Sample-optimal single-copy quantum state tomography via shallow depth measurements

Dohun Kim; Gyungmin Cho

REVIEW 1 major objections 4 minor 30 references

Sample-optimal single-copy quantum state tomography via shallow depth measurements

T0 review · 1 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Using two-layer Clifford circuits of depth O(log n), this paper proves single-copy quantum state tomography achieves optimal sample complexity for full-rank states and near-optimal for rank-r states.

desk verdict The full-rank shallow-depth QST result is solid and significant; the rank-r theorem has a real gap at the final trace-norm step and should be revised before it is cited as proven. read the letter →

arxiv 2509.12703 v2 pith:IQC4AQOL submitted 2025-09-16 quant-ph

Gyungmin Cho , Dohun Kim This is my paper

classification quant-ph MSC 81P4581P5081P68

keywords quantumstatetomographysamplecomplexityshadowchannelCliffordcircuitsshallow-depthrank-rstatestracedistancerandomizedmeasurements

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Quantum state tomography normally needs either collective measurements on many copies (hard on current hardware) or complicated single-copy POVMs. This paper shows that a two-layer brickwork of random Clifford gates, each layer acting on O(log n) qubits, is enough to make single-copy measurements sample-optimal. For a rank-r state in dimension d=2^n, the protocol reconstructs the state to trace error epsilon using O(d r^2 ln(d)/epsilon^2) copies, matching the known lower bound up to a factor of ln d. For full-rank states it removes the log factor and achieves the optimal O(d^3/epsilon^2) sample complexity. The practical point is that a depth-O(log n) circuit extracts enough information per copy to match the information-theoretic limits.

What carries the argument

The load-bearing object is the unbiased estimator rho_hat = M^{-1}(U^dagger |b><b| U), where M is the shadow channel of the random Clifford ensemble, defined as the expectation M(rho) = E_{U,b}[U^dagger |b><b| U]. The paper computes the second moment E[rho_hat^2] by mapping Pauli-correlation sums onto a transfer matrix acting on a 2-state chain and bounding its largest eigenvalue; this yields the operator-norm concentration bound via the matrix Bernstein inequality. For the full-rank case the machinery is a per-Pauli estimator whose Frobenius-norm concentration is controlled by McDiarmid's inequality.

What would settle it

For a fixed rank-r state and the two-layer brickwork ensemble with k = O(log n), compute the ratio of trace norm to operator norm of the single-sample estimator error. If this ratio exceeds 2r with non-negligible probability, the proof's conversion epsilon_op = epsilon_tr/(2r) fails, and Theorem 1's sample bound as proven would not follow; a direct simulation of the sample complexity versus d and r could then check whether the claimed rate still holds.

Watch

Extended reading notes

Core claim

The central discovery is that the unbiased estimator constructed from the shadow channel of a shallow two-layer Clifford ensemble concentrates fast enough for quantum state tomography. Theorem 1 states that for rank-r states, depth O(log n) two-layer brickwork random Clifford circuits perform (epsilon, delta)-QST with T = O(d r^2 ln(d)/epsilon^2), near-optimal because the single-copy lower bound is Omega(d r^2/epsilon^2). Theorem 2 states that the simpler block-random Clifford ensemble of depth O(log n) achieves T = O(d^3/epsilon^2) for full-rank states, which is exactly optimal. The proofs use the matrix Bernstein inequality (Theorem 1) and McDiarmid's inequality (Theorem 2), and they do no

Load-bearing premise

The rank-r sample bound stands only if the error matrix (estimator minus true state) has rank at most 2r, which justifies converting an operator-norm error bound into a trace-norm error bound; the paper's single-copy estimator is generically full rank and is not truncated, so this is the load-bearing step.

