{"id":"b3e50841-5fc9-48df-9561-52762a03d034","arxiv_id":"2412.20793","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"TDI-infinity, applied to LISA Bayesian parameter inference, preserves parameter recovery better than classical TDI when data gaps are present, though the all-in-one noise cancellation is not used in the headline simulations.","lead":"This paper applies the recently proposed TDI-infinity method to LISA gravitational-wave data analysis, showing it can handle data gaps more cleanly than classical time-delay interferometry. It also sketches an 'all-in-one' version that cancels several noise sources at once, but the main Bayesian simulations use a reduced version that suppresses only laser noise.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Chunked likelihood double-counts overlapping data and discards long-timescale noise correlations, so the claimed TDI-infinity posterior stability under gaps is not yet validated.","rationale":"The reader identified the chunked likelihood approximation as the weakest assumption, and my independent analysis converges on the same point. The chunking scheme of Eqs. (43) and (56) with l=500 and q=0.1 is used for all TDI-infinity results, yet the paper provides neither a convergence study nor a correction for the overlap double-counting. This matters because the central claim is comparative: TDI-infinity yields stable posteriors with gaps while classical TDI degrades. If the chunked pseudo-likelihood overweights data through overlaps and omits long-timescale noise correlations, the reported TDI-infinity credible intervals could be artificially narrow, making the method look more robust than it is. The no-gap/no-noise case and the agreement with time-domain TDI-2 in the no-gap case mitigate but do not eliminate this concern, because neither tests the gap scenario against a full-covariance reference. The all-in-one scope mismatch and single noise realization are additional limitations, but they are secondary: the reduced TDI-infinity gap-handling claim could still hold even if the all-in-one pipeline is unfinished. Therefore, the chunked-likelihood validation is the most load-bearing issue, and the appropriate verdict remains CONDITIONAL pending a convergence study.","tokens_in":34387,"tokens_out":6516,"duration_ms":73130,"concrete_test":"Run a convergence study for scenario (ii) (noise, no gaps): compute TDI-infinity posteriors for chunk lengths l ∈ {125, 250, 500, 1000, 2000} samples and overlap factors q ∈ {0, 0.05, 0.1, 0.2}, and compare the 68% credible interval widths and posterior means against a full-length (l = N) time-domain likelihood calculation. If intervals shrink systematically as l or q increases, or if they differ from the full-length result by more than the MCMC sampling uncertainty, the chunked likelihood is biased or double-counting. As a secondary check, generate several noise realizations at the true parameters and compute the mean of -2 log p(d | θ_true) under the chunked likelihood; if this mean is significantly below the effective number of independent TDI-infinity samples, overlapping data are being overcounted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that TDI-infinity handles gaps better than classical TDI rests on the posterior distributions shown in Figs. 13 and 14, which are computed with the chunked likelihood of Eq. (55). Chunks overlap by q=0.1 with l=500 samples (125 s), and the total log-likelihood is the plain sum over chunks, so data in the overlap regions enter multiple times. This double-counting inflates the effective sample size and can artificially narrow credible intervals, making the posteriors look artificially robust. The paper acknowledges the double-counting trade-off but provides no reweighting scheme and no convergence study over l and q (Sections 4.4, 5.5). Additionally, each chunk of 125 s cannot represent noise correlations on timescales longer than 125 s, even though the OMS noise model (Eqs. 57-58) has strong low-frequency power and the 1-hour observation contains frequencies down to ~0.28 mHz; the covariance matrix N built via Eq. (54) is therefore truncated. The no-gap/no-noise case (Fig. 11) cannot expose these errors, and the agreement with time-domain TDI-2 in Fig. 12 provides only partial reassurance, since that comparison uses no gaps. With gaps, there is no independent full-covariance benchmark. If the chunking bias is significant, the apparent superiority of TDI-infinity under gaps could be an artifact of the pseudo-likelihood rather than a genuine advantage.