{"id":"be31a50c-bf1d-4b02-941a-1fb9bd91cdf9","arxiv_id":"2608.05974","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A hierarchical search that uses the clustering of high-likelihood secondary maxima recovers generic EMRI signals in 14-dimensional parameter space from simulated LISA data.","lead":"This paper presents a search method that finds gravitational-wave signals from extreme-mass-ratio inspirals across a wide 14-parameter space without first knowing where the signal is. It works by using the pattern of secondary peaks in the likelihood landscape to shrink the search region step by step, demonstrated on two simulated LISA signals.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed broad-prior end-to-end search is not demonstrated: the 7D stage fixes the initial orbital frequency ν0 to within ~2 days of the true value (Sec. D.4), so the effective prior on ν0 is narrow and source-targeted.","rationale":"The reader's weakest assumption was the peak-pattern clustering. That is a legitimate concern, but the fiducial ν0 dependence is more directly disqualifying for the abstract's claim. Even if the peak pattern is real, the search's first stage cannot run blind over the full 14D space because ν0 is effectively fixed by the fiducial choice. The paper's own Sec. D.4 acknowledges this. Thus the two injections are not a demonstration of broad-prior end-to-end recovery; they are a demonstration with a heavily source-targeted initial frequency. This supports the reader's CONDITIONAL verdict, but for a different reason. A single additional experiment with a far-offset fiducial ν0 would settle the matter.","tokens_in":18114,"tokens_out":5247,"duration_ms":51009,"concrete_test":"Run the same 7D hierarchical search on an injection with the fiducial ν0 offset by, e.g., 10^5 lags (~17 days) or drawn from the full astrophysical prior on ν0, leaving all other settings identical. If the first-stage landmark LLR drops below threshold or the contracted ranges fail to enclose the true parameters, then the method's broad-prior claim is falsified. A softer version: map the 7D likelihood's peak LLR as a function of fiducial offset (e.g., 0, 10^3, 10^4, 10^5 lags) at the true θ7D; if the peak falls sharply beyond a few thousand lags, the effective ν0 prior is narrow.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is the handling of the initial orbital frequency ν0, not (or in addition to) the peak-pattern assumption. In the 7D likelihood, ν0 is removed from the search and replaced by a time-lag maximization over n (Eqs. 10–13). However, the template waveform is generated from a fixed fiducial ν0 chosen, for both injections, exactly 10^4 lags (≈2 days at 15 s cadence) before the true value (Sec. D.4). The lag search only compensates for small offsets: because the template's phase evolution is computed from the fiducial ν0, a lag n shifts the effective starting frequency to ν(t0+nΔt), but the chirp rate and the full frequency trajectory also change; for lags of weeks the template no longer matches the signal. Hence the effective prior on ν0 is a narrow window around truth, not the broad astrophysical prior claimed in the abstract and Table V. The paper admits this: 'choosing the fiducial ν0 that is 104 lags earlier than the true value is not sufficiently general for arbitrary EMRI sources' and 'This poses a challenge for a fully blind search.' The two demonstrations therefore do not establish the headline claim of end-to-end coherent parameter estimation over astrophysically broad priors; they show recovery when ν0 is already known to ~2 days. A blind search would need a method to set or marginalize over the fiducial ν0 from the data alone; none is demonstrated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a hierarchical, coherent search pipeline for generic (spinning, eccentric, inclined) extreme-mass-ratio inspirals in LISA-like data. The central technical contribution is a reduced-dimensional profile likelihood: an 8D likelihood over phase-evolution and detector-delay parameters, and a 7D likelihood that removes the initial orbital frequency ν0 by replacing it with a time-lag maximization. The search uses particle-swarm optimization to collect secondary maxima and iteratively contracts the prior volume, exploiting a claimed 'peak pattern' in which high-likelihood secondary maxima cluster progressively around the global maximum. Two half-year analytical-kludge injections at SNR ≈ 50 in stationary Gaussian LISA noise are recovered with fitting factors of 0.989 and 0.971 and small parameter errors. The paper claims this is the first end-to-end coherent parameter estimation for generic EMRIs over astrophysically broad priors.","tokens_in":18422,"tokens_out":8333,"duration_ms":85475,"significance":"If the central claim is correct, this would be a substantial advance for EMRI data analysis: it would demonstrate that coherent matched filtering can navigate the multimodal 14D likelihood landscape of generic EMRIs without source-targeted priors, at least for the tested SNR and waveform model. The reduced-dimensionality