acceptodds
Under review as a conference paper at ICLR 2027

Stronger Forcing Is Not Pattern Change in Environmental Streams under Equal Norms

Abstract

We often use field norm to summarise multivariate environmental streams such as ocean temperature, groundwater level, and soil moisture. A changing norm can lead us to mistake stronger forcing of a fixed pattern for a change in the pattern itself. On fully matched equal-norm pairs, both readings share the field-norm and fixed-dictionary residual paths, so those scalar summaries cannot separate them. We treat the active pattern as a subspace and score mean pattern motion on the Grassmannian. On noiseless rank-1 matched streams, we prove that endpoint and block comparisons fail on the stated path classes. Mean pattern motion identifies one-way drift, leave-and-return, and odd-lag equal hopping under the stated lag conditions. For one-way steady drift, we establish trailing-window rates and finite-sample exchangeable-null calibration. CORAL (Construction on Real Atlantic Latents) is our labelled equal-norm suite in North Atlantic EOF coordinates. In this suite, field norm and leave-and-return endpoints exhibit the predicted failures. Mean pattern motion separates drift and leave-and-return but separates equal hopping only partly when we fix the hop lag and window. We release code and artifacts reproducing the experiments and figures in our anonymous repository https://anonymous.4open.science/r/pattern.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Grassmannian geometry, subspace tracking, sequential monitoring, multivariate environmental streams, pattern change detection

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

The paper asks how to distinguish changes in signal strength from changes in spatial patterns when their norms are identical. It proposes averaging lagged Grassmann distances between estimated subspaces, shows where norm and endpoint statistics fail, and provides tracking and calibration guarantees under specific assumptions. Experiments use synthetic signals in North Atlantic EOF coordinates, additional noise tests, and an unlabelled heatwave case study.

Critical strengths
  1. The problem is well motivated. The matched examples clearly show why signal strength alone can miss spatial changes, and why comparing endpoints can miss an excursion.

  2. The calibration framework is useful. Under stream-level exchangeability, the paper provides finite-sample false-positive control. The sensitivity and calibration-shift experiments also make the scope of this guarantee clearer.

Critical weaknesses
  1. The incremental contribution relative to existing subspace detectors is insufficiently established. The proposed procedure combines windowed SVD, principal-angle distances, temporal averaging, and standard exchangeability-based calibration. The impossibility results concern a restricted feature class, including even-length angle proxies, rather than the complete cited detection procedures. Appendix C.7 adds useful comparisons, but its Grassmann-jump CUSUM is a heuristic inspired by Subspace-CUSUM. Consequently, the evidence does not yet establish what substantial new capability or knowledge the proposed method contributes beyond existing detectors and simple normalized-covariance or spectral statistics.

  2. The mechanism behind the empirical separation needs clarification. CORAL uses prescribed three-dimensional paths, with very low projected noise in the main table. For period-two hopping, an even tracking window gives a constant population leading subspace, yet ranking improves from approximately chance under frame-orthogonal residual noise to 100% under isotropic noise. This raises the possibility that differences in eigengaps and estimation fluctuations contribute to separation. The authors acknowledge the window-averaging limitation, but the mechanism behind the successful noisy results remains unclear.

Questions for authors
  1. For period-two hopping with an even window length w=36w=36, the population windowed leading subspace is constant. What explains the 100% isotropic-noise ranking? Can you separate genuine trajectory recovery from effects of eigengaps and estimation fluctuations, including the noise-to-zero limit and odd versus even windows?

  2. How does the method compare with an appropriate implementation of an existing subspace detector and a simple normalized-covariance or eigenvalue baseline under the same calibration protocol? For the noiseless CORAL hopping construction, the covariance eigenvalues differ between the branches despite equal norms, making this comparison particularly consequential.

  3. The main CORAL experiments fix ∣αt∣=1|\alpha_t|=1. Does the reported separation still hold when amplitudes vary over time, for example through trends or bursts? With calibration and test streams sharing the same amplitude law, can you report false-positive rates and detection power on amplitude-varying matched pairs?

3 comments
AuthorsOpenReview ↗

Response to Reviewer ziPA

DIRECT ANSWER — Contribution and existing detectors. We establish what selected summaries discard and when noisy estimates can distinguish fixed and moving patterns. Our matched constructions hide return excursions from endpoints and constant-amplitude alternation from even-block angles. For smooth one-way drift, we bound estimation error relative to the motion gap and calibrate decisions using exchangeable nulls. Our hopping trials show perfect noisy ranking even when windowing erases alternation (Proposition 4/Figure 9). Table 4 shows covariance and Subspace-CUSUM also succeeding. Our analysis distinguishes detecting a change from recovering its spatial path.

