acceptodds
Under review as a conference paper at ICLR 2027

When Is Residual EOF1 Certifiable?

Abstract

We study pipelines that compare an analysis window with a background reference and interpret the leading analysis mode as an anomaly—for example, a seismic event against ambient noise or a climate pattern against climatology. Write for the analysis-window covariance, for a quiet reference covariance in the same feature space, and for their background-adjusted contrast. We show that alone does not identify an anomaly: the same matrix admits pure-background and anomaly-plus-background decompositions, identical at second order when the anomaly lies along the background axis. Under partial contamination, the calibrated direction remains fixed while the whitened eigen-gap required for finite-sample certification shrinks to zero at same-record, where the direction is unidentified. Under independent sub-Gaussian sampling from the rank-one covariance model, a quiet held-out background recovers a residual spike above concentration scale, but that spike need not be EOF1 of . We prove observational equivalence and a sharp gap law, and introduce a Contrast Audit Procedure (CAP) that separates declared guard passes from conditional, theorem-backed direction certificates and otherwise abstains. An optional held-out-control calibration bounds per-record null returns under exchangeability without assuming independent guards. On the public INSTANCE benchmark, with station keys held out and guards frozen, both guards pass earthquakes and noise records. Known-direction injections into the held-out real-noise backgrounds reach accurate CAP returns at twice the frozen strength threshold; on real LOCATA recordings, frozen guards return sources versus inactive controls with median azimuth error. We release our code, prespecified splits, results, and reproduction instructions in our anonymous repository https://anonymous.4open.science/r/eof1 for replication.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
residual EOF1, covariance identifiability, eigen-gap certification, background covariance calibration, abstaining spectral inference

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_wCEqOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

From the paper: “We study pipelines that compare an analysis window with a background reference and interpret the leading analysis mode as an anomaly—for example, a seismic event against ambient noise or a climate pattern against climatology. Write Σ for the analysis-window covariance, C for a quiet reference covariance in the same feature space, and Σ − C for their background-adjusted contrast. We show that Σ alone does not identify an anomaly: the same matrix admits pure-background and anomaly-plus-background decompositions, identical at second order when the anomaly lies along the background axis. Under partial contamination, the calibrated direction remains fixed while the whitened eigen-gap required for finite-sample certification shrinks to zero at same-record, where the direction is unidentified. Under independent sub-Gaussian sampling from the rank-one covariance model, a quiet held-out background recovers a residual spike above concentration scale, but that spike need not be EOF1 of Σ. We prove observational equivalence and a sharp gap law, and introduce a Contrast Audit Procedure (CAP) that separates declared guard passes from conditional, theorem-backed direction certificates and otherwise abstains. An optional held-out-control calibration bounds per-record null returns under exchangeability without assuming independent guards. On the public INSTANCE benchmark, with station keys held out and guards frozen, both guards pass 240/299 earthquakes and 13/264 noise records. Known-direction injections into the 264 held-out real-noise backgrounds reach 98.1% accurate CAP returns at twice the frozen strength threshold; on real LOCATA recordings, frozen guards return 11/13 sources versus 1/13 inactive controls with 2.1◦ median azimuth error. We release our code, prespecified splits, results, and reproduction instructions in our anonymous repository for replication.”

Critical strengths

This is a very weird paper.

Critical weaknesses

My criticism is based on the following facts.

  1. The first paragraph of the paper reads as: “Background adjustment is meant to support a consequential decision: whether a reported direction is evidence of an event or merely structured background. A stable leading mode does not resolve that distinction. The reference determines what can be concluded. Analysis covariance alone and same-record calibration cannot certify the label. A quiet held-out reference supports a direction claim only after spike, gap, and sampling conditions clear under thresholds fixed in advance.”

The paragraphe has 5 sentences. Sentence 2: “A stable leading mode does not resolve that distinction.” What is “a stable leading mode”? Undefined and the sentence is hardly understandabl.

Sentence 3: “The reference determines what can be concluded”: What is the reference here? No definition or explanation of it has been given so far.

Sentence 4: “Analysis covariance alone and same-record calibration cannot certify the label.” What does “Analysis covariance” mean? Indeed the paper propose a lots of analyses on covariances: does this mean that none of the proposed analyses is useful? If so, what is the paper’s purpose?

Sentence 5: “A quiet held-out reference supports a direction claim only after spike, gap, and sampling conditions clear under thresholds fixed in advance.” The English use itself is problematic here. I cannot figure out the meaning of the verb-complemet combination in the complement “after …. clear under thresholds”.