derivation (Eqs. 1-14) is mathematically clean, the use of nested optimization and the treatment of extrinsic parameters are standard and sound, and the paper is careful to document computational costs and implementation details. However, the paper's broad-prior claim rests on two injections and a hand-tuned control parameter, and a key load-bearing assumption about ν0 is explicitly acknowledged to be non-general. The work is therefore best viewed as a promising proof-of-principle that requires substantially more validation and a revision of its headline claim.","major_comments":[{"comment":"The headline claim of 'astrophysically broad priors' is not supported for the initial orbital frequency ν0. In the 7D likelihood, ν0 is not searched; instead a fiducial value is fixed at 10^4 lags (≈2 days at 15 s cadence) before the true value for both injections, and the lag maximization n* absorbs only small offsets. For a source whose true ν0 differs from the fiducial value by weeks or more, the template's phase trajectory—computed from the fiducial ν0—will no longer match the signal even after the optimal time shift, because the chirp rate and higher-order frequency evolution differ. The manuscript itself states (D.1) that this fiducial choice is 'not sufficiently general for arbitrary EMRI sources' and (D.4) that a blind choice 'poses a challenge for a fully blind search.' The final 8D search also uses a ν0 interval derived from the 7D lag estimate, which is narrow (several weeks). The demonstrations therefore assume ν0 known to about two days, not the broad prior listed in Table V. The authors should either add a ν0-search stage (e.g., an outer grid or a preliminary ν0 localization) and demonstrate it, or revise the abstract and introduction to state the actual effective prior on ν0.","section":"D.1/D.4, Eqs. (10)-(13)"},{"comment":"The 'peak pattern'—that high-LLR secondary maxima concentrate progressively around the global maximum—is the central premise justifying every prior-contraction step, yet it is asserted as a general property and supported only by a qualitative harmonic-SNR argument and two injections, both at SNR ≈ 50, in the same stationary Gaussian noise model. No proof is provided, and no failure-rate study over source parameters, sky positions, SNRs, or noise realizations is presented. The paper's own language ('general global characteristic', 'it proves to be general') goes beyond what two examples can establish. A systematic validation, e.g., 20-50 injections drawn from the prior, with varying SNR and multiple noise realizations, quantifying how often the contracted range encloses the true parameters and how the landmark LLR converges, is needed to support the claimed generality.","section":"Supplemental B2, Figs. 10-11"},{"comment":"The LLR threshold ρthreshold is a control parameter that determines which secondary peaks are retained for the next prior contraction. Table VI shows that the threshold is chosen by hand, with different values for each iteration and each injection (e.g., 30, 30, 32, 38.5, 37, 46.4, 46.8 for injection 1). The success of the hierarchy depends on this choice, yet the manuscript gives no selection rule and no sensitivity study. If the results change substantially when the threshold is varied by a few units, the method is not robust; if they do not, that robustness should be demonstrated. A documented, reproducible criterion for setting ρthreshold, or a study of the method's performance under threshold perturbations, is required.","section":"Supplemental D.3, Table VI"}],"minor_comments":[{"comment":"The wording describing the fiducial ν0 offset is inconsistent: D.1 says '10^4 lags ahead of its true value' while D.4 says '10,000 lags before its true value', and the intended sign (earlier vs later in time) is ambiguous until one reads the backward-integration description. Please clarify the time ordering of the fiducial frequency relative to the true signal.","section":"D.1 vs D.4"},{"comment":"The notation for the degenerate sky location (π−θs, π+ϕs) is indicated with an asterisk, but the relative errors are computed against the degenerate value rather than the injected value, which is confusing. A better presentation would report the minimal sky error after accounting for the known degeneracy, as the text claims.","section":"Tables I-II"},{"comment":"The decreases in landmark LLR between some iterations (e.g., injection 1 between iterations 4 and 5) are attributed to PSO stochasticity, but the text does not quantify how often independent runs fail to find the previous landmark. Reporting the number of runs that recovered the landmark maximum at each stage would give readers a clearer sense of the method's reproducibility.","section":"Figs. 1-2 and D.5"},{"comment":"The computational cost (≈3×10^6 CPU hours, ~2 months per search) is presented as a limitation, and GPU acceleration is deferred to future work. This should be stated more prominently in the main text, not only in the Supplemental Material, because it affects the practical significance of the 