CORRECTION TO THE PAPER — We incorrectly grouped Subspace-CUSUM with endpoint/block comparisons in the submitted related work. We now describe its projected-energy mechanism separately and remove that attribution from the angle-proxy descriptions (p. 2, lines 79–87). The submitted remark after Theorem 1 already restricts its lower bounds to specified angle proxies (p. 4, line 215, through p. 5, line 218). Theorem 1(iii) now explicitly states constant amplitude (p. 4, lines 205–208), as its submitted proof required: equal visit counts alone do not balance squared amplitudes. Table 2's hop rows (p. 8) now mark fixed residual as unmatched; projected early/late alone is degenerate because it lacks a population eigengap. This corrects the labels without changing numerical values.

NEW EVIDENCE — Hopping mechanism (Q1). Different noise responses separate mean-motion scores after even windows erase noiseless alternation. In new Proposition 4 we prove the exact boundary at constant amplitude and zero noise: for directions (cos psi, ±sin psi), even windows have a constant leading line; odd windows alternate by arctan(sin(2 psi)/(w cos(2 psi))) for 0<psi<pi/4. Odd lags retain this angle and even lags cancel it. At psi=30 degrees, even-window hopping has spectrum (0.75,0.25,0), versus (1,0,0) on forcing; the gaps are 0.5 and 1. For w=36, we tested 100 pairs and obtained forcing/hopping means 0/0 at zero noise, 9.37e-8/1.51e-7 at noise 1e-6, and 8.34e-4/1.34e-3 at 0.0089. Rankings are 50%, 100%, 100%. Both scores approach zero as noise vanishes; the positive-noise rankings are finite-trial observations, not a trajectory-recovery guarantee. At zero noise, w=35 and 37 retain scores 0.0494 and 0.0468 radians, below the instantaneous hop angle pi/3. These mechanism scores are mean lagged angles in radians, not per-month rates. Section 6 explains the mechanism (p. 8, lines 418–423); Appendix C.10 gives the proof (p. 32, line 1723, through p. 33, line 1736), noise check (p. 33, lines 1739–1746) and Figure 9 (p. 33, caption lines 1762–1769). Figure 9's error bars use 100 independent pairs.

NEW EVIDENCE — Stronger comparisons (Q2). Motion, normalised covariance and projected-energy Subspace-CUSUM each attain 100% observed power across all nine new path–amplitude cells at in-frame noise 0.25. Spectral dispersion reaches 19–33.5% on drift, 32.5–35% on return and 100% on hopping, confirming the spectral distinction you identified. We implemented Subspace-CUSUM's projected-energy recursion with an independently trained null direction, a future window excluding the scored sample and its 36-sample delay. We calibrate the maximum over updates available by the common horizon on the same independent null streams as the other scores. We choose its drift parameter from separate null training data, without transferring Gaussian delay optimality to these different signal laws. The other scores average normalised-covariance deviation or spectral dispersion 1−lambda_max/trace. We fix initial azimuth at zero and train references on that null direction, using the submitted drift/return schedules; the original CORAL trials randomised initial azimuth. Section 6 reports these results (p. 8, lines 423–428), Appendix C.10 gives the implementation (p. 33, line 1770, through p. 34, line 1788), and Table 4 the outcomes (p. 34, lines 1796–1813). Covariance and projected-energy accumulation retain information that the specified angle proxies discard.

NEW EVIDENCE — Amplitudes (Q3). The observed full power for motion, covariance and Subspace-CUSUM holds for constant 1, trend 0.5+1.5t/T and a burst peaking at 4 in Table 4. Both latent branches share the amplitude path, with independent additive noises; calibration and test share the amplitude/noise law. We run 20 disjoint repetitions, each with 19 calibration nulls and 10 fresh streams per class at q=0.10 (p. 34, lines 1789–1794). Fresh-null rejection ranges from 4.5–14.5%; standard errors across the 20 independent repetitions range from 1.4 to 3.8 percentage points. The 200 test streams per class are not 200 calibration replicates. The matched-law design answers Q3 without assuming robustness to amplitude-law shift.

