{"id":"5e8eb86c-a08e-4a8a-9a67-1fe8c94ff90e","arxiv_id":"2412.01219","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A new likelihood function estimates GWB anisotropy angular power spectra directly from detector data, and cross-correlation with the CMB can make the quadrupole measurable with four years of LISA data.","lead":"The paper introduces a Bayesian method that estimates the angular power spectrum of gravitational wave background (GWB) anisotropy directly from detector data, optionally using cross-correlation with the cosmic microwave background to sharpen the measurement. On LISA mock data, the method finds that four years of data alone cannot detect low multipoles, but with a strong GWB-CMB correlation it can recover the quadrupole.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (3.8) marginalizes over an independent GWB realization for each time–frequency pixel, although the model (Eq. 2.3) has one frequency-independent realization shared across all data; the main likelihood (3.9) is not the marginal likelihood of the stated model.","rationale":"The paper's strongest claim is that Eq. (3.9) is a compact marginalized likelihood connecting detector data to the angular power spectrum, and that with r=1 LISA 4-year data give unbiased quadrupole estimates. For this claim to hold, Eq. (3.9) must be the marginal likelihood of the generative model described in Sections 2–3. The derivation in Eq. (3.8) integrates over the intensity multipoles separately for every (t,f) pair, which is only correct if the sky realization is independent in each time–frequency pixel. But the text is explicit that a^GW_ℓm are frequency-independent, and the injection uses a single sky map. The same realization therefore enters the means of all time segments and frequency bins, so the correct marginal likelihood has off-diagonal correlations in (t,f) that are absent from Eqs. (3.9)–(3.11). This is an internal inconsistency, not a disagreement with external consensus. The reader's weakest assumption identified a different misspecification (Gaussianity of D and the unvalidated second-order inverse expansion); both are real, but the independent-marginalization issue is more load-bearing because it affects the analytical likelihood itself. The proposed concrete test settles the question in a small controlled setting; if the factorized likelihood differs materially from the exact joint marginalization, the numerical results of Section 4 cannot be used to support the headline. This keeps the manuscript under the reader's CONDITIONAL verdict, with the condition sharpened: the marginalization must be performed over a single shared realization, or the equivalence of the factorized form to that model must be demonstrated.","tokens_in":17528,"tokens_out":20294,"duration_ms":211623,"concrete_test":"Construct a small two-level Gaussian model with one shared realization a: D_{tf}=Γ_{tf} K_f a + n_{tf}, where n_{tf} are iid N(0,C_D) and a~N(0,C_l). For a small grid (e.g., 2 time segments × 2 frequency bins, using a single ℓm or ℓmax=2), compute the exact marginal likelihood L_exact = N(0, blockdiag(C_D) + G C_l G†), with G_{tf}=Γ_{tf} K_f, and the paper's factorized likelihood L_paper = ∏_{tf} N(0, C_D + Γ_{tf} K_f^2 C_l Γ†_{tf}). If the posterior for C_l from L_paper differs from L_exact by more than, say, 20% in median or credible-interval width, then Eq. (3.9) is not the likelihood of the stated model and the quadrupole recovery claim in Fig. 3 is unsupported. Equivalently, compare the off-diagonal blocks of the exact data covariance to the block-diagonal form implied by Eq. (3.10) and check whether their norm exceeds 10% of the diagonal blocks in signal-dominated bins.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central result (3.9) is derived in Eq. (3.8) by taking a product over time frames and frequency bins of separate Gaussian integrals over the anisotropic intensity multipoles I_{ℓm,f}. This is equivalent to assuming that every (t,f) pixel has its own independent sky realization. That contradicts the model stated in Section 2: I_{ℓm}(f)=K_f a^GW_{ℓm} with a^GW_{ℓm} frequency-independent, and the sky is a single realization shared by all time segments. Conditioned on the shared realization, the data D_{t,f} have means Γ_{t,f} I_f and are independent across (t,f) within the Gaussian-D approximation of Section 3, but after marginalizing over one shared I the data are correlated: Cov(D_{t,f},D_{t',f'}) gains the off-diagonal block Γ_{t,f} K_f C_l K_{f'} Γ†_{t',f'} for (t,f)≠(t',f'). Eq. (3.10) contains only the diagonal blocks C_D + Γ K_f^2 C_l Γ† and