{"id":"70f9ff0e-5d50-400a-b919-c07d40c8cd11","arxiv_id":"2505.15243","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Classical shadow tomography can estimate K off-diagonal density matrix elements with O(log K / epsilon^2) samples, a logarithmic improvement over traditional direct measurement protocols.","lead":"This paper applies classical shadow tomography to directly estimate individual density matrix elements of a quantum state, claiming that K off-diagonal entries require only O(log K / epsilon^2) samples. It also proposes a full tomography protocol with O(d^3 log d / epsilon^2) sample complexity, close to the known single-copy lower bound.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The O(log K/ε²) Clifford result is sound, but the biased-MUB factor-1/2 reduction rests on an unproved partition of MUB unitaries; this needs rigorous proof or explicit verification before the advertised MUB advantage is accepted.","rationale":"The reader's weakest-assumption analysis correctly identifies Lemma 2's unproved MUB partition as the main vulnerability. I agree that this is load-bearing for the factor-1/2 MUB claim and for the paper's advertised practical advantage, but not for the central O(log K/ε²) scaling, which follows from Clifford classical shadows. The paper should be accepted only conditionally: the MUB lemma needs a rigorous proof or an explicit construction-based verification before the factor-1/2 claim is stated as a theorem. The asymptotic tomography claim O(d³ log d/ε²) survives either way because both Clifford shadows and the general MUB-sparse bound give O(1) shadow norms. Therefore I do not change the reader's CONDITIONAL verdict.","tokens_in":13370,"tokens_out":30524,"duration_ms":255548,"concrete_test":"Take the explicit complete MUB set from Ref. [36] for n=2 (d=4). Fix ℓ1=0, ℓ2=1, and O=(|0⟩+i|1⟩)/√2 with O0=O-I/4. For each of the 4 nontrivial MUB unitaries U_j, compute the values tr(O0 U_j†|k⟩⟨k|U_j) for k=0,1,2,3 and verify that exactly 2 unitaries have all four squared values equal to 1/16 (the UO,1 class), while the other 2 have all four values zero (the UO,2 class). If this holds, check that the nontrivial contribution in Eq. (11) is at most (2^{n+1})(2^{n-1})(1/4^n)=1. If instead all 4 unitaries have nonzero squared values, the nontrivial contribution is 2 and Lemma 2's factor-1/2 claim is false. Repeat for d=8 to confirm the partition scales as 2^{n-1}/2^{n-1}.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central Theorem 2 bound, O(log K/ε²), is a direct consequence of classical shadow theory for Clifford measurements and does not depend on the MUB analysis; that part of the paper is robust. The load-bearing weak point is Lemma 2, which claims biased MUB measurements reduce the sample complexity to roughly one half of the Clifford value. The proof hinges on the assertion that, for every observable O in the set O, the nontrivial MUB unitaries split into two equal-size classes UO,1 and UO,2 based on the relative phase between the two computational-basis components. The text says only \"Due to the symmetric structure of MUB states expression\" and provides no formal argument. This partition is not obvious; in dimension 4 it is easy to imagine all nontrivial bases contributing interference terms, which would double the nontrivial variance contribution in Eq. (11) and invalidate the bound ∥O0∥²_shadow < 3/2. The claim also depends on same-author references [35] and [36], one of which is an unpublished preprint. This concern does not threaten the asymptotic O(d³ log d/ε²) tomography result, since the general t-sparse bound gives an O(1) shadow norm even if the factor-1/2 claim fails, but it does threaten a prominently advertised practical and constant-factor contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DMP-ST, a classical-shadow-based protocol for direct estimation of density matrix elements. The central claims are: (i) estimating K off-diagonal elements of a d-dimensional density matrix to additive error epsilon requires only O(log K / epsilon^2) samples; (ii) full state tomography with trace distance error at most epsilon can be achieved with O(d^3 log d / epsilon^2) samples; and (iii) a biased mutually-unbiased-bases (MUB) measurement scheme reduces the sample complexity by an additional constant factor of about one half relative