Thank you for updating the paper. The new comparisons are helpful, but motion, covariance, and Subspace-CUSUM all reach 100% power in Table 4. Could you also compare these methods in the harder structured-noise settings in Appendix C.7, using the same calibration protocol? C.7 uses a Grassmann-jump CUSUM method rather than the Subspace-CUSUM baseline added in Table 4.

AuthorsOpenReview ↗

Follow-up on structured-noise comparisons

RESULTS On hopping, spectral dispersion and projected-energy Subspace-CUSUM have 100% observed power in all eight cells, versus 58.5–96.5% for mean motion and 81.5–100% for covariance. On return at noise 0.25, covariance gives 99.5–100% and CUSUM 93.5–99.5%, versus motion’s 55.5–64.5%; spectral dispersion gives 21.5–26.5%. At noise 0.50, covariance retains 53.5–64% return power while motion falls to 15.5–16.5%. We added all sixteen conditions in Tables 4–5, Appendix C.7, and report these comparative limitations in Section 7. The earlier isotropic/amplitude-law comparison is now Table 6 in C.10.

CALIBRATION AND UNCERTAINTY All methods score the same streams. Each of 20 independent repetitions uses 19 calibration nulls and 10 fresh streams per class, with strict rejection above the 18th ordered calibration score at q=0.10. Ten separate training-null streams set references and CUSUM drift without test tuning. We use the actual projected-energy recurrence from C.10, including its 36-sample reporting delay, rather than the Grassmann-jump heuristic. C.7 specifies the retained paths, windows, lags and noise laws/levels.

Fresh-null rejection ranges from 4.5% to 15%. For each of spectral dispersion and CUSUM, all eight paired hopping-power advantages over motion (3.5–41.5 percentage points) have pointwise repetition-bootstrap 95% intervals excluding zero. Resampling uses the 20 independent repetitions. These are not simultaneous intervals, and observed 100% power, including collapsed empirical intervals, is not a population guarantee.

REFERENCE CONVENTION We fix initial azimuth within a repetition and randomise it between repetitions; original C.7 randomised it per stream. For return nulls, this permits independent training of one fixed null direction. The calibration counts/rule and noise conditions are unchanged, but this is not an exact replay of the original orientation mixture.

SCIENTIFIC IMPLICATION These are power limitations of the motion decision rule in the tested settings. For constant-amplitude hopping, the information difference is explicit: the branches have different covariance spectra but the same noiseless even-window leading line (C.10). Spectral or projected-energy detection can succeed without resolving the alternation; successful noisy motion rejection likewise does not establish trajectory recovery. Our matched constructions identify information lost by specified summaries, the window-parity result identifies loss caused by averaging, and the smooth-drift bounds give sufficient conditions for preserving a motion gap through estimation. Together with exchangeable-null calibration, these results explain which comparisons can conceal change and when estimated motion supports a calibrated fixed-versus-moving decision.

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

The paper provides a method for separating and diagnosing norm change from pattern change in spatiotemporal cubes using subspaces. This makes sense because the change of a subspace is a pattern change and a change of magnitude within a subspace is a norm change. Then they apply their method to the CORAL dataset of North Atlantic sea surface temperature.

Critical strengths

The paper introduces an entirely novel set of terms and a framework for evaluating change in spatiotemporal data cubes. The authors prove things about this framework and apply it to a real world sea surface temperature dataset.

Critical weaknesses
  1. This paper is incredibly jargon heavy making it very difficult to follow. From the best of my knowledge the jargon used aligns neither with the climate science community nor the machine learning community. Introducing new jargon is OK, but it should be very carefully defined, especially when used for proofs etc. Examples include phrases from the abstract like "field-norm," "fixed-dictionary residual paths," "active pattern", "noiseless rank-1 matched streams", "endpoint and block comparisons", "one-way drift", leave-and-return", and "odd-lag equal hopping." A paper filled with these sorts of terms without substantial mathematical formalizations or links to prior literature is difficult to asses mathematically (thus my reduced confidence in my assessment) and to follow in general (thus my low score).

  2. It is difficult to assess the novelty and skill of the proposed method because the method is not placed in context of past scientific questions and the baselines are used in experiments do not come from previous works.

