polyhedra numerics inside the paper.

pith-pipeline@v1.2.0-grok45-kimik3 · 23161 in / 3409 out tokens · 70498 ms · 2026-07-30T16:45:05.633796+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Neural Network Learning of One-Bit Protocols for Qubit Measurement Simulation." pith.science (2026). https://pith.science/paper/VS63FUKJ

@misc{pith2026260723645,
  author       = {Pith},
  title        = {Pith review of: Neural Network Learning of One-Bit Protocols for Qubit Measurement Simulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VS63FUKJ}},
  note         = {Machine review of arXiv:2607.23645}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Communication complexity provides a natural framework for quantifying the classical resources required to reproduce quantum statistics. In the qubit prepare-and-measure scenario, two classical bits have been shown to be necessary and sufficient to simulate arbitrary qubit states and arbi- trary quantum measurements exactly. However, this result does not exclude the possibility that restricted families of measurements may admit accurate 1-bit classical approximations. We use a neural network procedure to demonstrate that a single bit can achieve high average accuracy for specific measurement families. A performance analysis of our neural network reveals that symmet- ric measurements with uniformly weighted elements, such as those forming regular polyhedra, are particularly amenable to this restricted communication. By analyzing the patterns learned by the neural network, we derive an analytical protocol that is extremely accurate for finite information- ally complete symmetric configurations and becomes exact in the limit of a continuous isotropic measurement.

Figures

Figures reproduced from arXiv: 2607.23645 by Estel Ferrer, Gael Sent\'is, Giulio Gasbarri, Josep Escrig, Mani Zartab, Ramon Mu\~noz-Tapia.

Figure 1
Figure 1. Figure 1: Quantum PM scenario. Alice prepares the qubit [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: 2-bit classical simulation protocol of Ref. [10] for [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: MAE of the probabilities calculated with the NN [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: MAE of the NN-procedure (dark gray) and of the [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: MAE of the probabilities obtained from the NN [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Conditional response probabilities for outcome [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: KLD of the analytical 1-bit protocol and of di [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Mean KLD for the analytical protocol, the NN [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Loss as a function of the number of epochs for the training and validation datasets. [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: MAE as a function of the batch size N for different sample sizes M, compared with the MAE of the protocol of Renner et al. [10]. Appendix B: Generation of random qubit POVMs Random rank-one qubit POVMs are generated by sets of random points, ⃗y in the unit sphere S 2 . These can be simply obtained from sampling two independent random variables u, v ∼ U(0, 1) and defining ϕ = 2πu and θ = arccos(1 − 2v). Th… view at source ↗
Figure 11
Figure 11. Figure 11: Conditional response maps for the six outcomes of the six-direction Cartesian (octahedral) eIC-POVM, conditioned [PITH_FULL_IMAGE:figures/full_fig_p017_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Conditional response maps for the six outcomes of the six-direction Cartesian (octahedral) eIC-POVM, conditioned [PITH_FULL_IMAGE:figures/full_fig_p018_12.png] view at source ↗

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

37 extracted references · 3 linked inside Pith

  1. [1]

    In addition, they share two random unit vectors ⃗λ1, ⃗λ2 ∈S 2

    Alice is given the Bloch vector⃗ r, and Bob is given the POVM specified byY. In addition, they share two random unit vectors ⃗λ1, ⃗λ2 ∈S 2

  2. [2]

    Alice computes the classical bits c1 =H( ⃗λ1 ·⃗ r), c 2 =H( ⃗λ2 ·⃗ r),(3) and sends them to Bob, whereH(x) is the Heavi- side step function

  3. [3]

    Bob defines the effective hidden variables ⃗λ′ i = (2ci −1) ⃗λi,(4) that is, he flips ⃗λi wheneverc i = 0

  4. [4]

    He then sets ⃗λ′ = ⃗λ′ 1 if| ⃗λ′ 1 ·⃗ yi| ≥ |⃗λ′ 2 ·⃗ yi|, and ⃗λ′ = ⃗λ2 otherwise

