REVIEW 2 major objections 5 minor 14 references
Procrustes Tomography -- reconstructing noisy quantum channels made easy
T0 review · 2 major / 5 minor · reviewed 2026-07-10 · grok-4.5
Pith's one-line read Procrustes tomography reconstructs noisy quantum channels from reconstructed states, matching or beating standard methods under realistic sampling and preparation noise.
desk verdict Clean pedagogical reformulation of process tomography as a Procrustes problem; competitive numerics, incremental novelty, worth a referee. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The Procrustes fit Lambda = min_L ||L A - B||_F (or its weighted form), solved by the Moore-Penrose pseudoinverse Lambda = B A^+; subsequent spectral truncation of states, Knee et al. CPTP projection, and Kronecker-product SVD for nearest-unitary projection.
What would settle it
Re-run the same two-qubit Lindblad channels, replace the linear-inversion state tomography step with maximum-likelihood or Bayesian state tomography at identical total shot counts, and check whether Procrustes still matches or beats Choi and linear-inversion process tomography on average gate infidelity.
Extended reading notes
Core claim
Process tomography can be reduced to an ordinary Procrustes least-squares problem on reconstructed density matrices: the process superoperator is recovered by Lambda = B A^+, where the columns of A and B are vectorized input and output states. With optional weighting, input-state tomography, CPTP projection, and unitary projection, this construction matches or outperforms both Choi reconstruction and linear-inversion process tomography for representative noisy two-qubit channels under the sampling and preparation conditions studied.
Load-bearing premise
The method's reported advantage rests on using linear-inversion state tomography (with negative-eigenvalue truncation) as the estimator of the output density matrices; better state estimators are not compared.
Editorial extensions
If this is right
- Any future improvement in state tomography (readout-error mitigation, machine-learning estimators, etc.) immediately improves process tomography with no change to the Procrustes algorithm.
- State-preparation errors can be removed by simply performing state tomography on both the prepared inputs and the process outputs, at the cost of roughly twice the experiments.
- When a gate is known to be nearly unitary, the Kronecker-product SVD projection yields a rank-1 process that saturates at lower infidelity and can be characterized with as few as d+1 states.
- Uneven sampling of cheap computational-basis preparations versus expensive superpositions can be optimally weighted by the fourth root of the shot counts, reducing over-bias from low-sample states.
Reading between the lines
- Because the method never requires the user to form virtual non-physical operators, it is immediately usable by experimental groups that already have a working state-tomography pipeline.
- The same A/B construction could be applied to continuous-time process characterization by treating intermediate-time density matrices as additional columns, turning tomography into a trajectory-fitting problem.
- If cloud providers expose only limited mid-circuit measurement or reset options, the modularity of Procrustes still allows users to characterize the effective channel they actually experience.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript introduces Procrustes tomography for reconstructing noisy quantum channels: after preparing a set of input states, performing state tomography on the outputs, and assembling the vectorized density matrices into matrices A (inputs) and B (outputs), the process matrix is obtained by the ordinary Procrustes solution Λ = B A⁺ (Eq. 10), optionally with a diagonal weighting matrix W that accounts for uneven shot counts (Eq. 11). The method is compared with the textbook Choi-matrix reconstruction and with linear-inversion process tomography under Lindblad noise (amplitude damping + dephasing) for two-qubit gates. Numerical experiments (Figs. 1–4) examine even and uneven sampling, imperfect state preparation, and optional projection onto the nearest unitary channel; they show that the Procrustes estimator matches or exceeds the fidelity of the two baselines, especially once input-state tomography or unitary projection is included. Finite-sample non-CPTP maps are corrected by the Knee et al. projection for all three methods.
Significance. If the reported performance advantage holds under realistic device conditions, the work supplies a lightweight, pedagogically transparent alternative to standard process tomography that is immediately usable by NISQ-cloud users who lack specialized tomography expertise. Strengths include an explicit, parameter-free least-squares estimator, a clean treatment of state-preparation errors (simply double the tomography), a natural weighting scheme for uneven sampling, and a unitary-projection step based on the Kronecker-product SVD. The numerical comparisons under controlled Lindblad dynamics are clear and support the claim within the simulated regime. The paper does not claim asymptotic optimality or experimental validation, so its contribution is best viewed as a practical, extensible reconstruction recipe rather than a fundamental advance in tomography theory.
major comments (2)
- Numerical simulations section (paragraph beginning “For the following examples…”) and the discussion after Eq. (10): all Procrustes results are generated with linear-inversion state tomography (plus optional negative-eigenvalue truncation). The paper itself notes that, under pure linear inversion, Procrustes and the linear-inversion baseline become mathematically identical. Consequently the observed advantage is attributable only to the optional PSD truncation, the weighting matrix W, or the input-state tomography step. Without at least one comparison against maximum-likelihood or Bayesian state tomography, it remains unclear whether the claimed superiority survives once a higher-quality state estimator is used for every method. A short additional panel or appendix addressing this point would substantially strengthen the central claim.