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper adapts TDI-infinity, a numerical null-space projection that removes laser and other suppressible noises from raw LISA beat-note measurements, to Bayesian parameter estimation for a one-hour massive-black-hole-binary signal. Gaps are handled by deleting invalid rows from the measurement vector and design matrix before computing the null space, so that the number of lost samples roughly equals the number of introduced gaps. The authors introduce an all-in-one formulation that includes optical-bench, clock, and modulation noises, present a chunked time-domain likelihood, and compare posterior distributions from TDI-infinity with classical TDI-2 in frequency-domain and time-domain implementations, under no-gap and gapped scenarios. The central simulation result is that TDI-infinity posteriors remain stable under six one-sample gaps, while classical TDI loses roughly 400 seconds of data and shows degraded posteriors. The inference runs, however, use a reduced version of TDI-infinity that suppresses only laser noise, and the chunking parameters l and q are not varied or checked for convergence.","tokens_in":34653,"tokens_out":5395,"duration_ms":55044,"significance":"If the gap-robustness result holds, this is a useful contribution to LISA low-latency data analysis: it replaces ad hoc gap filling or segment-wise likelihoods with a principled algebraic projection, and it avoids the data loss around gaps that plagues classical TDI. The inclusion of a time-domain classical TDI baseline in Figs. 12-14 strengthens the comparison, and the method involves no fitted constants or recycled equations; the gap advantage follows from a real difference in how invalid samples are treated. The main limitations are that the all-in-one setup is not used in the inference, the chunked likelihood is not validated for convergence, and the simulations use a single noise realization, injection, and gap pattern. These limitations are partly acknowledged in the text, but they currently bound the strength of the paper's central claim.","major_comments":[{"comment":"The inference results that support the paper's central claim use the 'reduced TDI-infinity framework,' where only the six laser noises are suppressed and optical-bench, clock, and modulation noises are deactivated; the all-in-one framework is validated only for noise suppression in Figs. 4-5 and 10. The title and abstract claim an 'all-in-one TDI-infinity' advantage under gaps, but no posterior comparison is made with all suppressible noises active. This is load-bearing because the all-in-one design matrix has a much larger dynamic range, and Fig. 10 shows that null-space algorithms leave substantially larger residuals in that setting. The authors should either add a gapped-inference case with the all-in-one setup after the re-normalization mentioned in the footnote to Sec. 5.4, or explicitly restrict the claims of the paper to the reduced framework.","section":"Section 5.4"},{"comment":"The chunked likelihood is used with l=500 and q=0.1 for all TDI-infinity results, but no convergence study over l and q is presented; the text itself defers this to future work in Sec. 5.5. Because chunks overlap, the log-likelihood in Eq. (55) double-counts data in the overlap regions, which inflates the effective sample size and can artificially narrow credible intervals, and the 125-s chunk length cannot represent OMS noise correlations below about 8 mHz even though the models in Eqs. (57)-(58) have a 2-mHz term. The no-gap comparison in Fig. 12 provides partial reassurance, but with gaps there is no full-covariance benchmark. A convergence test varying l and q, ideally including q=0 and a comparison against a full-covariance TDI-infinity likelihood for the gap case, is needed before the claimed gap robustness can be considered validated.","section":"Section 4.4, Eq. (55)"},{"comment":"All conclusions rest on one noise realization, one injected signal, and one gap pattern, with a relatively short MCMC run (60 walkers, 500 steps, burn-in 100, thinning by 15). The posterior shifts reported in Figs. 12-14 are therefore not separated from realization-specific noise fluctuations. At minimum, the authors should repeat the gap scenarios with several noise realizations, or report noise-whitened residuals, and ideally use a longer chain, so that the reader can assess whether the TDI-infinity stability in Fig. 13 is systematic rather than a single-draw artifact.","section":"Section 5.5"}],"minor_comments":[{"comment":"The statement that classical TDI loses 'approximately 400 seconds' of data should be quantified for the TDI-2 variables and the specific gap pattern, rather than presented as a visual claim from Figs. 7 and 8.","section":"Section 5.2"},{"comment":"The index expression 'ml·(i+1−qi)' is easy to misread; please define the integer ranges and the overlap convention explicitly in the text.","section":"Equations (43) and (56)"},{"comment":"The determinant normalization in Eq. (53) is not correctly preserved when overlapping chunks are summed in Eq. (55), so the quantity in Eq. (55) is a pseudo-likelihood that is not a properly normalized probability; this should be stated explicitly, since it affects model comparison even if parameter centroids are acceptable.","section":"Section 4.4, Eq. (55)"},{"comment":"The sentence about the discarded-bins check ('This plot is not shown here') should either be removed or replaced with the actual plot or a quantitative summary, because it makes a verification claim that the reader cannot check.","section":"Section 5.5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is readable and the algebraic null-space construction is sound, but the abstract and title overstate what is demonstrated: the gapped-inference comparisons use a reduced TDI-infinity setup rather than the all-in-one setup, and the chunked likelihood is not checked for convergence. If the authors add a chunk-convergence study and either demonstrate all-in-one inference or explicitly narrow the claims, the paper could be suitable for publication. The novelty relative to Vallisneri et al. 2021 and Houba et al. 2023 is incremental, but the gap-handling application for low-latency MBHB parameter estimation is of practical value to the LISA data-analysis community."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe essential point: this paper applies TDI-infinity to Bayesian inference for LISA data with gaps, and shows in simulations that it preserves posterior accuracy much better than classical TDI. That result is plausible, but the simulations depend on a chunked likelihood whose approximation error is not quantified. If the chunking bias is significant, the gap-robustness claim would be on shaky ground.\n\nWhat the paper does well: it formulates the likelihood in the time domain, removes invalid measurements from the design matrix before computing the null space, and handles gaps without interpolation or segment-wise likelihoods. The all-in-one TDI-infinity extension, covering optical bench, clock, and modulation noise, is a real new construction, though the authors admit it is numerically fragile on realistic orbits and therefore use only the laser-noise-reduced version for inference. That honesty is good. The comparison with time-domain TDI-2 in the no-gap case shows consistency and strengthens the reduced likelihood.\n\nThe soft spots: the chunked likelihood takes overlapping chunks (l=500, q=0.1), so overlapping data are counted multiple times. That inflates the effective sample size and can artificially narrow posteriors. Each chunk is 125 seconds, while the OMS noise model has power on much longer timescales; the covariance is truncated. The paper acknowledges the trade-off but gives no convergence study over l and q, and no independent validation of the gapped-case posteriors. The no-gap agreement with time-domain TDI-2 is reassuring but does not cover gaps, so the central comparison is not yet confirmed. One simulation, one noise realization, and no public code also limit strength.\n\nOverall, this is a serious, clear paper on a real problem. It deserves a serious referee, who should push for a chunking validation or a more modest claim. I would cite it as a promising direction, not a verified method.\n\nRecommendation: send to peer review.\n\nRegards,","headline":"Plausible Bayesian adaptation of TDI-infinity for gapped LISA data, but the headline gap-robustness claim rests on an unvalidated chunked likelihood.","tokens_in":35208,"tokens_out":3924,"would_cite":true,"duration_ms":37823,"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":"TDI-∞ keeps LISA parameter recovery stable through data gaps that cripple classical TDI.","keywords":["LISA","time-delay interferometry","TDI-infinity","data gaps","Bayesian inference","gravitational-wave parameter estimation","massive black hole binaries","null-space projection"],"falsifier":"Run the same Bayesian recovery on gapped data with chunk length $l$ varied from 250 to 2000 and overlap $q$ varied from 0 to 0.2; if the recovered mass and coalescence-time posteriors shift by more than the statistical uncertainty, or if biases grow monotonically as $l$ shrinks, the chunked-likelihood approximation is not unbiased.","tokens_in":1869,"feed_emoji":"🛰️","tokens_out":1935,"duration_ms":59905,"temperature":0.7,"pith_summary":"This paper claims that TDI-∞, a numerical variant of time-delay interferometry, handles LISA data gaps naturally and preserves astrophysical parameter estimation where classical TDI degrades badly. The authors build an all-in-one TDI-∞ pipeline that suppresses laser, optical-bench, clock, and modulation noise in a single null-space projection, then test Bayesian inference on one hour of gapped data containing a massive black hole binary merger. With six one-sample gaps near the merger, classical TDI corrupts roughly 400 seconds of data, while TDI-∞ loses only the gap samples themselves. The resulting posterior distributions for coalescence time and component masses remain stable with TDI-∞, whereas classical TDI shows significant spreading and bias, especially in the frequency-domain likelihood. A sympathetic reader would care because low-latency LISA alerts depend on analyzing short, possibly interrupted data streams where classical TDI's gap sensitivity is most