    Bob selects⃗ yi with probabilityp i. He then sets ⃗λ′ = ⃗λ′ 1 if| ⃗λ′ 1 ·⃗ yi| ≥ |⃗λ′ 2 ·⃗ yi|, and ⃗λ′ = ⃗λ2 otherwise

  5. [5]

    Bob reports the outcomeiwith probability PB(i|⃗λ′, Y) = pi(⃗ yi · ⃗λ′)H(⃗ yi · ⃗λ′)Pm j=1 pj(⃗ yj · ⃗λ′)H(⃗ yj · ⃗λ′) .(5) As shown in Ref. [10], the statistics generated by this protocol can be written as PC(i|⃗ r, Y) = Z S2 d⃗λ ρ(⃗λ|⃗ r, Y)PB(i|⃗λ, Y),(6) whered ⃗λ=dΩ/(4π) denotes the normalized uni- form measure on the sphere. The effective distributio...

  6. [6]

    In addition, they share a random unit vector ⃗λ∈S 2, uniformly distributed over the Bloch sphere

    Alice receives the description of a qubit stateρ ⃗ r, represented by its Bloch vector⃗ r, and Bob receives the description of an e-POVMM Y specified byY (Remember,p i = 1/m∀i). In addition, they share a random unit vector ⃗λ∈S 2, uniformly distributed over the Bloch sphere

  7. [7]

    Alice computes the bit c=H( ⃗λ·⃗ r),(15) and sends it to Bob

  8. [8]

    Bob defines the effective hidden variable ⃗λ ′ = (2c−1) ⃗λ.(16) Thus, Bob flips the shared vector wheneverc= 0

  9. [9]

    6 shows the response of (17) for the octahedral outcome associated toi= 5

    Bob reports the outcomeiwith probability P1−bit(i|⃗λ ′, Y) = (⃗ yi · ⃗λ ′)H(⃗ yi · ⃗λ ′)Pm j=1(⃗ yj · ⃗λ ′)H(⃗ yj · ⃗λ ′) .(17) The right panel of Fig. 6 shows the response of (17) for the octahedral outcome associated toi= 5. Its geomet- ric structure closely reproduces the ensamble-averaged neural network response: the probability of outcomeiis maximal ...

  10. [10]

    Brassard, Foundations of Physics33, 1593 (2003)

    G. Brassard, Foundations of Physics33, 1593 (2003)

  11. [11]

    Buhrman, R

    H. Buhrman, R. Cleve, S. Massar, and R. De Wolf, Re- views of Modern Physics82, 665 (2010)

  12. [12]

    Gallego, N

    R. Gallego, N. Brunner, C. Hadley, and A. Ac ´ ın, Physical Review Letters105, 230501 (2010)

  13. [13]

    de Gois, G

    C. de Gois, G. Moreno, R. Nery, S. Brito, R. Chaves, and R. Rabelo, PRX Quantum2, 030311 (2021)

  14. [14]

    Paw lowski and N

    M. Paw lowski and N. Brunner, Phys. Rev. A84, 010302(R) (2011)

  15. [15]

    Bowles, N

    J. Bowles, N. Brunner, and M. Paw lowski, Phys. Rev. A 92, 022351 (2015)

  16. [16]

    N. J. Cerf, N. Gisin, and S. Massar, Physical Review Letters84, 2521 (2000)

  17. [17]

    B. F. Toner and D. Bacon, Physical Review Letters91, 187904 (2003)

  18. [18]

    Degorre, S

    J. Degorre, S. Laplante, and J. Roland, Physical Review A72, 062314 (2005)

  19. [19]

    M. J. Renner, A. Tavakoli, and M. T. Quintino, Physical Review Letters130, 120801 (2023)

  20. [20]

    Montina, Physical Review A—Atomic, Molecular, and Optical Physics84, 042307 (2011)