'end-to-end' claim for LISA data over several years.","section":"Table IV"},{"comment":"The reference list is extensive but the Introduction's claim of being 'the first coherent search for a generic, spinning EMRI across the complete 14-dimensional parameter space' should be carefully calibrated against the ν0 limitation discussed above; as written, the claim overstates what is demonstrated.","section":"Introduction"}],"recommendation":"major_revision","confidential_remarks":"The paper contains a genuinely interesting algorithmic idea and a clean likelihood-reduction derivation, but the manuscript as it stands overclaims its central result. The ν0 issue is acknowledged in the text and is load-bearing: the demonstrated searches do not use a broad ν0 prior. I would be willing to reconsider a revised version in which the authors either add a ν0-search stage and validate it on a broader set of injections, or explicitly frame the contribution as a hierarchical coherent search conditional on a known or narrowly constrained ν0. I also suggest giving the editor's attention to the novelty claim: several recent works (e.g., Ye et al., Strub et al., Cole et al., Wang et al.) have demonstrated end-to-end EMRI searches over broad priors in restricted models; the present paper's contribution is the extension to generic spinning, eccentric, inclined waveforms, so the 'first' claim should be worded to reflect this distinction."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe one thing to know: this paper has a genuinely new architecture for coherent EMRI search, and the likelihood decomposition is clean, but the headline claim—first end-to-end coherent search over astrophysically broad priors—does not survive a close reading. The stress-test note gets it right. In the 7D stage, the fiducial ν0 is fixed to 10^4 lags, about 2 days, before the true value for both injections. The lag maximization can compensate only small offsets, because the template phase evolution is computed from the fiducial ν0; a fiducial offset of weeks or months would break the match. The paper admits this in Sec. D.4: choosing the fiducial that close is “not sufficiently general” and “poses a challenge for a fully blind search.” So the effective prior on ν0 is narrow and source-targeted, not the broad population prior advertised in the abstract. That is a load-bearing gap, not a minor caveat.\n\nWhat is genuinely new and good: the reduction of the 14D search to a 7D global-localization stage plus an 8D refinement, with extrinsic parameters profiled out by nested optimization, is mathematically sound and computationally meaningful. The idea of using the clustering of high-LLR secondary peaks to contract the prior is well motivated and is visually supported for these two injections. The paper is honest about the cost and about the tuning of ρthreshold. The citation pattern looks fine.\n\nThe soft spots beyond ν0: the peak-pattern property is asserted rather than derived, and it is demonstrated on only two SNR-50 injections in stationary Gaussian noise, with thresholds selected using knowledge of the injections. There is no failure-rate study, no lower-SNR case, no multiple noise realizations, and no test with a more realistic waveform. The Supplemental also notes a shortcut in injection 2’s fourth iteration that slightly compromises generality. These do not falsify the central idea, but they mean the paper is a compelling proof of concept, not a validated pipeline.\n\nFor EMRI algorithm developers and LISA data-challenge participants, this is worth reading and discussing. I would send it to peer review rather than desk reject, but I would expect a major revision that either fixes the ν0 initialization or reframes the end-to-end claim, and adds broader validation before the strong statement can stand.","headline":"A genuinely new coherent EMRI search strategy with clean math, but the broad-prior end-to-end claim outruns the evidence: the fiducial ν0 is tuned to each injection and the peak-pattern heuristic is shown on only two SNR-50 cases.","tokens_in":18957,"tokens_out":4857,"would_cite":true,"duration_ms":51170,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A coherent hierarchical search over the full 14-dimensional EMRI parameter space succeeds without source-targeted priors by letting high-likelihood secondary maxima guide adaptive prior contraction, recovering two test signals with…","keywords":["extreme-mass-ratio inspiral","gravitational-wave data analysis","coherent search","profile likelihood","multimodal likelihood","particle swarm optimization","LISA","adaptive prior contraction"],"falsifier":"Run the pipeline on a population sample—say dozens to hundreds of EMRI injections spanning the broad priors with independent noise realizations—and count how often the hierarchical contraction's final region contains the true parameters and the end-to-end fitting factor stays above a threshold such as 0.95; a single injected source whose high-likelihood peaks cluster around a remote secondary maximum, so that the contracted prior excludes the