damaging.","feed_headline":"TDI-infinity keeps LISA parameter recovery stable through data gaps","feed_subtitle":"Classical TDI corrupts roughly 400 seconds of data around six one-sample gaps; the new method loses only the gap samples.","key_machinery":"The load-bearing object is the null-space projection of the raw measurement vector. The measurement vector $y$ contains the telemetered beat-note signals from LISA's interferometers, the noise vector $p$ collects the suppressible noise sources, and the design matrix $M$ encodes their time-delayed coupling through fractional-delay finite-impulse-response filters. The TDI-∞ observable $o = T y$ is defined by the null space condition $T M = 0$, so it cancels all noises in $p$; gaps are handled by deleting invalid rows before the null space is computed. For long datasets the design matrix is chunked into overlapping submatrices with length $l$ and row overlap $q$, and the time-domain Gaussian likelihood uses a noise covariance $N = F^\\dagger \\tilde{N} F$ built from the raw-measurement power spectral densities. The turnback algorithm provides the sparse, banded null-space matrix $T$ that keeps the computation feasible and reveals six repeating generators, three of which preserve gravitational-wave signals and three of which are local and signal-insensitive.","core_discovery":"The paper's central claim is that TDI-∞ makes LISA gravitational-wave parameter inference robust to measurement interruptions without gap filling or interpolation. TDI-∞ replaces the algebraic TDI combinations with a numerical projection: the raw beat-note measurements are collected in a vector $y$, the modeled noise couplings form a design matrix $M$, and the TDI-∞ observable is $o = T y$ where $T$ spans the null space of $M^\\top$, so $T M = 0$ cancels all suppressible noises. Data gaps are handled by simply removing the corresponding rows of $y$, $M$, and the noise covariance before computing the null space, so the lost information is essentially just the missing samples. In the authors' simulations, classical TDI loses about 400 seconds of data around six one-sample gaps, while TDI-∞ loses roughly the six gapped samples; Bayesian posteriors for a massive black hole binary's coalescence time and masses stay stable with TDI-∞, and classical TDI posteriors spread and shift markedly. The paper also presents an all-in-one extension that suppresses optical-bench displacement, clock, and modulation noise in addition to laser noise, though the Bayesian inference runs use a reduced laser-only version to keep the design matrix numerically tractable.","pith_inferences":["If the chunked likelihood is unbiased, TDI-∞ could also simplify global-fit pipelines, where months of gapped data are currently handled with gap-filling or data augmentation; the same null-space trick would apply to any interruption pattern.","The paper does not study convergence over the chunk length $l$ and overlap $q$; a natural extension would map how posterior width and bias vary with these tuning parameters, which would tell whether the reported stability is robust.","The dynamic-range problem that forced the Bayesian runs to use a laser-only reduced framework suggests that full all-in-one inference needs renormalization of $M$; the authors note an initial version already works, so a testable next step is repeating the gapped-merger MCMC with all noise sources active.","Because TDI-∞ works in the time domain and removes gaps before projecting, it may naturally extend to non-stationary noise or time-varying arm lengths without the spectral-leakage corrections that plague frequency-domain TDI likelihoods."],"forward_implications":["LISA low-latency analyses of short data streams can proceed with essentially no extra data loss: each one-sample gap costs about one sample in TDI-∞, whereas classical TDI costs hundreds of seconds.","Bayesian parameter recovery with TDI-∞ remains stable with gaps near the merger, so early alerts for massive black hole mergers would not require waiting for retransmission of lost data packets.","Classical TDI in the frequency domain is particularly vulnerable to gaps and low-frequency noise leakage; time-domain classical TDI improves but still underperforms TDI-∞, so TDI-∞ removes the need for specialized gap-handling techniques.","The all-in-one formulation, once numerically stabilized, could cancel laser, optical-bench, clock, and modulation noise in a single projection, simplifying the L0-L1 processing chain.","Chunking makes TDI-∞ computationally practical for long LISA datasets, with the overlap factor $q$ trading off cross-correlation information against data double-counting."],"supporting_citations":[{"why":"Introduces TDI-∞ for a toy model, defining the null-space observable and its likelihood that this paper extends to LISA.","marker":"[34]"},{"why":"Earlier LISA adaptation of TDI-∞ for tilt-to-length noise, providing the basis for the null-space construction with realistic