{"id":"cf10e028-b703-413b-a3b5-38ab52b45239","arxiv_id":"2502.03062","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The authors extend selective inference to frequency-domain change point detection, yielding valid p-values for changes that appear across multiple frequencies.","lead":"This paper gives a way to attach statistically valid p-values to change points found in a signal's frequency content, so engineers can tell real shifts from noise. It extends selective inference, a method that prevents using the same data twice, to detect changes shared across multiple frequencies.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"P_k in Eq. (12) is not an orthogonal projection for middle frequencies (c_sym=2), so Eq. (14) and Theorem 1's conditioning argument fail; the proof is repairable by using the real symmetric projection Re(P_k), but as written the uniform p-value guarantee does not follow.","rationale":"The paper's central claim is that Eq. (17) gives uniformly distributed selective p-values, so the detected frequency-domain change points control type I error. The reader's weakest assumption was the exactness of the parametric-programming truncation-region enumeration. I find a more fundamental, earlier failure: the matrix P_k in Eq. (12), built from complex STFT basis vectors, is not an orthogonal projection for the typical middle frequencies with c_sym=2. The proof of Theorem 1 in Appendix B.1 relies on P_k being idempotent and Hermitian to conclude that conditioning on Q(X) = Q(x) restricts X to the one-dimensional line a + bz. Since P_k² ≠ P_k for middle frequencies, the line-conditioning argument breaks and the truncated-χ distribution is not established. This directly undermines Theorem 2's uniformity statement, not merely the power improvement of the parametric search. The failure is concrete and checkable: for c_sym=2, P_k has eigenvalue 2 on span(v^{(d)}), and ||P_k X||² = 2 X^T P_k X, contradicting Eq. (14). I want to credit the paper on other points: the STFT-based construction is otherwise natural, the code is released, and the synthetic experiments are extensive. The flaw is likely fixable because Re(P_k), built from the real and imaginary parts of the v^{(d)} vectors, is a genuine real symmetric projection, and using it in place of the complex P_k may recover the intended proof. Because the main theorem as written is unsupported but the repair seems plausible and the experiments suggest the method may work, a conditional verdict is appropriate, requiring the authors to correct the projection definition and re-verify Theorem 1 and the empirical claims.","tokens_in":55,"tokens_out":24187,"duration_ms":264572,"concrete_test":"From the released repository, take a tested CP whose D_k contains one middle frequency (e.g., M=4, d=1, τ_pre=1, τ_k=2, τ_suc=3), assemble P_k from Eq. (12), and compute P_k² − P_k and X^T P_k X vs ||P_k X||²; both will fail (the norm-squared ratio will be 2). Then recompute the p-values using the corrected projection P = Re(P_k) in the parametric line search (a = (I−P)x, b = σ P x / ||P x||) and rerun the Figure 5 type-I-error experiment. If corrected p-values still control type I near α, the flaw is a fixable formulation error; if not, Theorem 2's uniformity claim is empirically false.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Equation (12) defines P_k = Σ_{d∈Dk} (a_len^{(d)} c_sym / M) v^{(d)} v^{(d)*⊤}. For a middle frequency, c_sym=2 and ||v^{(d)}||² = M/a_len^{(d)}, so P_k v^{(d)} = 2 v^{(d)}; hence P_k² v^{(d)} = 4 v^{(d)} ≠ P_k v^{(d)}. The assertion in Eq. (14) that P_k = P_k² and P_k = P_k^⊤ is therefore false outside the special cases d=0, M/2 where c_sym=1. Consequently the displayed equality T_k(X) = σ⁻¹||P_k X|| is also false; for a middle frequency ||P_k X||² = 2 X^T P_k X. This is not a notation quibble: Appendix B.1's proof of Theorem 1 derives X = a + bz from Q(X) = Q(x), which requires P_k to be a projection (P_k² = P_k and P_k Hermitian) so that (I-P_k)P_k = 0 and P_k b = b. Without idempotence, V(X) = V(x) does not imply that P_k X is parallel to b, the conditioning set is not the line a + ℝ b, and the truncated-χ distribution used in Eq. (17) is unsupported. The real symmetric part Re(P_k) = Σ α_d [Re(v^{(d)})Re(v^{(d)})^⊤ + Im(v^{(d)})Im(v^{(d)})^⊤] is a genuine projection on DFT blocks, so the intended result is likely salvageable; but the theorem as written does