truth, would refute the central claim.","tokens_in":17872,"feed_emoji":"🛰️","tokens_out":7027,"duration_ms":69064,"temperature":0.7,"pith_summary":"Extreme-mass-ratio inspirals (EMRIs) are long, multiharmonic signals that carry precise maps of the spacetime around massive black holes, but their likelihoods are a needle in a haystack: tiny parameter shifts destroy phase coherence, producing a narrow global peak buried among many secondary maxima. This paper argues that those secondary maxima are not random noise but a usable structure—higher-likelihood peaks cluster more tightly around the true global maximum. Exploiting that peak pattern, the paper builds a hierarchical coherent search: independent particle-swarm runs collect high-likelihood peaks, the envelope of those peaks contracts the prior for the next stage, and a reduced-dimensional profile likelihood (7D for localization, 8D for refinement) keeps matched-filtering sensitivity while taming dimensionality. In two half-year analytical-kludge injections with signal-to-noise ratio near 50 in stationary Gaussian LISA noise, the pipeline recovers all 14 source parameters from population-scale broad priors, with fitting factors 0.989 and 0.971, distance errors near 3%, and sky-location errors under 0.1 radian. If right, this is the first end-to-end coherent parameter estimation for generic spinning, eccentric, inclined EMRIs without source-targeted priors.","feed_headline":"Blind EMRI search recovers all 14 parameters from broad priors","feed_subtitle":"Secondary likelihood peaks guide the hierarchy to 0.989 and 0.971 fitting factors in LISA-like noise.","key_machinery":"The load-bearing mechanism is the peak pattern—the empirical property that the highest-likelihood secondary maxima of an EMRI likelihood become progressively concentrated around the primary maximum, because partial matches to different waveform harmonics create ever-denser peaks as the template approaches the true parameters. The paper deploys this pattern through three coupled devices: a reduced-dimensional profile likelihood in which distance is maximized analytically, the time-independent extrinsic parameters are separated from the phase-coupled parameters by a linear decomposition, and the initial orbital frequency $\\nu_0$ is handled by a fast Fourier-based lag search; a seven-dimensional likelihood for global localization combined with an eight-dimensional likelihood for refinement; and local-best particle-swarm optimization that collects candidate peaks across independent runs. The peak pattern supplies the criterion for what to keep at each contraction step, the profile likelihoods make those steps computationally tractable, and the particle-swarm runs supply the ensemble of peaks whose coordinate envelope defines the next prior.","core_discovery":"The paper's central claim is that a fully coherent matched-filtering search of the complete 14-dimensional EMRI parameter space can succeed over astrophysically broad priors, provided the search treats the likelihood's secondary maxima as guides rather than obstacles. The authors characterize the likelihood landscape of a multiharmonic EMRI waveform: harmonics carry different signal-to-noise ratios, so templates far from the true parameters match only the loudest harmonics, while templates nearer the peak match progressively more of them; consequently, the highest-log-likelihood-ratio secondary maxima concentrate around the global maximum. They convert this concentration into an iterative contraction scheme: each stage runs independent local-best particle-swarm searches over a seven-dimensional profile likelihood, in which extrinsic parameters are profiled out, distance is maximized analytically, and the initial orbital frequency is replaced by a time-lag shift evaluated efficiently in the Fourier domain, and the collected peaks above a rising threshold define the next, smaller prior. A final eight-dimensional search treats the initial orbital frequency as free and removes the lag-induced bias. On two injected generic, spinning analytical-kludge signals of 0.5-year duration with target SNR 50 in stationary Gaussian LISA noise, the hierarchy converges in eight and six stages and returns maximum-likelihood estimates with fitting factors 0.989 and 0.971, fractional errors of $10^{-3}$ to $10^{-2}$ in the phase-evolution parameters, distance errors of about 3%, and absolute sky-location errors below 0.1 radian after accounting for the known sky degeneracy.","pith_inferences":["The peak-pattern assumption, if valid beyond two injections, implies a general design principle for multimodal phase-coherent searches: the risk that a contracted prior excludes the true peak can be reduced by collecting more high-likelihood peaks, because the probability that every collected peak cluster misses the truth falls as the peak count grows.","A natural testable extension the paper does not pursue is to turn the hierarchy into a detection statistic: the number and concentration of high-likelihood peaks above threshold in noise alone versus signal-plus-noise