Editorial extensions

If this is right

For full-rank states, single-copy tomography with shallow circuits uses the same number of samples as the information-theoretic optimum; no single-copy protocol can beat it asymptotically.
For rank-r states, the gap to the lower bound is only a logarithmic factor in dimension, so the protocol is essentially sample-optimal.
The circuits use only Clifford gates and are ancilla-free, so the depth O(log n) requirements are within reach of current noisy quantum devices.
The concentration proof does not use approximate unitary design properties, indicating that the circuit architecture (overlapping blocks of random Clifford gates) is itself what drives sample efficiency.
Open boundary conditions, which are more realistic than the periodic boundary conditions used in the main proof, still give the same near-optimal sample complexity (Appendix E).

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural next step is to test numerically whether truncating the estimator to rank r before applying the trace-norm bound restores the optimal d r^2/epsilon^2 scaling without the ln d factor.
The transfer-matrix technique for second moments is likely reusable for other randomized-measurement estimators under shallow circuits, such as purity estimates or out-of-time-order correlators where unbiased estimators may be needed.
Because the full-rank optimality uses a simple block ensemble that is not an approximate design, sample-optimal QST may require only local scrambling within O(log n)-sized blocks, not global pseudorandomness.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The full-rank half of this paper is the real news. Theorem 2 shows that sample-optimal single-copy QST, O(d^3/eps^2), can be done with depth-O(log n) local Clifford circuits, and the proof via McDiarmid is clear and convincing. The appendix F extension to the brickwork ensemble also checks out. That alone is a worthwhile result and removes a practical obstacle for near-term tomography.

The rank-r part, Theorem 1, is where the trouble sits, and the stress-test note is right. The proof runs matrix Bernstein and gets an operator-norm bound on rho_hat_T - rho. Then it sets eps_op = eps_tr/(2r) and claims trace-norm control. That step requires rank(rho_hat_T - rho) <= 2r. But the single-shot estimator M^{-1}(U^\dagger |b>

The gap is specific and looks addressable. Adding a rank-r projection step would break unbiasedness, so one would need to redo the bias and variance analysis, or find a different argument that uses rank-r structure before concentration. Until then, Theorem 1 should be treated as a conjecture with a plausible but incomplete proof.

Otherwise the paper is honest and well-contained. The appendices give enough detail to reconstruct the computations, and the discussion of open boundary conditions is welcome. The reliance on the authors' own prior work [18] for classical post-processing is fine; that is a separate result, not a circular load. No data or code are needed for this type of theory paper.

For peer review: yes, an editor should send this to referees. The full-rank theorem is significant and appears correct, and even if Theorem 1 needs revision, the paper deserves referee time. I would suggest the referee focus on the rank-r trace-norm step and on whether a repaired argument can preserve the claimed scaling.

Referee Report

1 major / 4 minor

Summary. The paper studies single-copy quantum state tomography with shallow local Clifford circuits. It claims: (i) for rank-r states, a two-layer brickwork Clifford measurement of depth O(log n) achieves (ε, δ)-QST with O(d r^2 ln(d/δ)/ε^2) copies, matching the single-copy lower bound up to ln d; (ii) for full-rank states, a simpler block Clifford measurement of depth O(log n) achieves O(d^3/ε^2), which is optimal. The proofs use the classical-shadow formalism with an unbiased estimator; Theorem 1 is proved via the matrix Bernstein inequality and a transfer-matrix calculation, Theorem 2 via McDiarmid's inequality. Appendix B gives a counterexample to using the Haar-shadow biased estimator for shallow circuits.