  3. There are multiple mathematical inconsistencies and ambiguities regarding notation, definitions and intuition.

Questions for authors
  1. Jargon

    • How are the terms steady drift, leave-and-return, and equal hopping from Table 1 linked to the terms in Figure 1 (stronger forcing, pattern change, and equal norms)?
    • Non standard terms like steady drift, leave-and-return, and equal hopping should be at least intuitively defined before they are used.
    • 011 What is a field norm?
    • 013 Do you mean forcing or scaling?
    • 017 What is a matched stream?
    • 017/018 What is a block comparison and a stated path class?
    • 019 what is one-way drift leave-and-return and odd-lag equal hopping?
    • 020 what is the difference between one-way drift and one-way steady drift?
    • 020 what is a trailing-window rate?
    • 148 what is a rank-1 schedule?
  2. Background/ context/ novelty

    • Has this exact problem been addressed before in prior works? If not, why?
    • Related to the previous question: are there already methods for separating scale and pattern change? If so, can you add these as baselines?
    • 033 How does one “confuse these readings?”
  3. Mathematical questions

    • 100 In Definition 1, what is S_t and S_1, are they subspaces?
    • 103 How do you find the orthonormal frame U_t?
    • 107 what is P_R?
    • 112 Do you need to define both v and \eta? Maybe just one of these is enough.
    • 122 What is Pr?
    • 125 What is \Pi_{drift}?
    • 145 A drawing or intuitive explanation of construction 1 or a tie to a real world scenario is required to motivate it.
    • 156 equation 3 follows immediately and does not need to be a part of the proposition
    • 160 what is the risk of a test?
    • 252 online scores just comes from PCA of the data in a time window of width w as long as the data are mean-centered on this window. Then your statistic is just the shift of the principal subspace as we change the window, averaged across varying window locations. Why choose d=3?
    • 305 What is \hat{\tau}_t?

Small items - Is the title is grammatically correct? L121 Remove bold from definition name. L123 Construction 1 is referenced before it is defined.

5 comments
AuthorsOpenReview ↗

Response to Reviewer pFip

DIRECT ANSWER — Question and contribution. We show why unchanged endpoints need not imply a stable pattern, and when window estimates can preserve evidence of movement. Our matched constructions hide return excursions from endpoints and constant-amplitude alternation from even-block leading spaces. Motion separates the stated noiseless paths at suitable lags; our noisy recovery result covers smooth one-way drift. We connect these distinctions to estimation error and calibrated decisions using established tracking and calibration tools. Subspace-CUSUM accumulates projected energy, while our score averages lagged angles; Table 4 shows both succeeding. The submitted post-Theorem 1 remark already limits the lower bounds to angle proxies (p. 4, line 215, through p. 5, line 218).

CLARIFICATION — Terminology and intuition. The field-norm monitor uses the squared norm, the sum of squared measurements, averaged over time where stated. “Stronger forcing” means changing coefficients within a fixed spatial subspace: excitation/scaling, not an inferred physical cause. Confusing these readings means interpreting amplitude change as spatial reorganisation. Matched streams share norms; fully matched streams also share energy outside a fixed reference subspace. Steady drift moves progressively away, leave-and-return makes an excursion, and hopping alternates between two positions. These are Figure 1(b) paths paired with Figure 1(c)'s common norm. A block contains consecutive observations; a path class contains subspace sequences with specified movement constraints. “One-way drift” and “one-way steady drift” name the same restricted subclass. A trailing-window rate bounds estimation error from recent observations. We explain these terms in the abstract (p. 1, lines 11–24), introduction (p. 1, line 52, through p. 2, line 73) and Section 2.

CLARIFICATION — Mathematical objects. S_t and S_1 are spatial subspaces. U_t is a latent orthonormal frame spanning S_t; window SVD estimates the subspace, not a unique latent basis. P_R projects orthogonally onto R. Pr denotes probability; balanced risk averages false-positive and false-negative probabilities. Pi_drift denotes a set of paths. Eta_t measures one-step movement; v_t measures lag-Delta movement, so both are needed when Delta differs from 1. Definition 1 and nearby text explain these objects (p. 3, lines 108–127). A rank-1 schedule is a sequence of unit directions spanning lines. Within Construction 1 we now explain how moving around a cone preserves norm and reference residual while direction changes (p. 4, lines 164–167). This is a controlled diagnostic example, not an ocean dynamics model.

FACTUAL CORRECTION — The review assumes window centring; the submitted estimator and code already use uncentred window second moments after record preprocessing. Separately centring each window would eliminate a constant noiseless forcing stream and define a different estimator. We now state this convention in Section 5 (p. 6, lines 278–283). We chose three coordinates for a cone axis and two transverse directions, not to estimate the ocean's intrinsic dimension. The same passage defines tau-hat_t as the held-out null quantile at horizon t; the submitted calibration corollary gives its order-statistic formula. Longer windows reduce noise but can erase rapid movement (lines 284–286). Fixed-horizon and repeated-monitoring calibration differ (p. 7, lines 329–331).