to random Clifford measurements. The reconstruction uses the identity that each off-diagonal element rho_jk is a linear combination of expectation values of three rank-one projectors, together with the classical shadow sample-complexity theorem. The Clifford-based O(log K / epsilon^2) result and the d^3 log d tomography scaling follow from standard shadow tomography plus a Frobenius-to-trace-norm bound.","tokens_in":13639,"tokens_out":8751,"duration_ms":79394,"significance":"If correct, the paper gives a clean and practically relevant application of classical shadow tomography: a single shared dataset can be reused to estimate many off-diagonal density matrix elements with logarithmic overhead, and the resulting full tomography protocol matches the best known single-copy upper bound. The Clifford-based part is a direct and sound application of Huang--Kueng--Preskill, and the trace-distance reduction in Theorem 3 is a standard Frobenius-norm argument. A rigorous proof of the claimed MUB factor-one-half improvement would be a genuine constant-factor advance in measurement efficiency, and the proposed H--S--CZ implementation is attractive for near-term devices. The main weaknesses are that the MUB factor claim rests on an unproved partition of MUB unitaries, and that the tomography protocol as stated does not produce a physical density matrix without an additional projection step.","major_comments":[{"comment":"The factor-1/2 MUB reduction rests on the assertion that, for every O in the set O, the nontrivial MUB unitaries split into two equally sized classes U_{O,1} and U_{O,2} according to the relative phase between the two computational-basis components. The text supports this only with the sentence \"Due to the symmetric structure of MUB states expression,\" followed by an example; no rigorous counting argument is given. This partition is load-bearing: if it fails, the nontrivial-MUB contribution in Eq. (11) can be twice as large and the bound ||O_0||^2_shadow < 3/2 is not established. Please provide a complete proof of the partition, or state it as a lemma with a full derivation; an explicit check for d=4 would be a useful sanity test. The proof also relies on Results 6(i) and 6(ii) of the unpublished preprint [36]; these should be stated explicitly and proved, or replaced by a peer-reviewed reference.","section":"III.B, Lemma 2"},{"comment":"The estimator rho_tilde defined by entrywise estimates of rho_jk is not guaranteed to be positive semidefinite or trace-one, so it is not a density matrix and D_tr(rho, rho_tilde) is not literally a trace distance between quantum states. Since Theorem 3 is presented as full quantum state tomography, the protocol should include a projection of the Hermitian estimate onto the set of density matrices, together with a proof that the trace-distance bound is preserved (a Frobenius-norm projection suffices given the current bound), or the theorem should be restated as a trace-norm approximation by a Hermitian matrix rather than as a tomographic reconstruction of a quantum state.","section":"IV.C, Theorem 3"}],"minor_comments":[{"comment":"There are several typos: \"notrivial\" in Section III.B, \"meaasurements\" in Section IV.C, and \"closed to\" instead of \"close to\" in the abstract and conclusion; these should be corrected.","section":"Throughout"},{"comment":"The proof invokes Eq. (B12) of [35] without restating its assumptions; since the prefactor 2^{n+1} in Eq. (11) is essential to the final bound, please state the variance formula and the biased sampling model explicitly in a self-contained way, or give a precise pointer to the exact statement in [35].","section":"III.B, Lemma 2"},{"comment":"The text says DMP-ST requires O(log(K/δ)/epsilon^2) distinct random unitary settings, but since each sample uses a freshly drawn unitary, this is the same as the number of samples; the phrase \"measurement configurations\" may misleadingly suggest a distinct fixed device setting per unitary, so the terminology should be clarified.","section":"III.C"},{"comment":"Reference [36] is an arXiv preprint; if it is the source of a load-bearing technical statement, please cite the published version if one exists, or include the relevant results explicitly in the present paper.