not prove it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper extends selective inference (SI) to change point detection in the frequency domain. Frequencies are obtained via short-time DFT; candidate change points are selected by a simulated-annealing heuristic that encourages alignment across frequencies, and each selected candidate is tested using a truncated-χ selective p-value. The authors claim exact type I error control (Theorem 2) and support this with synthetic experiments and a bearing dataset. The central construction represents the test statistic as the norm of a projection of the data, computes the truncation region by parametric programming along the line a+bz, and reports p-values that are shown empirically to control type I error while retaining power.","tokens_in":24913,"tokens_out":13588,"duration_ms":126353,"significance":"If the technical gaps are resolved, this is a timely and useful contribution: it is the first SI-based method for frequency-domain change point detection, it targets the practically important setting of changes that appear simultaneously across multiple frequencies, and it ships reproducible code and extensive experiments, including robustness checks for non-Gaussian and correlated noise. The general idea of conditioning on the SA selection event and removing over-conditioning by parametric programming is valuable and may transfer to other heuristic solvers. However, the current manuscript contains a load-bearing error in the definition of the projection matrix, so the main theorems are not established as written.","major_comments":[{"comment":"The projection matrix P_k in Eq. (12) is not an orthogonal projection for middle frequencies. For d not in {0, M/2}, c_sym=2 and ||v^(d)||^2 = M/a_len^(d), so P_k v^(d) = 2 v^(d), hence P_k^2 v^(d) = 4 v^(d) ≠ P_k v^(d). Consequently Eq. (14)'s assertion 'P_k = P_k^2, P_k = P_k^T' is false, and the displayed equality T_k(X)=σ^{-1}||P_k X|| does not hold. This invalidates the proof of Theorem 1 in Appendix B.1, which relies on (I−P_k)P_k=0 and on X=a+bz being a real line. The same complex-vs-real problem appears in Appendix C.1: the cost matrix C^{(d)}_{s+1:e} is complex Hermitian, but Eq. (23) treats a^T C b as a real coefficient e1, which is not generally true. The construction is repairable by defining the real symmetric projection \\tilde P_k = Σ_{d∈D_k} (a_len^(d)c_sym^(d)/M)(Re(v^(d))Re(v^(d))^T + Im(v^(d))Im(v^(d))^T), which satisfies \\tilde P_k^2 = \\tilde P_k and X^T \\tilde P_k X = Σ (a_len^(d)c_sym^(d)/M)|v^(d)T X|^2, and by redoing the conditioning algebra and quadratic inequalities with this matrix. As written, Theorems 1 and 2 are unproved.","section":"§4.1, Eq. (12)–(14); Appendix B.1"},{"comment":"The paper claims that Algorithm 3 computes the exact truncation region Z = {z : A(a+bz)=A(x)} by parametric programming, and this exactness is essential for Theorem 2's equality p_selective ~ Unif(0,1). However, no theorem or proof is given that the iterative interval-enumeration procedure terminates and recovers the whole region. The update 'S ← S ∪ Z_oc(a+bz) for z ∈ S_c' is not an implementable rule for a continuum of intervals; it is not specified how z is selected or how the breakpoints of A(a+bz) are characterized beyond the over-conditioned quadratic inequalities. If the computed region is only a subset of Z, the p-values are computed under over-conditioning and the claimed uniform property fails, although type I error control may remain conservative. This needs to be either rigorously proved or clearly stated as an approximate/exact-in-practice method.","section":"Appendix C.2, Eq. (24), Algorithm 3"},{"comment":"The paper repeatedly claims that 'the probability of the final detected CPs being false positives is theoretically guaranteed to be below the specified significance level' and that the final CP set is reliable. Theorem 2, however, provides a per-selection property for an individual selected candidate k, conditional on that candidate being selected; it does not control the probability that at least one of the multiple detected CPs is false. The experiments in §5.2 test only one randomly selected detected CP candidate per trial, so they do not measure the set-level false positive rate. The authors should clarify the exact nature of the guarantee (per-candidate conditional on selection) or incorporate a multiple-testing correction over