Therefore in the very first opening paragraph with 5 sentences, four of them are unclear with undefined or ambigorous terms. This shows how low is the presentation quality of the paper.

  1. On the technical side: Lines starting from 108 describe 5 main contributions of the paper. The one reads as: “ (1) We formalise analysis/reference covariance decompositions (§2) and prove observational non-uniqueness with aligned collapse (§3).” I thus paid some attention to the “observation non-uniqueness result”, namely Proposition 3.1, stated from line 165 (page 4. Its proof is given on page 15, lines 767-776. All is about a simple calculation on a rank-1 perturbed spiked covariance matrix: it is at the level of a first undergraduate course in linear algebra. The result of the proposition is a simple linear algebra result. I cannot find any significant new insight compared to the abundant literature cited by the authors, e.g., in the two paragraphs in Introduction, lines 041-081 and lines 093-104, respectively.

This shows the low level of the claimed “main contributions” of the paper.

Questions for authors

Your AI statement: “In this work, we used generative AI tools to assist with grammar and readability, coding, and feedback on experimental design and the interpretation of results. The authors independently reviewed all AI-assisted contributions and take full responsibility for the methods, analyses, interpretations, and final manuscript.” I suspect that AI has wrote most of the presentation of the paper including literature search. For example, the first paragraph I analyzed in details above is written by AI to me. However, the authors failed to declare such AI use properly according to the conference requirement.

3 comments
AuthorsOpenReview ↗

Response to Reviewer wCEq

CLARIFICATION We rewrote the opening to introduce the window, background record and leading eigenvector before stating the inference question. The elementary example identifies missing background information. We show why recovering an added covariance direction and establishing its separation relative to background are different claims.

Here “analysis covariance” means the covariance of the window being examined, not the act of analysing a covariance matrix. A “reference” is a separate record intended to describe background variability. We removed “stable leading mode” and “clear under thresholds.” Under our rank-one model, a quiet reference identifies the contrast even when the window covariance alone cannot justify an anomaly label. Revised locations: Section 1 opening, p. 1, lines 036–041; Figure 1 caption, p. 2, lines 067–071; worked contrasts, p. 4, lines 182–184.

DIRECT ANSWER Proposition 3.1 is elementary. We use it as a minimal non-identifiability example, not evidence of technical difficulty by itself. We revised contribution (1) and the Section 3 opening to give it that role. The spiked background illustrates alignment and direction disagreement; the finite-sample bounds allow general C≻0C\succ0 under their stated conditions. We corrected Section 2's overly narrow description of where the example family is used (p. 3, lines 156–159).

For a concrete reading, Σ=diag(3,1)\Sigma=\mathrm{diag}(3,1) has leading direction e1e_1 both when C=ΣC=\Sigma (pure background) and when C=IC=I with a rank-one addition 2e1e1⊤2e_1e_1^\top. A quiet reference distinguishes zero contrast from that addition; more precise estimation of Σ\Sigma alone does not. The revised worked example states the two contrasts explicitly (p. 4, lines 182–184). The later bounds answer the next question: how much sampling error those reference-supported direction and separation claims tolerate under the stated model and independence conditions.

We retain cPCA at weight one, which extracts top(Σ−C)\mathrm{top}(\Sigma-C). We establish when a quiet reference identifies the added covariance direction and how sampling error limits direction and separation claims under our rank-one and independence assumptions. For analysts comparing event and background windows, this distinguishes recovering an added direction from resolving its separation relative to background. Our exact leakage law shows why an unchanged generalised direction can conceal vanishing separation. Our bounds account for covariance and whitening error for general positive-definite backgrounds, controlling the contrast and whitened directions separately. We separate fixed-cutoff returns from certificates with angle bounds; CAP withholds the direction when its reference or separation requirements fail. Optional exchangeable controls bound the null-return probability for one record.

We distinguish a conditional contrast-angle bound from measured azimuth error. In LOCATA, we compare returned azimuths with those derived from optical source positions. Revised locations: contribution (1), p. 3, lines 112–122; Section 3 opening, p. 4, lines 168–171; the interpretation in Section 7, p. 10, lines 502–539.