- Figs. 1–4 and the associated text: all numerical evidence is restricted to two-qubit channels generated by a single Lindblad model (T1/T2 ratios and absolute T2 values listed in the figure captions). While the algebraic construction is dimension-independent, the performance ranking versus Choi and linear inversion could change for larger systems or for noise that includes coherent control errors, leakage, or non-Markovian effects. The manuscript should either (i) state explicitly that the superiority claim is limited to the two-qubit Lindblad setting examined, or (ii) supply at least one higher-dimensional or differently noised example.
minor comments (5)
- Abstract and opening paragraph: the rhetorical question “What is the most expensive part…?” and the unqualified claim that Procrustes “outperforms established methods in a number of aspects” should be tempered to match the more careful language used in the concluding remarks.
- Eq. (2) and surrounding text: the linear combination that recovers |n⟩⟨m| from physical states is standard, yet the precise coefficients for the two-qubit case used later are never written out; a short explicit formula would aid reproducibility.
- Fig. 2 caption and text: the non-monotonic infidelity at low shot counts is attributed to “over-biasing toward Clifford operations,” but no quantitative diagnostic (e.g., diamond-norm distance to the nearest Clifford) is supplied; a one-sentence clarification would help.
- Typographical issues: “breifly” (p. 2), “porocess” (p. 4), “Figu. 2” (p. 4), and inconsistent hyphenation of “state-preparation” / “state preparation” appear throughout.
- References: several recent works on GST, compressed sensing, and ML-based process tomography are cited, yet the classic maximum-likelihood process tomography literature (e.g., Śšvanda et al., Fiurášek) is under-represented; adding one or two standard citations would improve context.
Circularity Check
No circularity: Procrustes estimator is ordinary least-squares against independent simulated ground truth; performance metric is external.
full rationale
The paper defines Procrustes tomography as the ordinary least-squares solution Λ = B A^{+} (Eq. 10) of the matrix equation B = Λ A formed from reconstructed input and output density matrices. This is a standard Procrustes / pseudoinverse construction; no free parameters are fitted to the fidelity data that later serve as the figure of merit. The numerical comparisons (Figs. 1–4) evaluate average gate infidelity against independently generated Lindblad channels whose Hamiltonians and jump operators are fixed before any reconstruction is performed. When linear-inversion state tomography is used for both Procrustes and the linear-inversion baseline, the paper itself notes that the two estimators become mathematically identical (text after Eq. 10); residual advantages arise only from optional, explicitly derived extensions (PSD truncation, weighting matrix W, input-state tomography, unitary projection) that the baselines do not receive. Self-citations are limited to prior work on related topics (efficient Hamiltonians, etc.) and do not close any definitional loop. The derivation is therefore self-contained against external benchmarks and exhibits no circular reduction.
Assumptions & free parameters
free parameters (3)
- T1/T2 ratios and absolute T2 values used in simulations =
T1=5 T2; T2 in {10,100,1000}
- shot-allocation ratio for uneven sampling =
Ni = 10 N_pm
- weighting exponent for Procrustes weights =
1/4
assumptions (4)
- domain assumption Completely-positive trace-preserving maps are linear and can be represented by a process matrix Lambda acting on vectorized density operators.
- standard math The Frobenius-norm least-squares solution Lambda = B A^+ is the unique minimizer of the Procrustes objective when A has full row rank (or the minimal-norm solution otherwise).
- ad hoc to paper Negative eigenvalues of reconstructed density matrices may be zeroed and the matrix re-normalized without destroying the subsequent process estimate.
- domain assumption The Knee et al. projection maps any estimated superoperator onto the nearest CPTP map in a way that preserves the ranking of reconstruction methods.
Cite this review
Pith. "Pith review of Procrustes Tomography -- reconstructing noisy quantum channels made easy." pith.science (2026). https://pith.science/paper/ARUQYT2W
@misc{pith2026260707988,
author = {Pith},
title = {Pith review of: Procrustes Tomography -- reconstructing noisy quantum channels made easy},
year = {2026},
howpublished = {\url{https://pith.science/paper/ARUQYT2W}},
note = {Machine review of arXiv:2607.07988}
}
read the original abstract
What is the most expensive part of quantum device characterization? Clearly, the answer is quantum process tomography. However, especially for noisy intermediate-scale quantum (NISQ) computers, a comprehensive understanding of the noisy quantum dynamics is essential in interpreting the computational output. In this work, we introduce an efficient method -- Procrustes tomography -- that outperforms established methods in a number of aspects. After a pedagogical and constructive introduction, we demonstrate the utility of the method for representative examples of noisy quantum channels.