could provide a false-alarm estimate for the search.","The fiducial-$\\nu_0$ fine-tuning problem—how much earlier the template starts than the true signal—is acknowledged as unresolved for a fully blind search, and a population-informed choice of this lag is an obvious follow-up that the current demonstration sidesteps."],"forward_implications":["If the peak pattern holds for realistic sources, coherent matched filtering can be run over population-scale priors on all 14 parameters, not just over source-targeted or Schwarzschild-restricted models.","The 7D lag-shift likelihood consistently reveals the peak structure in the two demonstrations, indicating that global localization can be achieved before the expensive 8D refinement stage is launched.","The hierarchy converged to the true parameters from both a boundary-near and a central injection, suggesting the contraction is not tuned to a particular corner of the prior.","The reported cost is high—about two months for a half-year search on the current CPU pipeline—but the paper identifies GPU acceleration as a direct route to runs on the order of a week.","Because the structure of the method is not tied to the specific waveform family, the paper states it should transfer from analytical-kludge templates to faster and more accurate FEW templates with computational adaptation."],"supporting_citations":[{"why":"Supplies the analytical-kludge waveform model, with post-Newtonian eccentricity and Lense-Thirring precession, used for both injections and templates.","marker":"[3]"},{"why":"Shows that the source distance can be maximized analytically, removing one parameter from the profile likelihood.","marker":"[30]"},{"why":"Characterizes the nonlocal degeneracies of multiharmonic EMRI likelihoods that the peak pattern builds on.","marker":"[33]"},{"why":"Provide the nested-optimization and particle-swarm profile-likelihood machinery from which the hierarchical search is constructed.","marker":"[40, 41]"},{"why":"Define the particle-swarm and local-best particle-swarm algorithms used to collect candidate peaks at each stage.","marker":"[47, 48]"},{"why":"Supplies the SciRDv1 LISA noise power spectral density used in the likelihood computation.","marker":"[50]"},{"why":"Generated the stationary Gaussian LISA noise into which the two test signals are injected.","marker":"[56]"}],"fun_headline_variants":["Secondary peaks guide full EMRI parameter recovery","Coherent EMRI search nails all 14 parameters","EMRI search turns likelihood peaks into guides","Full 14-D EMRI recovery from secondary maxima"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole contraction rests on the empirical claim that higher-likelihood secondary peaks gather increasingly tightly around the true peak; if a sufficiently loud wrong peak cluster ever sits far from the truth, the contracted prior can enclose it and exclude the real signal—and this claim is supported only by two injections and a qualitative harmonic argument, not by proof or a population study.","fun_headline_variants_meta":{"raw":{"variants":["Secondary peaks guide full EMRI parameter recovery","Coherent EMRI search nails all 14 parameters","EMRI search turns likelihood peaks into guides","Full 14-D EMRI recovery from secondary maxima"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000243,"raw_usage":{"total_tokens":1635,"prompt_tokens":1155,"completion_tokens":480,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":771,"completion_tokens_details":{"reasoning_tokens":421}},"tokens_in":771,"tokens_out":480,"duration_ms":5328,"temperature":1.0,"reasoning_tokens":421,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T20:02:49.354737+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline on a population sample—say dozens to hundreds of EMRI injections spanning the broad priors with independent noise realizations—and count how often the hierarchical contraction's final region contains the true parameters and the end-to-end fitting factor stays above a threshold such as 0.95; a single injected source whose high-likelihood peaks cluster around a remote secondary maximum, so that the contracted prior excludes the truth, would refute the central claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the analytical-kludge waveform model, with post-Newtonian eccentricity and Lense-Thirring precession, used for both injections and templates."},{"cited_title":"Babak, J","cited_arxiv_id":null,"evidence_quote":"Shows that the source distance can be maximized analytically, removing one parameter from the profile likelihood."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Characterizes the nonlocal degeneracies of multiharmonic EMRI likelihoods that the peak pattern builds on."},{"cited_title":"Babak, A","cited_arxiv_id":null,"evidence_quote":"Supplies the SciRDv1 LISA noise power spectral density used in the likelihood computation."},{"cited_title":"Petiteau, G","cited_arxiv_id":null,"evidence_quote":"Generated the stationary Gaussian LISA noise into which the two test signals are injected."}],"review_version":1}