Significance. Sample-optimal single-copy QST is an active problem, and a shallow-depth construction would be practically valuable. The full-rank result (Theorem 2) is notable: it only needs per-block mutually unbiased bases, not approximate unitary designs, and would match the known Ω(d^3/ε^2) lower bound. The transfer-matrix method in Theorem 1 is elegant, and the concentration proofs are self-contained given standard shadow-channel facts. The paper contains no post-hoc fitting or circular argument. The main caveat is that the rank-r theorem's proof has a missing rank bound at the operator-to-trace-norm conversion; until this is supplied, the near-optimal rank-r claim is unproven.

major comments (1)

[Section IV, Theorem 1 proof, Eq. (47)–(48)] The proof obtains a high-probability bound on ||ρ̂_T − ρ||_op and then sets ε_op = ε_tr/(2r) to replace the operator norm by the trace norm. This step implicitly uses ||A||_tr ≤ rank(A)||A||_op together with rank(ρ̂_T − ρ) ≤ 2r. No such rank bound is proved. For the block-Clifford inverse shadow map one has M_k^{-1}(|b⟩⟨b|) = (2^k+1)|b⟩⟨b| − I, with eigenvalues 2^k and −1, so a single snapshot is full rank; for the two-layer brickwork ensemble the inverse map is diagonal in the Pauli basis with coefficients m_P^{-1} > 0, and no projection or truncation is applied to ρ̂_T. Thus rank(ρ̂_T − ρ) is not bounded by 2r and can be as large as d. The claimed T = O(d r^2 ln d / ε^2) therefore does not follow from the displayed argument; the trivial rank bound would give only O(d^3/ε^2). This is load-bearing for the rank-r theorem.

minor comments (4)

[Section IV, Theorem 2 proof, Eq. (55)–(56)] The displayed fraction in the McDiarmid exponent does not simplify to Eq. (56) as written. With T = M(2^k+1)^{n/k} and per-coordinate changes C_i = 2(4^k+2^k−1)^{n/(2k)}/(M(2^k+1)^{n/k}), the standard bound gives exponent −M t^2/(2A(n,k)); the final expression is correct, but the fraction in Eq. (55) should be written as 2t^2/(T C_i^2) or equivalently without the extra M(2^k+1)^{n/k} in the numerator.
[Theorem 2 proof, norm conversion] Please clarify the conversion from Frobenius to trace norm. The correct inequality is ||A||_tr ≤ √(2^n) ||A||_F, so one should take ε_F = ε_tr/√(2^n). The main text appears to contain a typesetting ambiguity here.
[Abstract / Section V] The abstract claims 'nearly optimal classical runtime for explicit matrix output', but the paper does not analyze classical runtime. Either provide a runtime statement or soften the claim.
[Appendix E] The open-boundary-condition treatment is terse. The definitions of F̃ and G̃ are given, but the derivation that h(F G̃) and h(F̃ G) are Θ(λ_+^m) is only sketched; additional detail would help the reader verify the OBC claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the sampling bounds are derived from self-contained concentration proofs; the flagged issue at Eq. (47) is a proof gap, not an input-output equivalence.

full rationale

The paper's central claims are derived rather than imported. Theorem 1 proves concentration of the unbiased shadow estimator by bounding ||rho_hat||_op and ||E rho_hat^2||_op using the Clifford shadow channel, Weingarten calculus, and a transfer-matrix computation, then applies the matrix Bernstein inequality. Theorem 2 and its full-rank variant prove concentration via McDiarmid's inequality using explicit bounds on the Frobenius error and its bounded-difference constants. No parameter is fitted to data, no estimator is calibrated from the quantity it later 'predicts', and no lower bound is used as an upper bound by construction. The citations to the authors' prior work [18] are not load-bearing: the Clifford-Pauli properties stated with citation [18] are restated and proved in Appendix A, and [18] is otherwise used only for practical classical post-processing remarks, not for the sample-complexity theorems. The serious issue in the rank-r proof is the step around Eq. (47) where epsilon_op = epsilon_tr/(2r) is used to convert an operator-norm bound into a trace-norm bound without proving that rho_hat_T - rho has rank at most 2r; the inverse-shadow estimator is generically full rank. That is a mathematical gap in the claimed theorem, not a circularity: the conclusion does not reduce to an input by definition, and no fitted quantity is renamed as a prediction. Under the provided criteria, no circular step is exhibited, so the appropriate finding is no significant circularity.