the detected CPs and evaluate the corresponding error rate.","section":"§1, §4, §5.2"}],"minor_comments":[{"comment":"There is a typo: 'desrcribe' should be 'describe' in the first sentence of Section 2.1.","section":"§2.1"},{"comment":"'altough' should be 'although' in the paragraph on decrement of temperature.","section":"§3.4"},{"comment":"There are repeated misspellings: 'truely' should be 'truly' in the Figure 1 caption and figure descriptions in §5.3; 'T ype I Error Rate' in Figure 5 should be 'Type I Error Rate'.","section":"§5 and figures"},{"comment":"The sentence 'The results theoretically confirmed that our method provided an unbiased evaluation based on SI framework' overstates what theorems and simulations establish; 'theoretically' should be removed or the scope clarified, since the paper's theoretical guarantee is exactly what needs repair.","section":"§6"},{"comment":"The correlated-noise experiment is reported as failing to control type I error for larger ρ; this limitation is honest and should be retained, but it would be useful to state in the main text that the theoretical guarantee is restricted to i.i.d. Gaussian noise and that the real-data analysis relies on the estimated-variance variant whose validity is only empirical.","section":"Appendix D.3"}],"recommendation":"major_revision","confidential_remarks":"The projection error is a genuine mathematical flaw, not a matter of taste or convention; it is fixable, but the revised proof must use a real symmetric projection and revisit the conditioning algebra. In addition, the authors should either prove the exactness of the parametric-programming search or temper the exactness claim, and they should reconcile the set-level false-positive claim with the per-selection nature of Theorem 2."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has a load-bearing flaw in the central theorem, but the idea is new and the intended result is recoverable. The contribution is the first selective-inference p-values for frequency-domain change points, with aggregation across multiple frequencies via simulated annealing. That is a genuine first, and the empirical work is careful: null simulations, power comparisons against over-conditioning and Bonferroni, robustness to unknown variance and non-Gaussian noise, and honest reporting of the correlated-noise failure. They also ship code.\n\nThe problem is Theorem 1. The test statistic is presented as sigma^{-1}||P_k X|| with P_k claimed to be a projection. As defined in Eq. (12), it is not. For a middle frequency, c_sym=2 and ||v^{(d)}||^2 = M/a_len, so P_k v^{(d)} = 2 v^{(d)}; hence P_k^2 != P_k. The matrix is also complex, while the paper asserts P_k = P_k^T. Consequently Eq. (14) is false, and the Appendix B.1 proof that the conditioning set is a line a+Rb does not go through: U(X) and P_k X are not independent, b is not a 1-eigenvector, and the truncated-chi distribution is unsupported. The reader's conditional verdict is right.\n\nThat said, the result is salvageable with a real fix, not a cosmetic one. Define P_k using the real and imaginary parts of the v^{(d)} vectors: P~k = Sum (a_len c_sym/M)(Re v^{(d)} Re v^{(d)}^T + Im v^{(d)} Im v^{(d)}^T). For the DFT basis, Re and Im are orthogonal and equal-norm, so this is a genuine orthogonal projection onto the real subspace spanned by them. Its quadratic form matches X^T P_k X (the imaginary part of P_k is antisymmetric and drops out), so the test statistic is sigma^{-1}||P~k X||, trace is Sum c_sym, and the standard selective-inference argument works. The theorem needs this correction.\n\nA second, smaller soft spot: Algorithm 3's parametric programming enumeration is asserted to give the exact truncation region, but there is no proof of termination or that the over-conditioned regions cover the full selection region. With a fixed seed, the simulated-annealing process can change candidate sets at points that are not just Metropolis inequality roots. The method may well be correct, since it follows Duy & Takeuchi 2022, but the paper should argue this explicitly.