We test the resulting claims against event labels, injected directions and physical measurements. INSTANCE's frozen guards return 240/299240/299 held-out earthquakes and 13/26413/264 noise records, but those event labels do not validate directions. Known-direction injections supply that target on real backgrounds. LOCATA supplies a physical check: 11/1311/13 active recordings return versus 1/131/13 inactive controls, with 2.1∘2.1^\circ median error among returns against optical tracking. These endpoints show what the reporting rule can retain, while the matched comparisons preserve losses: Roy selects more INSTANCE events, and bootstrap has more accurate returns at CAP's frozen spike cutoff. The combined evidence supports a useful distinction between a reproducible variance pattern, an identified covariance addition and a physically accurate returned direction (Table 3, p. 10, lines 487–495). LOCATA has only thirteen recordings from one room and array, which limits transfer.

CLARIFICATION The submission form already disclosed writing assistance, retrieval/discovery and research ideation/execution. The manuscript statement was less explicit about the tasks. We expanded it to enumerate literature discovery and checking, conceptual and theoretical development and proof assistance, experimental design and interpretation, coding, synthetic-data simulation and data processing, and drafting or editing prose and displays. The revised statement also covers the current response preparation and retains author responsibility for the work. We address the identified writing defects directly; we do not use prose style as evidence for or against particular tool use. Revised location: AI use statement, p. 11, lines 540–547.

Reviewer_wCEqOpenReview ↗

Follow-up

Thank you for your clarification. However, there still exists a lot of similar clarity issues in the revised article which makes difficult to recommend acceptance.

AuthorsOpenReview ↗

Clarified task, directions and guarantees

CLARIFICATION Your follow-up prompted a closer check of the remaining terminology and conceptual links. We have prepared focused exposition revisions to the 11 November version. The problem, estimator, assumptions and numerical results are unchanged.

Our task is to recover a direction of added variation relative to a background record. Weight-one contrastive PCA supplies the covariance-difference direction. Under the rank-one model, we derive an exact law for gap loss as signal leaks into the reference while the variance-ratio direction stays fixed (Theorem 3.1). With independent sub-Gaussian samples and the stated reference conditions, we bound contrast and whitened-direction errors separately (Theorems 4.1 and 3.3). A failed gap check need not imply an inaccurate contrast direction.

WHAT WE CLARIFIED

  • Task and directions. To address the undefined and overloaded terms identified in your original review, we now place “preprocessed window covariance” beside the first use of EOF1 (abstract, p. 1, lines 010–014). Section 2 separately explains the window mode, contrast direction and variance-ratio direction, and gives the two actual contrasts in the existing diagonal example (p. 3, lines 126–151). We also distinguish reference quietness, holding observations out and sampling independence. The introduction now names the exact gap law and separate angle bounds (p. 2, Research gap) and groups the same scientific components around these questions (p. 3, lines 108–123).
  • Checks and guarantees. The earlier phrase “every return inherits the reference's side information and inferential scope” compressed the conditions and output. Section 5 now describes each threshold comparison and defines a margin as score minus threshold. It states the theorem's actual probability event: under its assumptions, the probability that CAP returns and an angle bound fails is at most the stated failure level; this is not coverage conditional on returning (p. 6, lines 270–288; Table 2, lines 297–305). The coordinate mapping after Theorem 3.3 makes clear that its angle bound concerns whitened vectors (p. 5, lines 229–230). We preserve the distinction between the spike-only contrast-angle bound and CAP's stronger separation requirement.
  • Evidence and physical meaning. Instead of “development-frozen declared-mode guard passes,” Section 6 says that we fix thresholds on development records and apply them unchanged to held-out records (p. 7, lines 339–347). The existing injection construction is now described in the main text (p. 8, lines 425–427). LOCATA's covariance features and largest-loading-to-azimuth mapping explain what the optical comparison measures (p. 9, lines 460–467). Figure 6 defines Etna's repeatability statistic and assigns the 0.5 cutoff to block-return rates (p. 9, lines 479–482). The abstract also makes the 98.1% denominator explicit: all 264 injected backgrounds, rather than only returned directions.

These revisions make the contribution directly assessable: how a reference identifies the added direction, how sampling error limits recovery, and why accurate recovery can coexist with weak background-relative separation. The physical comparison provides a concrete use case: CAP returns 11/13 active LOCATA recordings versus Roy's 7/13 at the same 1/13 inactive count, with 2.1° median error among CAP's returns (Table 3). The model-based bounds quantify direction error under their assumptions; optical tracking independently evaluates the selected acoustic directions.