Assumptions & free parameters 1 free parameters · 6 assumptions · 0 invented entities

The paper introduces no new physical entities or fitted constants. The main free choice is the block size k, set from theory. The axioms are standard probabilistic inequalities and Clifford-group design properties, plus the rank-r assumption. No hidden phenomenological parameters appear.

free parameters (1)

Block size k = k = 2 log n (Theorem 1); k = log n (Theorem 2)
Algorithm design parameter controlling circuit depth and concentration bounds; chosen by hand from theory, not fitted to external data.

assumptions (6)

standard math Matrix Bernstein inequality
Used in Appendix C to bound the operator norm of the estimator error in Theorem 1.
standard math McDiarmid's inequality
Used in Theorem 2 to turn Frobenius-norm expectation and bounded differences into a high-probability bound.
standard math Clifford group is a unitary 2-design
Used to evaluate E_{U1} contractions in Eq. (27) of Theorem 1.
domain assumption Shadow channel has Pauli eigenoperators
Assumes the measurement ensemble is Clifford so M(P) = m_P P; derived in Appendix A using Pauli algebra.
ad hoc to paper Block size assumptions: k divides n and k is even
The theorems state O(log n) depth but require k = 2 log n or log n to divide n; not guaranteed for arbitrary n, although padding or rounding likely handles it.
domain assumption Rank-r state prior
Theorem 1 assumes the unknown state has rank at most r and uses r in the error conversion, without proving the estimator difference has rank bounded by 2r.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Sample-optimal single-copy quantum state tomography via shallow depth measurements." pith.science (2026). https://pith.science/paper/IQC4AQOL

@misc{pith2026250912703,
  author       = {Pith},
  title        = {Pith review of: Sample-optimal single-copy quantum state tomography via shallow depth measurements},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IQC4AQOL}},
  note         = {Machine review of arXiv:2509.12703}
}

abstract

Quantum state tomography (QST) is a central task in quantum information, and its efficiency is commonly characterized by sample complexity. Although collective measurements on multiple copies achieve optimal performance, they are difficult to implement on near-term devices, motivating the study of single-copy approaches. Here, we introduce a ancilla-free single-copy QST protocol based on logarithmic-depth local circuits on an $n$-qubit system. For rank-$r$ states in dimension $d=2^n$, our protocol achieves trace-norm error $\epsilon$ using $\mathcal{O}(dr^2\log d/\epsilon^2)$ copies, matching the single-copy lower bound up to a logarithmic factor. For full-rank mixed states, it removes this logarithmic overhead and achieves the optimal scaling $\mathcal{O}(d^3/\epsilon^2)$, with nearly optimal classical runtime for explicit matrix output. These results show that sample-optimal QST can be realized using experimentally accessible shallow-depth measurements.

Figures

Figures reproduced from arXiv: 2509.12703 by the authors.

**Figure 2.** FIG. 2. (a) Block random unitary ensemble, where each block is uniformly sampled from Cl( [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** FIG. 3. Case classification by matrix elements of [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: FIG. 4. (a) Two-layer patched brickwork circuit with periodic boundary. (b) and (c) Open boundary [PITH_FULL_IMAGE:figures/full_fig_p022_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 2 linked inside Pith

[18]

Schuster, J

T. Schuster, J. Haferkamp, and H.-Y. Huang, Science389, 92 (2025)

2025
[1]

M. A. Nielsen and I. L. Chuang,Quantum Computation and Quantum Information: 10th Anniversary Edition(Cambridge University Press, 2010)

2010
[2]

Note thatk= 1 2 lognin the two-layer brickwork unitary ensemble has the same circuit volume ask= lognin the block unitary ensemble

f(ρ;X)−f(ρ;X (i))≤ 2L NU NS .(F27) Applying these two properties to McDiarmid’s inequality, we obtain Pr(f(ρ)≥ϵ F)≤δ(F28) whenever NU NS = L ϵ2 F p 2n +N S +L 1/2 ln(2/δ) 2 .(F29) Let us assumeN S ≤ O(2n), then NU NS =O L ϵ2 F max(2n, L) (F30) =O L2 ϵ2 F .(F31) Therefore, by takingϵ F =ϵ tr/ √ 2n, we can ensure that Pr(∥ˆρT −ρ∥ tr ≥ϵ tr)≤δ,(F32) whenT=N U...
[3]