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The central asymptotic O(log K/epsilon^2) result is a direct consequence of the classical shadow framework once one observes that the relevant rank-one projectors have constant shadow norm; the mathematical core is therefore sound and incremental. The advertised MUB factor-one-half improvement depends on two same-author references, one of them an unpublished preprint, and on an unproved partition claim; the editor may want to ensure this is fully justified before publication. The physicality issue in Theorem 3 is fixable but should be addressed explicitly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my take on Wang's DMP-ST paper. The Clifford part is essentially correct and useful as an explicit protocol, but it is a direct application of the classical shadow framework to rank-one projectors. The O(log K / ε²) bound for K off-diagonal entries follows immediately from Huang–Kueng–Preskill once you note the shadow norm of such projectors is O(1). The paper is honest that this is the mechanism, though it could say more plainly that this is a corollary rather than a new theorem. The diagonal estimation bound (Theorem 1) is just Hoeffding, and the trace-distance conversion in Theorem 3 is a standard Frobenius argument; both are fine.\n\nThe genuinely new piece is the claim that biased MUB sampling halves the sample complexity due to a shadow norm of 3/2 for observables of the form |j⟩±|k⟩ and |j⟩±i|k⟩. This is where I have real concerns. First, the biased distribution defined in Eq. (9) does not normalize: if there are 2^n nontrivial MUB bases, the probabilities sum to 1/2 + 2^n/(2^n+1) ≈ 1.5, so the protocol as written is not a probability distribution over measurements. That alone invalidates the reconstruction channel and the variance calculation in Eq. (11). Second, the partition of MUB unitaries into U_{O,1} and U_{O,2}, each of size 2^{n-1}, is asserted with “Due to the symmetric structure” but no proof; it is load-bearing for the 3/2 bound. The reliance on the author's own refs [35] and [36], one an unpublished preprint, makes verification harder. I suspect the factor-1/2 claim may be salvageable with a corrected distribution and a proper argument, but as written it is not established.\n\nThe paper does have value: the Clifford-based protocol for directly estimating many off-diagonal elements from one dataset is clean, the reuse argument is explicit, and Theorem 3 is a pleasant, if not deep, packaging of entrywise tomography with trace-distance guarantees. That part deserves a serious referee. The MUB lemma, however, needs to be either removed or completely reworked.\n\nMy recommendation: send to peer review, but with the expectation of major revision on the MUB portion. The Clifford core will survive; the advertised constant-factor improvement will not, unless the author fixes the normalization and supplies a rigorous partition proof.\n\nWould I bring it to reading group? Maybe, if we want to discuss how much of shadow tomography is just \"plug in the right observables.\" I wouldn't cite it in its current form.","headline":"Clifford half is correct but derivative; the MUB constant-factor claim has an unnormalized distribution and an unproved partition, so the advertised improvement is not yet supported.","tokens_in":14160,"tokens_out":5910,"would_cite":false,"duration_ms":47174,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that $K$ off-diagonal density-matrix entries can be estimated from one shadow dataset with $O(\\log K/\\epsilon^2)$ samples","keywords":["classical shadow tomography","direct measurement protocols","density matrix elements","off-diagonal coherences","mutually unbiased bases","sample complexity","quantum state tomography","trace distance"],"falsifier":"Take $d=4$ or $d=8$, fix a target observable $O=|\\phi\\rangle\\langle\\phi|$ with $|\\phi\\rangle=(|\\ell_1\\rangle+i|\\ell_2\\rangle)/\\sqrt{2}$, and enumerate all nontrivial MUB projectors in Eq. (11), counting how many unitaries fall in $U_{O,1}$ and $U_{O,2}$; if the counts are unequal or the computed variance sum exceeds $3/2$ for some state $\\sigma$, the factor-two MUB claim fails.","tokens_in":13163,"feed_emoji":"⚛️","tokens_out":10869,"duration_ms":87951,"temperature":0.7,"pith_summary":"The paper's goal is to make direct estimation of density-matrix entries a global, reusable procedure rather than a per-entry experiment. It