{"id":"ef280300-9604-4fbf-aae6-abe2e938d9c0","arxiv_id":"2501.07891","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A new QPCA algorithm replaces quantum phase estimation with a block-encoding and quantum power method, with complexity depending on the eigenvalue gap rather than the largest eigenvalue.","lead":"This paper proposes a new quantum algorithm for principal component analysis that combines density matrix exponentiation with quantum singular value transformation and a quantum power method, claiming better performance when the largest eigenvalues are well separated. It also proposes a protocol for preparing covariance matrices from classical data, including uncentered datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Density matrix exponentiation yields a non-unitary channel, not the unitary exp(-iρt); Lemma 2's requirement of a unitary U is therefore unmet, so the block-encoding step (Lemma 4) and Theorem 1 are unsupported.","rationale":"The reader's weakest assumption identifies the same load-bearing flaw: Lemma 1 produces a channel, not a unitary. The trace in Eq. (1) makes the reduced map on σ a general CPTP map, and the purity counterexample shows it is not a unitary channel. Even if the channel is close to unitary for small Δt, Lemma 2 (Corollary 71 of [43]) is a statement about unitary operators, and QSVT requires block-encoded matrices implemented by unitaries. No dilation is supplied that yields a unitary acting on the target while preserving the coherence needed for controlled-U and QSVT. Since Lemma 4 and Theorem 1 depend on this, the complexity claims are unsupported. I also note the separate flaw in the covariance preparation (the garbage term in Eq. (14) prevents Lemma 9 from being applied exactly), but it is not the central claim. The proposed test is a direct check of non-unitarity and settles the matter.","tokens_in":15132,"tokens_out":9832,"duration_ms":107815,"concrete_test":"Implement the circuit from Eq. (1) for d=2 with ρ=I/2 and σ=|0⟩⟨0|; compute the reduced state after one time step. Its purity is (c² + s²/2)² + (s²/2)² < 1 for any t≠0, so the map is not a unitary channel; equivalently, its Choi matrix has rank > 1. Since Lemma 2 is only defined for a unitary U, this directly falsifies the assertion that Lemma 1 provides the unitary needed by Lemma 4. Re-deriving Lemma 4 with the actual channel will show that QSVT cannot be applied as written.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Lemma 4 is the keystone of Theorem 1: it claims that O(1/ε²) copies of ρ yield an ε-approximated block encoding of πρ/4 by combining Lemma 1 with Lemma 2. The problem is that Lemma 1 does not implement a unitary. Its output is the CPTP channel Λ_t(σ) = Tr_1[e^{-iS t}(ρ⊗σ)e^{iS t}], obtained by tracing out the copy of ρ. This channel is not a unitary conjugation whenever ρ is not pure. For example, take ρ = I/2 and σ = |0⟩⟨0|; one time step gives c²|0⟩⟨0| + (s²/2)I with c=cos t, s=sin t, which is mixed for t≠0, whereas a unitary channel preserves purity. Lemma 2 (logarithm of a unitary) requires a unitary U = exp(-iH) with controlled-U and U† as primitives. The circuit from Lemma 1 acts jointly on the target and the consumed copy; there is no unitary on the target system alone that equals the implemented map, and the copy cannot be uncomputed or discarded inside a coherent QSVT routine. Consequently the assertion 'as any unitary trivially block encodes itself' does not apply, and no block encoding of πρ/4 follows from the stated procedure. Theorem 1 inherits this gap because it feeds the purported block encoding into the quantum power method. Thus the central claim is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a new quantum algorithm for principal component analysis (QPCA). The central idea is to use density matrix exponentiation to obtain an approximation to exp(-iρt), then use the quantum singular value transformation (QSVT) framework to construct a block encoding of πρ/4, and finally apply a quantum power method (Lemma 6) to extract the largest eigenvalue and eigenvector of ρ. The claimed complexity of Theorem 1 depends on the spectral gap γ between the two largest eigenvalues and improves on the original QPCA in certain regimes. The paper also proposes a method for preparing a covariance matrix from classical data, including the contribution from the centroid μμ†, using block-encoding techniques. The manuscript is primarily a complexity-theoretic argument with no numerical simulations or machine-checked proofs.","tokens_in":15416,"tokens_out":8241,"duration_ms":79184,"significance":"If the central claims were valid, the algorithm would be a useful complement to the original QPCA, particularly when the largest eigenvalue gap is large, and the covariance-matrix preparation would address the centroid term that was not fully handled in prior work. However, the significance is undermined by a fundamental technical gap: the procedure cited as implementing exp(-iρt) actually implements a non-unitary quantum channel, and the paper does not provide a coherent unitary block encoding of πρ/4. Additionally, the covariance-matrix construction contains a concrete normalization error. The paper honestly compares its complexity with existing work, but the main theorem and the covariance-preparation lemma are not established by the presented arguments.","major_comments":[{"comment":"The construction of the block encoding of πρ/4 relies on Lemma 1 providing a unitary U = exp(-iρ/2) that can be used in Lemma 2 (logarithm of a unitary). However, Lemma 1, as derived in Section II (Eqs. (1)-(3)), implements the CPTP channel Λ_t(σ) = Tr_1[e^{-iS t}(ρ⊗σ)e^{iS t}], which is not a unitary conjugation on the target system whenever ρ is mixed. For example, with ρ = I/2, the channel maps a pure state to a mixed state for t ≠ 0, whereas