Reviewer_bjnpOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Contrast Audit Procedure (CAP), a screening process for anomaly detection based on covariance matrix principal direction analysis. A rank-one model for the covariance decomposition into rank-one anomaly + background is analyzed. The paper shows that under certain conditions, with quiet background and sub-gaussian samples, the gap and residual spike can be controlled. CAP is built on these analysis, incorporating multiple guards against spike or gap. Experiments on both synthetic and real datasets are presented to illustrate the performance, and compare with baseline methods.

Critical strengths
  • The finite sample gap control results (Theorem 3.3, lemmas in the Appendix) are correct and clearly stated. There is some technical novelty in piecing the results together, however it's not so strong as to be called a "critical strength" in my opinion.
Critical weaknesses
  • The sample size required by the theoretical results is very large, which also shows in the simulations, see Fig 4(a). Even at n=2 dim, there is no certificate with 0.25M samples. This is lower than the 0.737M number from the Appendix Table 10, so the theory and simulations align, however this sample requirement is easily beyond what's typically available in real data, making the theoretical result hard to apply.
  • The simulation results comparing the proposed CAP with other methods in Table 3 do not suggest that CAP has better performance. This brings questions to whether the guards introduces are actually suitable for real data.
Questions for authors
  • What is mm in Algorithm 1? If it's "mode" in the "Require", state it clearly.
  • The "contamination" case is essentially a misspecification of the background that is completely non-identifiable, and trivially weakens the anomaly. Is this the only motivation for doing the gap guard?
1 comment
AuthorsOpenReview ↗

Response to Reviewer bjnp

CLARIFICATION We quantify the theorem's sufficient sample size per covariance and measure azimuth error among LOCATA's fixed-cutoff returns. Table 10 gives a sufficient planning size of about 0.740.74 million samples at n=2n=2; LOCATA returns 11/1311/13 active recordings versus 1/131/13 inactive controls, with 2.1∘2.1^\circ median error among returns. These are separate claims: the real-data studies use fixed cutoffs without nominal direction coverage. The planning bound comes from covariance and whitening errors, not an information-theoretic lower bound. We clarify that cost beside Figure 4; the Gaussian experiment still uses unchanged constants.

NEW EVIDENCE To explain the source of the cost, we added a spike-only column to Appendix Table 10: with the same radii and planning model, its n=2n=2 sufficient crossing is 13,470, compared with 737,373 for the joint report. The joint rule requires the more conservative whitening screen. Locations: Section 6, p. 7, lines 373–377, continuing on p. 8, lines 395–398; Figure 4, p. 8, lines 411–414; revised Table 10 and its explanation, p. 30, lines 1566–1588.

EXISTING EVIDENCE Our comparisons show different rankings for event selection, known-direction recovery and physical-direction selection. Table 3 includes real INSTANCE and LOCATA observations as well as known-direction injections into held-out real noise. At matched INSTANCE development-null counts, Roy returns more events (254/299254/299 versus 240/299240/299) at the same test-noise count (13/26413/264); log-det returns 248/299248/299 events with 7/2647/264 noise. Both return more labelled earthquakes with equal or fewer noise returns. They do not dominate the directional endpoints: at 1×1\times injection, accurate-return rates are .470.470 for CAP, .337.337 for Roy and .027.027 for log-det, while bootstrap is strongest at .939.939. On physically tracked LOCATA recordings, CAP returns 11/1311/13 active and 1/131/13 inactive recordings, versus Roy's 7/137/13 and 1/131/13; CAP's returned azimuths have 2.1∘2.1^\circ median error. At 2×2\times injection CAP and Roy both reach .981.981 accurate returns: their ranking changes with strength. Under the original separately frozen INSTANCE cutoffs, joint CAP also gives up ten events for one fewer noise return relative to gap-only. We show physical accuracy among LOCATA joint returns; our INSTANCE results quantify the event-selection costs of the same two-margin reporting requirement. All compared rules return the same contrast eigenvector after selection; they differ in which records they retain. For injections, 1×1\times means β=τspike\beta=\tau_{\mathrm{spike}}, and an accurate return has absolute cosine at least .9.9 with the injected direction, over all 264264 backgrounds. LOCATA instead uses optical source positions. Its active/null Wilson 95% intervals are [.578,.957][.578,.957] and [.014,.333][.014,.333]; this is a small physical panel. Locations: Table 3, p. 10, lines 487–495; the original matched comparisons and physical-direction evidence in Section 6, pp. 8–9, lines 424–434 and 451–463; revised interpretation in Section 7, p. 10, lines 519–535.