    A. Montina, Physical Review A—Atomic, Molecular, and Optical Physics84, 042307 (2011)

  21. [21]

    Rudolph, arXiv preprint quant-ph/0608120 (2006)

    T. Rudolph, arXiv preprint quant-ph/0608120 (2006)

  22. [22]

    Zartab, G

    M. Zartab, G. Gasbarri, G. Sent ´ ıs, and R. Mu˜ noz-Tapia, Scientific Reports16, 21297 (2026)

  23. [23]

    Schl¨ osser and M

    S. Schl¨ osser and M. Kleinmann, arXiv preprint arXiv:2603.01255 (2026)

  24. [24]

    Sidajaya, A

    P. Sidajaya, A. D. Lim, B. Yu, and V. Scarani, Quantum 7, 1150 (2023)

  25. [25]

    Heinosaari, M

    T. Heinosaari, M. A. Jivulescu, and I. Nechita, Journal of Mathematical Physics61, 042202 (2020)

  26. [26]

    Eldar and G

    Y. Eldar and G. Forney, IEEE Transactions on Informa- tion Theory48, 599 (2002)

  27. [27]

    P. G. Casazza and G. Kutyniok,Finite frames: Theory and applications(Springer Science & Business Media, 2013)

  28. [28]

    Slomczynski and A

    W. Slomczynski and A. Szymusiak, arXiv preprint quant- ph/1402.0375 (2014)

  29. [29]

    Cobucci, R

    G. Cobucci, R. Brinster, S. Khandelwal, H. Kamper- mann, D. Bruß, N. Wyderka, and A. Tavakoli, Phys. Rev. Lett.136, 060201 (2026)

  30. [30]

    Curran, arXiv preprint quant-ph/2605.18291 (2026)

    F. Curran, arXiv preprint quant-ph/2605.18291 (2026)

  31. [31]

    Montina, Physical Review Letters109, 110501 (2012)

    A. Montina, Physical Review Letters109, 110501 (2012)

  32. [32]

    M. J. Renner and M. T. Quintino, Quantum7, 1149 (2023)

  33. [33]

    Kriv´ achy, Y

    T. Kriv´ achy, Y. Cai, D. Cavalcanti, A. Tavakoli, N. Gisin, and N. Brunner, npj Quantum Information6, 70 (2020)

  34. [34]

    C. Zhu, R. H. Byrd, P. Lu, and J. Nocedal, ACM Transac- tions on mathematical software (TOMS)23, 550 (1997). 10 Appendix A: Neural network training methodology and hyperparameter selection

  35. [35]

    First, we generateMrandom pure qubit states sampled uniformly on the Bloch sphere

    T raining data and batch-averaged loss For a given POVMM Y , the training dataset is constructed as follows. First, we generateMrandom pure qubit states sampled uniformly on the Bloch sphere. We denote their Bloch vectors by⃗ r(j), withj= 1, . . . , M. For each state⃗ r(j), the target quantum probability distribution is computed using the Born rule, PQ(i|...

  36. [36]

    The input layer has four neurons, corre- sponding to the three Cartesian components of ⃗λand the communicated bitc

    Network architecture and optimization For each POVM, we train an independent feed-forward neural network. The input layer has four neurons, corre- sponding to the three Cartesian components of ⃗λand the communicated bitc. The network has two hidden layers with 10 and 16 neurons, respectively. The output layer hasmneurons, wheremis the number of outcomes o...

  37. [37]

    To select suitable values ofMandN, we analyze the performance of the NN-procedure for a POVM withm= 6 elements

    Performance metric and selection ofMandN The performance of the NN-procedure is evaluated using MAE between the batch-averaged neural network predic- tion and the target quantum probability distribution: MAE = 1 M MX j=1 mX i=1 1 N NX k=1 ˆy(j,k) i −y (j) i = 1 M MX j=1 mX i=1 ¯p(j) i −P Q(i|ρ⃗ r(j) ,M Y ) .(A7) As in the loss function, the average over t...