\n\nThis paper is for readers in selective inference and frequency-domain change point detection. The novelty justifies serious refereeing. Send it, but make clear that the projection issue must be fixed before the validity claim can be accepted.","headline":"New and useful extension of SI to frequency-domain change points, but the central projection matrix is not a projection; the theorem is repairable with a real fix.","tokens_in":25428,"tokens_out":9733,"would_cite":false,"duration_ms":79025,"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":"Frequency-domain change points can now be tested with p-values that control false positives exactly, even when detection relies on heuristic search.","keywords":["frequency-domain change point detection","selective inference","valid p-values","discrete Fourier transform","truncated chi distribution","simulated annealing","parametric programming","type I error control"],"falsifier":"Run the method on many independent null sequences (known $\\sigma$, no true spectral change), record the selective p-value for the first detected candidate, and test uniformity; if $P(p \\le 0.05)$ deviates from 0.05 beyond Monte Carlo error, the exactness claim fails. A more direct check: for a fixed observed sequence, evaluate $A(a + bz)$ on a fine grid of $z$ and compare the union of returned over-conditioned intervals against the set of $z$ with $A(a+bz)=A(x)$; any mismatch disproves the truncation-region claim.","tokens_in":24283,"feed_emoji":"📊","tokens_out":6261,"duration_ms":53562,"temperature":0.7,"pith_summary":"The paper's goal is to attach a statistically valid p-value to every change point detected in the frequency domain of a time series. It argues that selective inference—conditioning on the very act of selecting a candidate—removes the double-dipping bias that makes naive p-values near zero even when nothing changed. The proposed method constructs a test statistic from the discrete Fourier transform, conditions on the full heuristic search that produced the candidate, and derives an exactly uniform null p-value. If this is right, engineers who monitor rotating machinery or biosignals can flag a spectral shift as real only when the number is small, with false-positive rate controlled at the chosen level.","feed_headline":"Frequency change points now carry exact p-values","feed_subtitle":"Selective inference conditions on the heuristic search, keeping false-positive rates at the chosen alpha.","key_machinery":"The load-bearing object is the test statistic $T_k(X) = \\sigma^{-1}\\|P_k X\\|$, where $P_k$ is the orthogonal projection onto the subspace spanned by DFT vectors $v^{(d)}$ for the frequencies $d \\in \\mathcal{D}_k$ at which the candidate $\\tau_k$ was detected, weighted by segment lengths and the symmetry constant $c^{(d)}_{\\mathrm{sym}}$. Conditioning on the sufficient statistic $Q(X) = (V(X), U(X))$ and on the selection event $A(X)=A(x)$ reduces the null distribution of $T_k$ to a truncated $\\chi$-distribution with $\\mathrm{tr}(P_k)$ degrees of freedom and a truncation region $Z$. The region $Z$ is the exact set of $z$ for which the heuristic returns the observed candidate set; Algorithm 3 finds it by parametric programming, walking along the parametrized line and gluing together the over-conditioned intervals where every Metropolis decision is fixed.","core_discovery":"The central claim is Theorem 2: the selective p-value defined in Eq. (17) satisfies $P_{H_{0,k}}(p^{\\mathrm{selective}}_k \\le \\alpha \\mid A(X)=A(x)) = \\alpha$ for every $\\alpha \\in (0,1)$, and hence also unconditionally, so the type I error rate is exactly the nominal level. In plain terms, whenever the null hypothesis holds—no mean shift in any of the frequencies where a candidate was detected—the reported p-value is uniform, and the false-positive probability among reported change points is controlled. The discovery is that this uniformity survives even though the candidate set is produced by a simulated-annealing heuristic, because the line search of Algorithm 3 reconstructs the exact selection event by unioning over-conditioned regions along the one-dimensional slice $X = a + bz$.","pith_inferences":["The line search's exactness is the linchpin; a cheap validation would be to compare Algorithm 3's $Z$ against a dense grid of $z$ values on real null data to confirm no interval is missed.","The same projection-and-condition recipe could test changes in phase, power, or cross-frequency coherence, since it only requires the statistic to be the norm of a Gaussian projection.","Correlated noise breaks the guarantee, as the paper's own experiments show; users should whiten or estimate the covariance before trusting $\\alpha$.","Multivariate extension