DIRECT ANSWER mm denotes the operating mode. Algorithm 1 now defines $m\in\lbrace \mathtt{none},\mathtt{declared},\mathtt{theorem}\rbrace $ in Require (p. 6, lines 309–310). Its reuse check now explicitly names input-window IDs (line 313); matching station/channel identifiers alone does not mean reusing the same observations.

DIRECT ANSWER The contamination path is not the only motivation for the gap screen; Theorem 3.3 also controls a distinct whitened eigendirection. The path models leakage into a reused reference. We do not infer the unknown contamination fraction from the covariance pair, and the path does not cover arbitrary background misspecification. Weakening the residual is expected. The exact result is that the generalised direction remains unchanged for every ε<1\varepsilon<1, while its dimensionless gap follows δ⋆(ε)=(1−ε)βQ/(1+εβQ)\delta_\star(\varepsilon)=(1-\varepsilon)\beta Q/(1+\varepsilon\beta Q) and vanishes at exact reuse. Thus a stable direction can conceal loss of background-relative separation. This motivates the extra screen, together with the finite-sample gap bound. It is not necessary for Theorem 4.1's residual-angle bound; we explicitly corrected that scope and proved the theorem-mode guard redundancy in Appendix Proposition A.2. The existing INSTANCE ablation shows that both separately frozen declared-mode thresholds affect decisions; LOCATA is gap-limited. Locations: the clarified leakage model in Section 2, p. 3, lines 146–154; the exact law in Theorem 3.1, p. 4, lines 188–196; the corrected guard interpretation in Section 5, p. 6, lines 284–290; the new implication proof in Proposition A.2, p. 20, lines 1053–1066. Section 7, p. 10, lines 514–530, discusses empirical guard tradeoffs and the limits of reference and threshold transport.

Reviewer_3pe1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies when a leading covariance direction can be interpreted as a signal relative to a background reference. It establishes non-identifiability without background information and derives an exact expression for how reference contamination reduces the generalized eigengap. Under a rank-one covariance model and independent sampling, it provides finite-sample bounds for estimating the residual direction. The proposed Contrast Audit Procedure checks reference provenance, residual strength, and a whitened eigengap before returning a direction. Experiments cover synthetic examples, seismic records, acoustic localization, and additional physical datasets.

Critical strengths

a) The paper makes a useful distinction between the leading variance direction, the direction of the covariance difference, and the generalized eigenvector, and illustrates why a dominant variance direction need not identify the signal of interest. The contamination formula clearly explains how spectral separation can disappear even while the generalized direction remains unchanged.

b) I also like the thorough experiments and that the authors report unfavorable comparisons and calibration failures since they can prove to be useful to build upon as well.

Critical weaknesses

Under the stated theorem thresholds, passing the whitened-gap guard appears to imply passing the spike guard on the concentration event. Meanwhile, Theorem 4.1 already bounds the returned residual direction using the spike condition alone. I understand that the authors intend to report both absolute strength and background-relative separation, but could they clarify what retaining both checks adds in the theorem, and why the additional whitened requirement is useful for certifying the returned contrast direction?

Questions for authors

see above.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 3pe1

DIRECT ANSWER On the concentration event, the theorem-mode gap condition implies the spike condition. We corrected the claim that both are independently required; the individual angle and separation bounds remain valid. The gap adds a background-relative reporting requirement, not a tighter bound for CAP's returned contrast direction. Theorem 4.1 needs only β>4η\beta>4\eta; Theorem 3.3 controls a different, whitened eigendirection. New Proposition A.2 proves the implication, and the added Table 10 column quantifies its planning cost. An analyst who needs only the contrast-angle bound can use Theorem 4.1 under its stated assumptions. Sections 4–5 now make the targets explicit (p. 5, lines 247–252; p. 6, lines 284–290).

CORRECTION TO THE PAPER You are right about the linked theorem thresholds. We overstated their independence as requirements. Under quiet C⪰μIC\succeq\mu I and η>0\eta>0, δ⋆=βv⋆⊤C−1v⋆≤β/μ\delta_\star=\beta v_\star^\top C^{-1}v_\star\le\beta/\mu, whereas ηM≥2η/μ\eta_M\ge2\eta/\mu. Hence δ⋆≥4ηM\delta_\star\ge4\eta_M implies β≥8η\beta\ge8\eta, which already meets Theorem 4.1's condition. The individual gap and angle bounds remain valid; the error was our statement that both conditions were independently needed in this theorem regime. We corrected Sections 4–5, the guarantee and its appendix argument, and the conclusion. This removes an overstated requirement; it preserves recovery of the contrast direction and control of the distinct whitened direction under their stated conditions.