a unitary channel preserves purity. Lemma 2 explicitly requires a unitary U with controlled-U and U† as primitives. The paper does not provide a coefficient or an uncomputation argument to turn this channel into a coherent unitary acting on the target system alone. Consequently, the assertion in Section III that 'any unitary trivially block encodes itself' does not apply to the output of Lemma 1, and the block encoding of πρ/4 is not established. Since Theorem 1 feeds this purported block encoding into the quantum power method (Lemma 6), the central claim is unsupported.","section":"Section III, Lemma 4 and Theorem 1"},{"comment":"The covariance-matrix preparation contains a concrete normalization error. In Eq. (14), Up is a unitary block encoding of Σ_i p_i U_i, so its action on |0⟩|0⟩ is |0⟩(Σ_i p_i |x_i⟩) + |Garbage⟩, where the total state is normalized. The vector |Φ⟩ defined in Eq. (15) as |0⟩(Σ_i p_i |x_i⟩) is generally not normalized, so no unitary can generate it. Moreover, Lemma 9, as stated in the manuscript, constructs a block encoding of the reduced density matrix ρ = Tr_A |Φ⟩⟨Φ|, not of the full projector |Φ⟩⟨Φ|; the text incorrectly says 'construct the block encoding of |Φ⟩⟨Φ|'. The reduced density matrix of the state produced by Up is not equal to μμ†, because the |Garbage⟩ component contributes additional terms. Therefore the claimed block encoding of μμ† and the resulting covariance-matrix preparation in Lemma 8 are not correct as written.","section":"Section V, Eqs. (14)-(17) and Lemma 8"},{"comment":"The quantum power method Lemma 6, which is the core subroutine for extracting the largest eigenvalue and eigenvector, is quoted verbatim from the author's own preprints [45,50] and is not proved, derived, or independently verified in this manuscript. Theorem 1 inherits this dependence, so the paper's central claim is not self-contained. While it is common to build on prior work, the cited sources are not peer-reviewed, and the lemma statement is the main technical engine of the algorithm. The manuscript should either provide a full proof of Lemma 6 in an appendix or explicitly frame Theorem 1 as conditional on the validity of [45,50].","section":"Section III, Lemma 6 and Theorem 1"}],"minor_comments":[{"comment":"The second line of Eq. (1) writes 'exp(-ρΔt)' instead of 'exp(-iρΔt)'; this is a typo that could confuse the reader.","section":"Section II, Eq. (1)"},{"comment":"The error term '4dp ε/α' in Lemma 3 appears dimensionally inconsistent; in the standard QSVT statement the error is proportional to d√(ε/α), and the bound |P(x)| ≤ 1/2 should be checked against the source. The manuscript later sets δ to a constant, but the dependence of the approximation error on δ is not made explicit.","section":"Section III, Lemma 3"},{"comment":"The symbol M in '|0⟩^{⊗ log(M)}' is not defined; it should be the dimension of the ancilla or the number of qubits, and the notation is inconsistent with the rest of the paper.","section":"Section V, Eq. (14)"},{"comment":"The displayed form of Ui shows only the first column as |xi⟩; the remaining columns are not described, which makes the subsequent discussion of the 'first column of Σ p_i U_i' less clear.","section":"Section V, Eq. (13)"},{"comment":"The manuscript contains many typos and grammatical errors, such as 'alternatively new quantum framework' in the abstract and inconsistent use of † and T for adjoints. A thorough editing pass is needed.","section":"General"},{"comment":"References [45] and [50] are arXiv preprints. The manuscript should indicate whether these have been peer-reviewed, or include enough detail in an appendix so that the key lemma can be evaluated independently.","section":"References"}],"recommendation":"reject","confidential_remarks":"The paper's main technical claims are not supported by the presented arguments. The gap between the density-matrix-exponentiation channel and the unitary required by QSVT is a fundamental issue that would require a genuinely new construction to fix, not a local revision. The covariance-matrix section contains additional errors that further weaken the paper. I would also note to the editor that the manuscript relies heavily on the author's own non-peer-reviewed preprints, so the novelty and rigor of the proposed work are difficult to assess from this submission alone."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper's headline idea is a QPCA variant that swaps the original's dependence on the largest eigenvalue r1 for the spectral gap gamma, using QSVT to turn density-matrix exponentiation into a block encoding and then feeding it into the author's earlier quantum power method. The gap-versus-r1 comparison is genuinely new and useful: in the large-gap regime this would complement the original QPCA. The covariance-preparation section also has a nice intent, handling the centroid term explicitly rather than ignoring it.\n\nUnfortunately the central construction does not hold together. The bridge is Lemma 4: it claims that O(1/eps^2) copies of rho give an eps-approximated block encoding of pi*rho/4 by combining Lemma 1 (density-matrix exponentiation) with Lemma 2 (logarithm of a unitary). Lemma 1 does not produce a unitary. It produces the channel Lambda_t(sigma)=Tr_1[e^{-i S t}(rho tensor sigma)e^{i S t}], which is not unitary unless rho is pure. A unitary block encoding is exactly what Lemma 2 and the QSVT machinery require. The line 'as any unitary trivially block-encodes itself' does not apply because U=exp(-i rho/2) is never implemented as a unitary on the target system. This is a load-bearing gap: Theorem 1 inherits it.\n\nThere is also a concrete error in the covariance preparation. The unitary Up defined in Eq. (14) acting on |0>|0> gives |0>|mu> + |Garbage>. That's fine as a pure state, but when you apply Lemma 9 to block-encode Tr_A|Phi>