C. F. Roos, G. P. T. Lancaster, M. Riebe, H. H¨ affner, W. H¨ ansel, S. Gulde, C. Becher, J. Eschner, F. Schmidt-Kaler, and R. Blatt, Phys. Rev. Lett.92, 220402 (2004)

2004
[4]

P. Yang, M. Yu, R. Betzholz, C. Arenz, and J. Cai, Phys. Rev. Lett.124, 010405 (2020). 26

2020
[5]

DiCarlo, M

L. DiCarlo, M. D. Reed, L. Sun, B. R. Johnson, J. M. Chow, J. M. Gambetta, L. Frunzio, S. M. Girvin, M. H. Devoret, and R. J. Schoelkopf, Nature467, 574 (2010)

2010
[6]

Hangleiter and M

D. Hangleiter and M. J. Gullans, Physical Review Letters133, 020601 (2024)

2024
[7]

S. Chen, W. Gong, Q. Ye, and Z. Zhang, inProceedings of the 57th Annual ACM Symposium on Theory of Computing, STOC ’25 (Association for Computing Machinery, New York, NY, USA, 2025) p. 429–438

2025

Show all 30 references

[8]

Grewal, V

S. Grewal, V. Iyer, W. Kretschmer, and D. Liang, inProceedings of the 56th Annual ACM Symposium on Theory of Computing(2024) pp. 1352–1363

2024
[9]

Leone, S

L. Leone, S. F. Oliviero, and A. Hamma, Quantum8, 1361 (2024)

2024
[10]

Landau and Y

Z. Landau and Y. Liu, inProceedings of the 57th Annual ACM Symposium on Theory of Computing(2025) pp. 1828–1838

2025
[11]

J. Haah, A. W. Harrow, Z. Ji, X. Wu, and N. Yu, inProceedings of the forty-eighth annual ACM symposium on Theory of Computing(2016) pp. 913–925

2016
[12]

S. Chen, B. Huang, J. Li, A. Liu, and M. Sellke, in2023 IEEE 64th Annual Symposium on Foundations of Computer Science (FOCS)(IEEE, 2023) pp. 391–404

2023
[13]

O’Donnell and J

R. O’Donnell and J. Wright, inProceedings of the forty-eighth annual ACM symposium on Theory of Computing(2016) pp. 899–912

2016
[14]

S. Chen, J. Li, and A. Liu, inProceedings of the 56th Annual ACM Symposium on Theory of Computing(2024) pp. 1331–1342

2024
[15]

Kueng, H

R. Kueng, H. Rauhut, and U. Terstiege, Applied and Computational Harmonic Analysis42, 88 (2017)

2017
[16]

Gut ¸˘ a, J

M. Gut ¸˘ a, J. Kahn, R. Kueng, and J. A. Tropp, Journal of Physics A: Mathematical and Theoretical53, 204001 (2020)

2020
[17]

Lowe and A

A. Lowe and A. Nayak, ACM Transactions on Computation Theory17, 10.1145/3717450 (2025)

2025 doi
[19]

Cho and D

G. Cho and D. Kim, arXiv preprint arXiv:2504.15698 (2025)

2025 arXiv
[20]

Collins, S

B. Collins, S. Matsumoto, and J. Novak, Notices of the American Mathematical Society69, 1 (2022)

2022
[21]

Onsager, Phys

L. Onsager, Phys. Rev.65, 117 (1944)

1944
[22]

Huang, R

H.-Y. Huang, R. Kueng, and J. Preskill, Nature Physics16, 1050 (2020). 27

2020
[23]