Figures
Reference graph
Works this paper leans on
-
[1]
[1]LaRose R.,Quantum,3(2019)
work page 2019
-
[2]
Aquila: QuEra's 256-qubit neutral-atom quantum computer
[2]Ser vices A. W.,Amazon braket supported regions and deviceshttps://docs.aws.amazon.com/braket/latest/ developerguide/braket-devices.htmlaccessed: 2026- 07-04 (2026). [3]IonQ,Introduction – ionq quantum cloud documentation https://docs.ionq.com/user-manual/introduction accessed: 2026-07-04 (2026). [4]Quantinuum,Quantinuum technical documentation https:/...
work page Pith review arXiv 2026
-
[3]
[13]Choo K., Von Keyserlingk C. W., Regnault N. andNeupert T.,Phys. Rev. Lett.,121(2018) 086808. p-6 Procrustes Tomography – reconstructing noisy quantum channels made easy [14]Iqbal M., Tantiv asadakarn N., Verresen R., Campbell S. L., Dreiling J. M., Figgatt C., Gae- bler J. P., Johansen J., Mills M., Moses S. A.et al.,Nature,626(2024)
work page 2018
-
[4]
[15]Bando Y., Susa Y., Oshiyama H., Shibata N., Ohzeki M., Gómez-Ruiz F. J., Lidar D. A., Suzuki S., Del Campo A.andNishimori H.,Phys. Rev. Res., 2(2020) 033369. [16]Robertson R., Doucet E., Spicer E.andDeffner S.,Entropy,27(2025)
work page 2020
-
[5]
J.andDeffner S.,arXiv preprint arXiv:2603.03801, (2026)
[17]Robertson R., Consiglio M., Stevens J., Doucet E., Apollaro T. J.andDeffner S.,arXiv preprint arXiv:2603.03801, (2026) . [18]Irf an A. K.andYu C.,arXiv preprint arXiv:2509.24213, (2025) . [19]Bautra E., Dimitrijevs M.andYakaryilmaz A., arXiv preprint arXiv:2603.14607, (2026) . [20]Di Bartolomeo G., Vischi M., Cesa F., Wixinger R., Grossi M., Donadi S....
-
[6]
J., Acharya A., Carleo G., Carrasquilla J.andAolita L.,Nat
[25]Torlai G., Wood C. J., Acharya A., Carleo G., Carrasquilla J.andAolita L.,Nat. Commun.,14 (2023)
work page 2023
-
[7]
[26]Gaikw ad A., Bihani O., Ar vindandDorai K.,Phys. Rev. A,109(2024) 012402. [27]Ahmed S., Quijandría F.andKockum A. F.,Phys. Rev. Lett.,130(2023) 150402. [28]Altepeter J. B., Branning D., Jeffrey E., Wei T., Kwiat P. G., Thew R. T., O’Brien J. L., Nielsen M. A.andWhite A. G.,Phys. Rev. Lett.,90(2003) 193601. [29]Xue S., W ang Y., Zhan J., W ang Y., Zeng...
work page Pith review arXiv 2024
-
[8]
B., Goss N., Marinelli B., Naik R
[40]Hashim A., Nguyen L. B., Goss N., Marinelli B., Naik R. K., Chistolini T., Hines J., Marceaux J. P., Kim Y., Gokhale P.et al.,PRX Quantum, 6(2025) 030202. [41]Innan N., Siddiqui O. I., Arora S., Ghosh T., Koçak Y. P., Paragas D., Galib A. A. O., Khan M. A.-Z.andBennai M.,Quantum Mach. Intell.,6(2024)
work page 2025
Show all 14 references
-
[9]
A., Phys
[42]Wei V., Coish W., Ronagh P.andMuschik C. A., Phys. Rev. Res.,6(2024) 023250. [43]Lohani S., Kirby B. T., Brodsky M., Danaci O.and Glasser R. T.,Mach. Learn.: Sci. Technol.,1(2020) 035007. [44]Schmale T., Reh M.andGärttner M.,npj Quantum Inf.,8(2022)
2024
-
[10]
S., Di Giov anni A., Rotzinger H., Usti- nov A
[45]Aasen A. S., Di Giov anni A., Rotzinger H., Usti- nov A. V.andGärttner M.,Commun. Phys.,7(2024)
2024
-
[11]
T., Gambetta J
[46]Merkel S. T., Gambetta J. M., Smolin J. A., Po- letto S., Córcoles A. D., Johnson B. R., Ryan C. A.andSteffen M.,Phys. Rev. A,87(2013) 062119. [47]Greenbaum D.,arXiv preprint arXiv:1509.02921, (2015) . [48]Huang T., Gaikw ad A., Moskalenko I., Aggar- w al A., Abad T., Kuzm...
2013 arXiv
-
[12]
[52]Hradil Z.,Phys. Rev. A,55(1997) R1561(R). [53]Aditi K.andBecker S.,Phys. Rev. A,112(2025) 052436. [54]Loan C. F.,J. Comput. Appl. Math.,123(2000) 85 nu- merical Analysis
1997
-
[13]
Phys.,24(2022) 055002
[56]Aifer M.andDeffner S.,New J. Phys.,24(2022) 055002. [57]Stevens J.andDeffner S.,Quantum Sci. Technol.,10 (2025) 04LT03. [58]Higham N. J.,Functions of matrices : theory and com- putation(Cambridge University Press, Cambridge, TAS, Australia)
2022
-
[14]
M.andKoch C
[59]Romer A., Reich D. M.andKoch C. P.,arXiv preprint arXiv:2507.17648, (2025) . p-7
2025 arXiv
Reviewed July 10, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.