factorizes over (t,f), so it cannot be the marginal likelihood of the stated model. The numerical injections in Section 4 use one a^GW realization and one sky map (lower panel of Fig. 1), so the analysis is mismatched to the injected data. This is more fundamental than the numerical expansion issue noted by the reader: even with exact matrix inverses, Eq. (3.9) does not correspond to the model being fit. Treating every time–frequency pixel as an independent sky realization artificially increases the number of independent sky modes, which can shrink the effective cosmic variance and bias the recovered angular power spectrum; the claimed unbiased quadrupole recovery in Fig. 3 is therefore not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a Bayesian likelihood for the angular power spectrum C_l^GW of a gravitational-wave background anisotropy, derived directly from short-time Fourier transform detector data and optionally conditioned on a cosmological tracer Y through a Gaussian cross-correlation. The central analytic result is Eqs. (3.9)-(3.11), obtained by marginalizing over the anisotropic intensity multipoles I_lm. The method is demonstrated on four-year LISA mock data generated with the schNell response model. The main numerical findings are that, without cross-correlations, the multipoles cannot be constrained, whereas with r^{GW x CMB}=1 the quadrupole C_tilde_2^GW is recovered at log10 C_tilde = -2.4, and at r=0.85 it is recovered within 1 sigma. The authors conclude that cross-correlations can boost the Bayesian inference of GWB anisotropy.","tokens_in":17872,"tokens_out":9528,"duration_ms":94074,"significance":"If the derivation and numerical validation are correct, the approach would be a useful alternative to sky-map-based Bayesian methods, avoiding high-dimensional map sampling and providing a way to incorporate cross-correlations with CMB or large-scale structure in GWB anisotropy studies. Strengths include the analytic Gaussian marginalization, the general detector-pair formalism that does not assume diagonal TDI channels, the use of a realistic time-dependent LISA response through the public code schNell, and the explicit reduction to earlier no-cross-correlation likelihoods when C^{GW x Y}=0. However, the numerical validation rests on a likelihood that is not the marginal likelihood of the stated single-realization model, and the unbiasedness claim is supported by a single injected realization. These issues materially weaken the evidence for the headline result, though the framework appears salvageable.","major_comments":[{"comment":"The marginalization in Eq. (3.8) is performed separately for each time-frequency pixel, with the prior (3.2) applied independently at each t and f. In the model of Sec. 2 and Eq. (2.3), however, I_lm(f)=K_f a_lm^GW with a single frequency-independent realization a_lm^GW. Conditional on that shared realization, D_{t,f} and D_{t',f'} are correlated across frequency, with off-diagonal covariance blocks Gamma_{t,f} K_f C^{GW|Y} K_{f'} Gamma^dagger_{t',f'}. The covariance in Eq. (3.10) contains only diagonal blocks and the likelihood (3.9) factorizes over (t,f), so it is the marginal likelihood of a model in which every time-frequency pixel has an independent sky realization, not the model used in the injections of Sec. 4, which uses one sky map (lower panel of Fig. 1). This can artificially increase the effective number of independent sky modes, leading to overconfident or biased posterior intervals. Even with exact matrix inverses, Eq. (3.9) does not correspond to the stated model; the numerical results in Figs. 2-4 therefore do not yet validate that model.","section":"Sec. 3, Eqs. (3.8)-(3.10)"},{"comment":"The numerical inversions replace (C_D + Gamma C_I Gamma^dagger)^{-1} by its second-order expansion, C_D^{-1} - C_D^{-1} Gamma C_I Gamma^dagger C_D^{-1}, and similarly the related expressions, without a convergence check. Since Eq. (3.9) uses both the inverse and the determinant of this matrix, uncontrolled truncation can bias the posterior, especially when Gamma C_I Gamma^dagger is not small compared with C_D. The paper should validate the expansion against exact inversion for the fiducial parameters, or use a numerically stabilized direct inversion, and quantify the error before the recovery plots are accepted.","section":"Sec. 4 and Appendix C, Eqs. (C.8)-(C.9)"},{"comment":"The claim of 'unbiased estimations' is supported by a single injected realization at log10 C_tilde_2 = -2.4. Unbiasedness is a frequentist property of an estimator and cannot be established from one posterior whose