Bravyi, R

S. Bravyi, R. Shaydulin, S. Hu, and D. Maslov, Quantum5, 580 (2021)

2021
[24]

Acharya, A

J. Acharya, A. Dharmavarapu, Y. Liu, and N. Yu, inProceedings of the 57th Annual ACM Symposium on Theory of Computing(2025) pp. 718–729

2025
[25]

Acharya, A

J. Acharya, A. Dharmavarapu, Y. Liu, and N. Yu, arXiv preprint arXiv:2507.22001 (2025)

2025 arXiv
[26]

Huang, R

H.-Y. Huang, R. Kueng, G. Torlai, V. V. Albert, and J. Preskill, Science377, eabk3333 (2022)

2022
[27]

Elben, S

A. Elben, S. T. Flammia, H.-Y. Huang, R. Kueng, J. Preskill, B. Vermersch, and P. Zoller, Nature Reviews Physics5, 9 (2023)

2023
[28]

Cioli, E

R. Cioli, E. Ercolessi, M. Ippoliti, X. Turkeshi, and L. Piroli, Quantum9, 1698 (2025)

2025
[29]

Bertoni, J

C. Bertoni, J. Haferkamp, M. Hinsche, M. Ioannou, J. Eisert, and H. Pashayan, Physical Review Letters133, 020602 (2024)

2024
[30]

H.-Y. Hu, A. Gu, S. Majumder, H. Ren, Y. Zhang, D. S. Wang, Y.-Z. You, Z. Minev, S. F. Yelin, and A. Seif, Nature Communications16, 2943 (2025). 28

2025

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [18]

Schuster, J

T. Schuster, J. Haferkamp, and H.-Y. Huang, Science389, 92 (2025)

2025

[2] [1]

M. A. Nielsen and I. L. Chuang,Quantum Computation and Quantum Information: 10th Anniversary Edition(Cambridge University Press, 2010)

2010

[3] [2]

Note thatk= 1 2 lognin the two-layer brickwork unitary ensemble has the same circuit volume ask= lognin the block unitary ensemble

f(ρ;X)−f(ρ;X (i))≤ 2L NU NS .(F27) Applying these two properties to McDiarmid’s inequality, we obtain Pr(f(ρ)≥ϵ F)≤δ(F28) whenever NU NS = L ϵ2 F p 2n +N S +L 1/2 ln(2/δ) 2 .(F29) Let us assumeN S ≤ O(2n), then NU NS =O L ϵ2 F max(2n, L) (F30) =O L2 ϵ2 F .(F31) Therefore, by takingϵ F =ϵ tr/ √ 2n, we can ensure that Pr(∥ˆρT −ρ∥ tr ≥ϵ tr)≤δ,(F32) whenT=N U...

[4] [3]

C. F. Roos, G. P. T. Lancaster, M. Riebe, H. H¨ affner, W. H¨ ansel, S. Gulde, C. Becher, J. Eschner, F. Schmidt-Kaler, and R. Blatt, Phys. Rev. Lett.92, 220402 (2004)

2004

[5] [4]

P. Yang, M. Yu, R. Betzholz, C. Arenz, and J. Cai, Phys. Rev. Lett.124, 010405 (2020). 26

2020

[6] [5]

DiCarlo, M

L. DiCarlo, M. D. Reed, L. Sun, B. R. Johnson, J. M. Chow, J. M. Gambetta, L. Frunzio, S. M. Girvin, M. H. Devoret, and R. J. Schoelkopf, Nature467, 574 (2010)

2010

[7] [6]

Hangleiter and M

D. Hangleiter and M. J. Gullans, Physical Review Letters133, 020601 (2024)

2024

[8] [7]

S. Chen, W. Gong, Q. Ye, and Z. Zhang, inProceedings of the 57th Annual ACM Symposium on Theory of Computing, STOC ’25 (Association for Computing Machinery, New York, NY, USA, 2025) p. 429–438

2025