to multiple sensors is the natural next step—the DFT basis already handles complex spectra, so stacking sensors only changes the projection matrix."],"forward_implications":["Reported frequency-domain change points carry p-values that are exactly uniform under the null, so a user who rejects at level $\\alpha$ will see a false positive with probability at most $\\alpha$.","The method identifies which frequencies actually shifted at a change point, enabling root-cause diagnosis rather than a single 'something changed' flag.","Because the selection-conditional p-value conditions on the heuristic process, the guarantee holds even though the candidate search is only approximately optimal.","Over-conditioned competitors control the error rate but waste power; the exact truncation region restores power while keeping the guarantee.","The conditioning machinery transfers to any optimizer whose decision boundaries along the slice are quadratic, which includes many simulated-annealing and piecewise-linear heuristics."],"supporting_citations":[{"why":"Establishes the selective-inference principle of conditioning on the selection event to obtain valid p-values.","marker":"Taylor & Tibshirani, 2015"},{"why":"Provides the parametric-programming technique used to identify the exact truncation region in Algorithm 3.","marker":"Duy & Takeuchi, 2022"},{"why":"Supplies the dynamic-programming-based change point selective inference that is extended to the frequency domain.","marker":"Duy et al., 2020"},{"why":"Introduces post-selection inference for change point detection, the foundation this paper builds on.","marker":"Hyun et al., 2018"},{"why":"Provides the simulated annealing heuristic whose selection event is conditioned on to compute the p-values.","marker":"Kirkpatrick et al., 1983"},{"why":"Documents over-conditioning in selective inference, the comparison point for the exact line-search method.","marker":"Lee et al., 2016"}],"fun_headline_variants":["Exact p-values for frequency change points","Selective inference yields exact spectral CP p-values","Frequency change points with controlled false positives","Spectral shift detection with exact significance","Change points in frequency domain now statistically exact"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The p-values are exact only if the parametric line search of Algorithm 3 enumerates every interval of the slice where the simulated-annealing search returns the observed change-point set; if an interval is missed, the conditioning set is too large and the p-values stop being uniform.","fun_headline_variants_meta":{"raw":{"variants":["Exact p-values for frequency change points","Selective inference yields exact spectral CP p-values","Frequency change points with controlled false positives","Spectral shift detection with exact significance","Change points in frequency domain now statistically exact"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000269,"raw_usage":{"total_tokens":1578,"prompt_tokens":860,"completion_tokens":718,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":476,"completion_tokens_details":{"reasoning_tokens":653}},"tokens_in":476,"tokens_out":718,"duration_ms":7061,"temperature":1.0,"reasoning_tokens":653,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T10:03:44.082487+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the method on many independent null sequences (known $\\sigma$, no true spectral change), record the selective p-value for the first detected candidate, and test uniformity; if $P(p \\le 0.05)$ deviates from 0.05 beyond Monte Carlo error, the exactness claim fails. A more direct check: for a fixed observed sequence, evaluate $A(a + bz)$ on a fine grid of $z$ and compare the union of returned over-conditioned intervals against the set of $z$ with $A(a+bz)=A(x)$; any mismatch disproves the truncation-region claim.","supporting_citations":[{"cited_title":"Computing valid p-value for optimal changepoint by selective inference using dynamic programming","cited_arxiv_id":null,"evidence_quote":"Supplies the dynamic-programming-based change point selective inference that is extended to the frequency domain."},{"cited_title":"Exact post-selection inference for the generalized lasso path","cited_arxiv_id":null,"evidence_quote":"Introduces post-selection inference for change point detection, the foundation this paper builds on."}],"review_version":1}