mode happens to lie near the fiducial value. The paper should provide an ensemble of injected realizations and a coverage check before making the unbiasedness claim.","section":"Sec. 4, Fig. 3"},{"comment":"Eq. (3.1) treats the cross-spectrum estimator D_AB = d_A d_B* as a complex Gaussian variable with covariance (2.8). For a single STFT segment, the product of two Gaussian variables is not Gaussian, and the covariance (2.8) keeps only monopole contributions to the intensity. The text acknowledges the Gaussian approximation but does not quantify its accuracy for the short T_seg used here; a validation against simulated D distributions would substantiate this assumption.","section":"Sec. 3, Eq. (3.1)"}],"minor_comments":[{"comment":"The heading of Appendix A reads 'F rom the anisotropic...' with an extra space; please correct the typographical error.","section":"Appendix A"},{"comment":"The barred quantities denoting the conditional mean are visually very similar to the unbarred intensity symbols in the compiled text; please use a more distinct notation and define all barred variables explicitly.","section":"Eq. (3.5)"},{"comment":"The priors used for log10 C_tilde_l and r in the MCMC sampling are not stated; please specify them.","section":"Sec. 4"},{"comment":"The statement that 'an upper limit for the relative correlation cannot be established' would be more precise with a quantitative criterion, such as a reported 95% upper bound.","section":"Sec. 4, Fig. 4"},{"comment":"No code availability statement is provided; releasing the likelihood implementation and MCMC scripts would improve reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The skeptical concern about Eq. (3.8) is, in my reading, correct and is the main obstacle to accepting the paper. I do not recommend rejection because the Gaussian-marginalization technique can in principle be rederived for the shared-realization model and the numerical tests redone, but the current numerical results should not be relied upon until that is done."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the central likelihood in Eq. (3.9) marginalizes over a different model than the one Section 2 states. The stress-test note is right, and this is not a minor fix.\n\nThe genuinely new piece is the idea of building a Bayesian likelihood for the GWB angular power spectrum that can ingest a cross-correlation with an external tracer like the CMB. The conditional prior (3.2)-(3.4) is standard Gaussian conditioning and is written cleanly. The no-cross-correlation limit reducing to Gair et al. and van Haasteren & Levin is correctly acknowledged. Using realistic LISA response from schNell in a mock demonstration is also a reasonable choice.\n\nThe problem is that Eq. (3.8) integrates a separate I_lm,f for every (t,f) pixel. The model in Eq. (2.3) has one frequency-dependent realization shared across all time segments, and one a_lm shared across frequency. Marginalizing over one shared realization couples all pixels through the response; Eq. (3.10) contains only diagonal blocks in (t,f) and cannot be that marginal likelihood. The injections in Section 4 use one sky map, so the analysis is mismatched to the injected data. Treating every pixel as independent inflates the effective number of sky modes, which can shrink variance and bias the recovered C_l; the claimed unbiased quadrupole recovery in Fig. 3 is therefore not established. This is more fundamental than the reader's second-order expansion concern, though that concern is also real: the 2nd-order expansion in (3.10) is unvalidated and could matter. There is no released code, and the \"unbiased\" statement rests on a single realization.\n\nWhat should be credited: the derivation up to Eq. (3.7) is fine, and the conditional prior machinery is portable. If the marginalization is redone correctly, the method could be useful for LISA, TianQin, and PTA analyses.\n\nBottom line: this deserves serious refereeing, not desk rejection, because the core idea is relevant and the flaw is identifiable. But the paper in its current form overstates what it demonstrates. The fix is a proper joint marginalization, followed by a multi-realization check and code release.