measurements.","marker":"[42]"},{"why":"Documents how data gaps affect detectability and parameter estimation of massive black hole binaries with LISA, motivating the comparison.","marker":"[29]"},{"why":"Presents a Bayesian data augmentation method for LISA gaps, representing the classical-TDI gap-handling approach this paper seeks to avoid.","marker":"[30]"},{"why":"Supplies the unified model for LISA measurements and instrument simulations used to build the design matrix $M$.","marker":"[40]"},{"why":"Provides the sparse null-space factorization algorithm used to construct the TDI-∞ projection matrix $T$.","marker":"[62]"},{"why":"Defines the A, E, T TDI channels used in the classical TDI likelihood and the concept of independent TDI generators.","marker":"[61]"},{"why":"Simulation tool for the LISA instrument used to generate noise models and TDI-2 data for the comparisons.","marker":"[78]"},{"why":"Python TDI package used to compute classical TDI channels and noise propagation for the benchmark cases.","marker":"[79]"}],"fun_headline_variants":["TDI-infinity keeps LISA parameter recovery stable through data gaps","New TDI method for LISA shrugs off data gaps in Bayesian inference","Gapped LISA data? TDI-infinity loses only the missing samples","All-in-one TDI-infinity: robust LISA inference despite gaps","LISA data gaps? TDI-infinity preserves parameter estimates"],"cache_read_input_tokens":37248,"weakest_assumption_plain":"The chunked likelihood approximation with chunk length $l=500$ and overlap $q=0.1$ preserves enough cross-correlation information to keep the posterior unbiased; the paper does not test convergence over $l$ and $q$.","fun_headline_variants_meta":{"raw":{"variants":["TDI-infinity keeps LISA parameter recovery stable through data gaps","New TDI method for LISA shrugs off data gaps in Bayesian inference","Gapped LISA data? TDI-infinity loses only the missing samples","All-in-one TDI-infinity: robust LISA inference despite gaps","LISA data gaps? TDI-infinity preserves parameter estimates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000893,"raw_usage":{"total_tokens":3945,"prompt_tokens":1138,"completion_tokens":2807,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":754,"completion_tokens_details":{"reasoning_tokens":2712}},"tokens_in":754,"tokens_out":2807,"duration_ms":19033,"temperature":1.0,"reasoning_tokens":2712,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:10:20.674886+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same Bayesian recovery on gapped data with chunk length $l$ varied from 250 to 2000 and overlap $q$ varied from 0 to 0.2; if the recovered mass and coalescence-time posteriors shift by more than the statistical uncertainty, or if biases grow monotonically as $l$ shrinks, the chunked-likelihood approximation is not unbiased.","supporting_citations":[{"cited_title":"Time-delay interferometry without delays","cited_arxiv_id":null,"evidence_quote":"Introduces TDI-∞ for a toy model, defining the null-space observable and its likelihood that this paper extends to LISA."},{"cited_title":"Time-delay interferometry infinity for tilt-to-length noise estimation in LISA","cited_arxiv_id":null,"evidence_quote":"Earlier LISA adaptation of TDI-∞ for tilt-to-length noise, providing the basis for the null-space construction with realistic measurements."},{"cited_title":"Effect of data gaps on the detectability and parameter estimation of massive black hole binaries with LISA","cited_arxiv_id":null,"evidence_quote":"Documents how data gaps affect detectability and parameter estimation of massive black hole binaries with LISA, motivating the comparison."},{"cited_title":"Gravitational-wave parameter estimation with gaps in LISA: A Bayesian data augmentation method","cited_arxiv_id":null,"evidence_quote":"Presents a Bayesian data augmentation method for LISA gaps, representing the classical-TDI gap-handling approach this paper seeks to avoid."},{"cited_title":"Unified model for the LISA measurements and instrument simulations","cited_arxiv_id":null,"evidence_quote":"Supplies the unified model for LISA measurements and instrument simulations used to build the design matrix $M$."},{"cited_title":"turnbackLU: Sparse matrix LU factorization, 2024","cited_arxiv_id":null,"evidence_quote":"Provides the sparse null-space factorization algorithm used to construct the TDI-∞ projection matrix $T$."},{"cited_title":"Prince, Massimo Tinto, Shane L","cited_arxiv_id":null,"evidence_quote":"Defines the A, E, T TDI channels used in the classical TDI likelihood and the concept of independent TDI generators."},{"cited_title":"LISA Instrument, November 2023","cited_arxiv_id":null,"evidence_quote":"Simulation tool for the LISA instrument used to generate noise models and TDI-2 data for the comparisons."},{"cited_title":"PyTDI, 2023","cited_arxiv_id":null,"evidence_quote":"Python TDI package used to compute classical TDI channels and noise propagation for the benchmark cases."}],"review_version":1}