proves that a single classical shadow dataset, built from random Clifford or biased mutually-unbiased-basis measurements, determines any set of $K$ off-diagonal entries to additive error $\\epsilon$ using $O(\\log K/\\epsilon^2)$ samples. It also proves a full state-tomography version: estimating every entry with error $O(\\epsilon/d^{3/2})$ gives trace-distance error at most $\\epsilon$ with sample complexity $O(d^3\\log d/\\epsilon^2)$, near the tight single-copy lower bound. A reader should care because entrywise reconstruction with this scaling would make high-dimensional state verification practical and remove the post-selection losses that limit older direct measurement protocols.","feed_headline":"One dataset recovers K off-diagonal entries in O(log K) samples","feed_subtitle":"Shadow tomography makes off-diagonal estimation logarithmic in K and full tomography near-optimal in samples.","key_machinery":"The carrying object is the classical shadow inverse channel and its shadow norm. For each measurement ensemble the protocol estimates observables via $M^{-1}(U^{\\dagger}|k\\rangle\\langle k|U)$ and applies median-of-means; the sample count is set by $\\max_i \\|O_i\\|^2_{\\mathrm{shadow}}$. For the observables $\\{|\\phi_{jk}\\rangle\\langle\\phi_{jk}|, |\\psi_{jk}\\rangle\\langle\\psi_{jk}|\\}$, the Clifford shadow norm is bounded by $3(1-1/2^n)$. For the biased MUB ensemble, with the computational basis chosen with probability $1/2$ and the other $2^n$ MUBs uniform, the proof splits the nontrivial MUB unitaries into two equal classes $U_{O,1}$ and $U_{O,2}$ by relative phase between the two support components; tracked through the variance formula, this yields shadow norm below $3/2$, the entire source of the factor-two saving and of the simpler $H$--$S$--$CZ$ circuit implementation.","core_discovery":"The central discovery is that off-diagonal density-matrix elements are not structurally harder to estimate in bulk than diagonal ones. Any entry $\\rho_{jk}$ can be recovered from projections onto $|j\\rangle$, $|k\\rangle$, $(|j\\rangle+|k\\rangle)/\\sqrt{2}$, and $(|j\\rangle+i|k\\rangle)/\\sqrt{2}$, so the estimation task becomes a shadow-norm question about rank-one observables supported on two computational basis states. For random Clifford measurements the shadow norm of each such observable is at most $3(1-1/d)$, giving $O(\\log K/\\epsilon^2)$ samples for $K$ entries by median-of-means. For biased MUB measurements the paper derives a shadow norm below $3/2$, a constant-factor improvement that halves the sample count. Setting the per-entry error to $\\epsilon_1 = 2\\epsilon/d^{3/2}$ converts entrywise estimates into a full density matrix with trace distance at most $\\epsilon$ using $O(d^3\\log d/\\epsilon^2)$ samples, with Clifford or MUB circuits and lightweight post-processing instead of global convex optimization.","pith_inferences":["Beyond the paper, the same shared-data reasoning applies to any sparse set of linear observables: the $\\log K$ overhead is generic to classical shadows, so the entrywise scaling is not special to the four-projector representation.","The proof asserts rather than demonstrates the equal partition of MUB unitaries; a numerical enumeration for a representative observable would settle that subclaim without affecting the main $O(\\log K)$ scaling.","If the trace-distance protocol holds with reasonable constants, it offers a calibration and verification routine for near-term devices that avoids semidefinite solvers, which could matter when classical compute is the bottleneck."],"forward_implications":["For any collection of $K$ target coherences, only $O(\\log K/\\epsilon^2)$ copies of $\\rho$ are needed, so the data cost stops growing linearly with the number of entries one wants.","Full tomography to trace distance $\\epsilon$ is achieved with $O(d^3\\log d/\\epsilon^2)$ samples and no convex-optimization reconstruction step.","The number of distinct measurement settings needed for entrywise reconstruction of all $d^2$ entries drops from $\\Omega(d)$ for conventional direct protocols to $O(\\log d/\\epsilon^2)$ random settings, an exponential reduction in settings when $K$ is large.","Once the dataset is collected, any additional off-diagonal element can be estimated from the same data without new measurements.","Biased MUB sampling