NEW EVIDENCE New Appendix Proposition A.2 proves the population implication above and the observable counterpart (p. 20, lines 1053–1066). Write A=Σ^−C^A=\hat\Sigma-\hat C and s=∥PSD(A)∥ops=\lVert \mathrm{PSD}(A)\rVert _{\mathrm{op}}. Since A⪯sIA\preceq sI, a positive empirical gap satisfies δ^≤λ1(M^)−1≤s/λmin⁡(C^)\hat\delta\le\lambda_1(\hat M)-1\le s/\lambda_{\min}(\hat C). With ηˉ≤μ/4\bar\eta\le\mu/4 on the concentration event, λmin⁡(C^)≥3μ/4\lambda_{\min}(\hat C)\ge3\mu/4 and the observable covariance-norm upper bound is at least ∥Σ∥op≥μ\lVert \Sigma\rVert _{\mathrm{op}}\ge\mu. Thus ηˉM≥4ηˉ/μ\bar\eta_M\ge4\bar\eta/\mu, and passing δ^≥m0ηˉM\hat\delta\ge m_0\bar\eta_M with m0>6m_0>6 implies s≥3m0ηˉs\ge3m_0\bar\eta. The theorem-mode spike guard therefore cannot reject a gap-passing observation on that event. The algorithm and thresholds are unchanged.

We also added a spike-only column to Table 10 to quantify the cost of the stronger report. In the same quiet Gaussian planning model (C=IC=I, β=2\beta=2, t=3t=3, m0=6.01m_0=6.01), the spike-only sufficient crossing is 13,470 samples at n=2n=2, versus 737,373 for the joint report. Both use the same population plug-ins to the observable radii. These are sufficient planning sizes, not information-theoretic limits, empirical acceptance guarantees or evidence of real-data direction coverage. The comparison and its interpretation are on p. 30, lines 1566–1588.

CLARIFICATION We distinguish absolute residual strength from dimensionless background-relative separation. For example, let CL=I+(L−1)v⋆v⋆⊤C_L=I+(L-1)v_\star v_\star^\top and keep β\beta fixed. The contrast remains βv⋆v⋆⊤\beta v_\star v_\star^\top, with the same strength and direction, while δ⋆=β/L\delta_\star=\beta/L tends to zero. An accurately estimated covariance addition can therefore be small relative to background variation. Our analysis therefore supports reporting both quantities. This population example already appeared in Proposition A.1 (now p. 20, lines 1039–1048); it motivates the reporting choice rather than proving independent theorem requirements.

EXISTING EVIDENCE Declared-mode cutoffs are frozen separately and need not obey this theorem implication. The submitted INSTANCE ablation uses the same held-out records and unchanged cutoffs: gap-only returns 14/26414/264 noise and 250/299250/299 events, spike-only returns 15/26415/264 and 247/299247/299, and joint CAP returns 13/26413/264 and 240/299240/299. Relative to gap-only, the joint rule removes one noise return but loses ten events; relative to spike-only, it removes two noise returns but loses seven events. Both empirical checks change decisions, but these count differences do not establish a statistically superior operating point. We brought these existing counts into Section 6 (p. 8, lines 424–431).

LOCATA tests physical accuracy: CAP returns 11/1311/13 active and 1/131/13 inactive records, with 2.1∘2.1^\circ median error among returned directions. Joint CAP equals gap-only there, so this panel evaluates the returned set rather than establishing that adding both guards is necessary (p. 9, lines 451–463). For analysts comparing event and background windows, we distinguish accurate contrast recovery from resolved background-relative separation. Our bounds state the conditions for each claim; our ablations count the events and noise records rejected by the joint thresholds. Section 7 connects this distinction to the empirical tradeoffs (p. 10, lines 506–518).

I thank the authors for their detailed response. However, I think incorporating these changes requires substantial changes to the exposition and the manuscript. I am therefore inclined to maintain my original score.

AuthorsOpenReview ↗

Clarification of completed revisions

Thank you for considering our response.

We incorporated these corrections and exposition changes in the revised manuscript uploaded on 16 November. Sections 4–5 now state that the gap condition implies the spike condition under the linked theorem thresholds. This corrects the independence claim; the individual angle bounds remain valid. Our problem and estimator are unchanged.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.