CORRECTION TO THE PAPER — We replaced I_n with I_d in three noise identities in the reduced-coordinate proof (Appendix A.5, p. 19, line 998, through p. 20, line 1027); the bounds are unchanged. We also corrected the related-work attribution: Subspace-CUSUM accumulates projected energy rather than endpoint/block angles (p. 2, lines 79–87).

DIRECT ANSWER — Presentation choices. Proposition 1 keeps norm and residual equality together because their joint restriction supports Corollary 1. The forward reference identifies that argument; Construction 1 is not needed to define the path classes. The title is grammatical: it distinguishes excitation within a fixed pattern from spatial change under equal norms. The revised opening explains these terms. Bold definition headings help readers locate the objects.

NEW EVIDENCE — Motion, covariance and Subspace-CUSUM each attain 100% observed power in the nine tested path–amplitude cells. We compared projected-energy Subspace-CUSUM, normalised covariance and spectral dispersion. Section 6 summarises them (p. 8, lines 423–428); Appendix C.10 gives implementation details (p. 33, line 1770, through p. 34, line 1788), and Table 4 reports outcomes (p. 34, lines 1796–1813). Our comparison shows that covariance and projected-energy scores distinguish these streams using information beyond the angle summaries; our motion score averages lagged subspace angles.

Thank you for your detailed response. I'm still working through the response so I am not fully ready to reply yet.

However, I came across this work: Tippett, M.K., L’Heureux, M.L. Low-dimensional representations of Niño 3.4 evolution and the spring persistence barrier. npj Clim Atmos Sci 3, 24 (2020). https://doi.org/10.1038/s41612-020-0128-y

It may be of interest to the authors and deserve a place in related work and perhaps even in experiments as a baseline.

Thank you for addressing most of my questions. The CUMSUM baseline and extra experiment are nice additions to this work. still have some lingering questions/ comments.

  • Is it intentional to have bold only for some definition names
  • “Steady drift moves progressively away, leave-and-return makes an excursion, and hopping alternates between two positions.”
    • What is this moving away from and what is the difference between making an excursion vs moving progressively away?
  • “One-way drift” and “one-way steady drift”
    • “One-way drift” is mentioned twice in the manuscript and I was unable to find its formal definition. Is it just a drift of the subspace pattern in one direction? If so, this should be stated clearly.
  • “Lag” vs “window size” it might be worth it to define each of these explicitly as the time between windows and size of the window used to estimate the subspace
  • Did I miss further motivation for construction1? Maybe it is here, but I'm not sure how to interpret this phrase on L149 “provide the needed separation between the readings.”
Reviewer_PQA2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to distinguish stronger forcing of an existing environmental pattern from changes in the pattern itself. The authors show that field norms and some conventional subspace comparisons can fail to distinguish these cases, even when the underlying patterns behave differently.

They propose measuring pattern motion using Grassmannian distances and provide theoretical results for steady drift, leave-and-return, and alternating patterns. The method is evaluated on the CORAL benchmark, which uses synthetic signals constructed in North Atlantic ocean-temperature representations. The experiments show good separation for drift and leave-and-return patterns, but more limited performance for alternating patterns under some noise conditions.

Critical strengths

The paper makes an important distinction between changes in signal magnitude and changes in spatial patterns. The theoretical results help explain why norm-based and endpoint-based methods can miss certain types of pattern changes. The authors test several types of pattern changes and examine the effects of noise, window size, and lag selection. I appreciate that the paper reports cases where the proposed method fails rather than presenting only successful results.

Critical weaknesses

Most experiments use artificially constructed patterns embedded in ocean-temperature data. Although this provides controlled comparisons, it does not fully demonstrate that the method can reliably identify pattern changes in real environmental systems. The actual North Atlantic heatwave analysis is mainly descriptive and does not provide ground-truth labels. The theoretical guarantees rely on restrictive assumptions about noise, rank, and pattern motion. The experiments also show that performance can drop substantially with incorrect rank selection, stronger noise, or unsuitable window settings. This raises questions about how reliably the method would work on more complex environmental streams.

Questions for authors

Q1. The main experiments use synthetic pattern changes constructed from ocean-temperature representations. Could the authors provide further evidence that the method can detect meaningful pattern changes in real environmental data?