offers a constant-factor (roughly two) reduction in sample count and uses simpler one- and two-qubit gate circuits than general Clifford circuits."],"supporting_citations":[{"why":"Supplies the classical shadow framework, the inverse channel, and the $O(\\log(K/\\delta)/\\epsilon^2 \\max_i \\|O_i\\|^2_{\\mathrm{shadow}})$ sample bound used throughout.","marker":"[19]"},{"why":"Provides the four rank-one observables whose expectation values reconstruct any off-diagonal element, which is Lemma 1 of the paper.","marker":"[34]"},{"why":"Supplies the biased-MUB shadow-norm machinery and the MUB-sparse observable framework used to bound the variance for these two-sparse observables.","marker":"[35]"},{"why":"Supplies the MUB structural facts about evenly distributed amplitudes and the $H$--$S$--$CZ$ implementation used to split the unitaries and simplify circuits.","marker":"[36]"},{"why":"Gives the approximate 4-design sample complexity $O(d r^2 \\log d/\\epsilon^2)$ that the full-tomography result aims to match with simpler measurements.","marker":"[28]"},{"why":"Establishes the tight single-copy lower bound $\\Omega(d^3/\\epsilon^2)$ that the trace-distance tomography claim approaches.","marker":"[29]"}],"fun_headline_variants":["Shadow tomography: O(log K) samples for K density matrix entries","Exponential improvement in off-diagonal quantum state estimation","Near-optimal full tomography from one shadow dataset","One shadow estimate yields all off-diagonal density-matrix entries","Biased MUBs cut shadow-tomography samples by a constant factor"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The halving claim for biased MUB measurements rests on an unproved symmetry assertion that the nontrivial MUB unitaries split into two equal-size classes for each target observable.","fun_headline_variants_meta":{"raw":{"variants":["Shadow tomography: O(log K) samples for K density matrix entries","Exponential improvement in off-diagonal quantum state estimation","Near-optimal full tomography from one shadow dataset","One shadow estimate yields all off-diagonal density-matrix entries","Biased MUBs cut shadow-tomography samples by a constant factor"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00089,"raw_usage":{"total_tokens":3873,"prompt_tokens":1009,"completion_tokens":2864,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":625,"completion_tokens_details":{"reasoning_tokens":2779}},"tokens_in":625,"tokens_out":2864,"duration_ms":21184,"temperature":1.0,"reasoning_tokens":2779,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:21:16.003683+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $d=4$ or $d=8$, fix a target observable $O=|\\phi\\rangle\\langle\\phi|$ with $|\\phi\\rangle=(|\\ell_1\\rangle+i|\\ell_2\\rangle)/\\sqrt{2}$, and enumerate all nontrivial MUB projectors in Eq. (11), counting how many unitaries fall in $U_{O,1}$ and $U_{O,2}$; if the counts are unequal or the computed variance sum exceeds $3/2$ for some state $\\sigma$, the factor-two MUB claim fails.","supporting_citations":[{"cited_title":"Vallone and D","cited_arxiv_id":null,"evidence_quote":"Supplies the classical shadow framework, the inverse channel, and the $O(\\log(K/\\delta)/\\epsilon^2 \\max_i \\|O_i\\|^2_{\\mathrm{shadow}})$ sample bound used throughout."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the four rank-one observables whose expectation values reconstruct any off-diagonal element, which is Lemma 1 of the paper."},{"cited_title":"Horodecki, P","cited_arxiv_id":null,"evidence_quote":"Supplies the biased-MUB shadow-norm machinery and the MUB-sparse observable framework used to bound the variance for these two-sparse observables."},{"cited_title":"Ringbauer, T","cited_arxiv_id":null,"evidence_quote":"Supplies the MUB structural facts about evenly distributed amplitudes and the $H$--$S$--$CZ$ implementation used to split the unitaries and simplify circuits."},{"cited_title":"Seif, Z.-P","cited_arxiv_id":null,"evidence_quote":"Gives the approximate 4-design sample complexity $O(d r^2 \\log d/\\epsilon^2)$ that the full-tomography result aims to match with simpler measurements."},{"cited_title":"Jnane, J","cited_arxiv_id":null,"evidence_quote":"Establishes the tight single-copy lower bound $\\Omega(d^3/\\epsilon^2)$ that the trace-distance tomography claim approaches."}],"review_version":1}