Q2. The results appear sensitive to rank selection, noise, and window settings. How should these parameters be chosen in practice when the true pattern structure is unknown? Could the authors clarify the method's reliability under such conditions?

1 comment
AuthorsOpenReview ↗

Response to Reviewer PQA2

DIRECT ANSWER — Real environmental evidence (Q1). We measured calibrated power under varying amplitudes; the heatwave remains a descriptive observational example. We have not added independently labelled environmental transitions. CORAL tests identification with known paths, while the heatwave shows magnitude, cross-era angles and local motion giving different readings on a measured record. OISST corroborates warmth, without labelling spatial change. The revised abstract states this scope (p. 1, lines 11–24), as does Section 6 (p. 7, lines 349–352; p. 9, lines 475–480; Table 3/Figure 5, p. 10). Together they connect controlled identification with interpretation of observations.

CLARIFICATION — What the controlled evidence establishes. We prove a specific information loss: on our fully matched noiseless pairs, norms and reference residuals coincide despite movement, and a returning pattern preserves endpoints. We then establish when lagged motion retains the distinction and when noisy window estimates preserve it for smooth one-way drift. For environmental interpretation, we distinguish information discarded by a summary from error introduced by estimation. CORAL uses prescribed paths in frozen ocean-temperature EOF coordinates; its two eras are coordinate frames, not independently labelled field validations. Known paths test this distinction; labelled field transitions would test broader environmental accuracy. Section 6 and Appendix C.9 keep these roles separate.

CORRECTION TO THE PAPER — We corrected our heatwave prose: squared-norm and residual ratios exceed their displayed baseline ranges; only motion lies within its range. Table 3 already shows squared norm 2.09 and residual 2.23 above [0.70,1.37] and [0.76,1.23]; only motion 0.77 lies within [0.37,1.70]. We corrected the prose (p. 9, lines 475–480), leaving the numerical values unchanged. These descriptive ranges are not significance tests valid under temporal dependence. Appendix C.9 now explains that scores ending in the event use trailing 36-month windows and lag 12 that extend before it (p. 32, lines 1694–1698). They do not isolate movement occurring only within the twelve-month heatwave.

DIRECT ANSWER — Parameter choice and reliability (Q2). We recommend choosing a candidate rank from an independent reference period with a resolved eigengap, tuning nearby windows/lags on separate streams, freezing those choices, then calibrating on held-out nulls at the target horizon. Section 7 states this order (p. 10, lines 534–539). Longer windows reduce sampling noise but can erase fast movement; the lag must retain the movement of interest. These choices require independent scientific and tuning evidence when the true structure is unknown. This guidance does not guarantee unknown-rank recovery. A fixed-horizon threshold also does not control repeated monitoring; the submitted stream-maximum calibration addresses a finite monitoring horizon (Section 5, p. 7, lines 329–331).

EXISTING EVIDENCE — Rank errors can destroy power even while a fixed score retains valid exchangeable-null calibration. The submitted recovery/power theorem assumes correct rank, an eigengap, bounded amplitudes, an isotropic signal Gram, smooth one-way drift and independent isotropic sub-Gaussian noise. It does not cover arbitrary recurrent paths or dependent environmental noise. Calibration requires the score to be fixed beforehand and calibration and fresh-null streams to remain exchangeable; it does not require correct rank. Appendix C.6 already shows power falling from 79.5% at rank 1 to 10% at rank 2 at its elevated-noise default. Structured noise also reduces return power in C.7, and C.8 documents false-positive inflation under null-law shift. Section 5, Remark 2 now states the recovery/calibration distinction (p. 7, lines 340–345). We retain these adverse findings: calibration controls false alarms under its assumptions, rather than ensuring sensitivity or robustness to a changed null law.

NEW EVIDENCE — Under matched calibration and test laws, motion attains 100% observed power in all nine path–amplitude cells, with fresh-null rejection 4.5–12% (new Table 4). We tested constant, trend and burst amplitudes at in-frame noise 0.25. Each cell has 20 independent repetitions, each with 19 calibration nulls and 10 fresh streams per class; the resulting 200 test streams per class are not 200 calibration replicates. Appendix C.10 gives the protocol (p. 34, lines 1789–1794) and Table 4 the outcomes (p. 34, lines 1796–1813). These tests support the prescribed laws, not arbitrary amplitude-law shift. New Proposition 4/Figure 9 further shows why successful noisy hopping scores can coexist with erased noiseless alternation. Our evidence shows when motion is informative and why high ranking alone does not establish trajectory recovery.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.