\n\nWho should read it: methodologically inclined GWB data-analysis people; it is a good discussion piece for a group meeting, mostly as a cautionary example.","headline":"The cross-correlation idea is worth a look, but the main likelihood marginalizes over independent per-pixel skies, so the claimed quadrupole recovery is not established.","tokens_in":18437,"tokens_out":4248,"would_cite":false,"duration_ms":40129,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A Bayesian likelihood connects LISA detector data straight to the GWB angular power spectrum, and shows that cross-correlating with the CMB is what makes the quadrupole detectable in 4-year data.","keywords":["stochastic gravitational wave background","GWB anisotropy","angular power spectrum","Bayesian inference","cross-correlation","LISA","cosmic microwave background","quadrupole"],"falsifier":"A decisive check would be to run many independent 4-year LISA mock realizations at the fiducial $r=1$ quadrupole and measure how often the posterior's 68% credible interval contains the injected value; if the true value lands outside that interval in substantially more than 32% of runs, the Gaussian or monopole-only covariance assumption is wrong.","tokens_in":17278,"feed_emoji":"🌌","tokens_out":13300,"duration_ms":105662,"temperature":0.7,"pith_summary":"This paper sets out to establish that the angular power spectrum of the gravitational wave background (GWB) can be inferred directly from a detector's short-time Fourier transform data, without first reconstructing a sky map, through an analytic likelihood that marginalizes over unobserved sky realizations. It also tries to show that using the known cross-correlation between GWB anisotropies and another cosmological tracer such as the CMB materially strengthens the inference. With realistic LISA response and noise models, the authors argue that 4-year LISA data alone cannot significantly constrain low multipoles, but that a perfectly correlated CMB tracer enables an unbiased recovery of the quadrupole moment, with a partial correlation of $r=0.85$ still recovering the fiducial value within $1\\sigma$. If correct, this gives a generic, detector-agnostic pipeline for anisotropy science across current and future gravitational wave observatories.","feed_headline":"GWB-CMB correlation unlocks LISA's quadrupole detection","feed_subtitle":"New likelihood reads multipoles straight from detector data; 4-year LISA suffices if the GWB tracks the CMB.","key_machinery":"The load-bearing object is the compact marginalized likelihood of eqs. (3.9)--(3.11), assembled from three pieces: the detector likelihood for the cross-power spectra, whose covariance keeps only monopole contributions to the intensity; the conditional prior on the harmonic coefficients, whose mean is the tracer-scaled cross-spectrum and whose variance is reduced by the factor $1-(r^{\\mathrm{GW}\\times Y})^2$; and the analytic marginalization over the $I_{\\ell m}$ using the complex Gaussian integration formula. The result adds a response-weighted harmonic-prior covariance to the detector covariance, and subtracts the tracer-conditional mean and the noise from the data before weighting by the inverse. In the numerical demonstration the inverse is evaluated by expanding the response-weighted prior term to second order, and the monopole spectrum is fixed before sampling the multipoles.","core_discovery":"The central claim is that the marginalized likelihood for the full dataset factorizes over time frames and frequency bins as $L(D|\\mathrm{GW}\\times Y) \\propto \\prod_{t,f} |\\pi C_{ab}|^{-1} \\exp(-J_a^\\dagger C_{ab}^{-1} J_b)$, where $C_{ab} = (C_D)_{ab} + \\Gamma_{a\\mu}(C_I)_{\\mu\\nu}\\Gamma^*_{\\nu b}$ and $J_a = D_a - \\Gamma_{a\\mu}Z_\\mu - N_a$ (eqs. 3.9--3.11). Here $C_D$ is the noise-plus-monopole covariance of the detector cross-power spectra, $C_I$ is the prior covariance of the harmonic coefficients conditioned on the tracer, and $\\Gamma_{a\\mu}$ encodes the detector response to each multipole. The form is obtained by multiplying the detector likelihood (3.1) with the conditional harmonic-space prior (3.2) and marginalizing over all $I_{\\ell m}$ with a complex Gaussian integral plus a standard matrix-inversion lemma. Applied to injected LISA data with $\\Omega_{\\mathrm{GW}}(f)=10^{-10}(f/10^{-3}\\,\\mathrm{Hz})$, scale-invariant $\\widetilde{C}^{\\mathrm{GW}}_\\ell$, and a perfectly correlated CMB map, the paper finds the posterior on $\\log_{10}\\widetilde{C}^{\\mathrm{GW}}_2$ is unbiased at the fiducial value $-2.4$ and remains possible down to about $-4$; without cross-correlation, 4-year data gives no significant multipole constraint.","pith_inferences":["The paper notes that the multi-tracer generalization is straightforward but does not implement it; chaining the conditional prior over CMB, CMB lensing, and galaxy-overdensity maps together would likely tighten the quadrupole constraint beyond any single tracer.","The Gaussianity of the cross-power spectra is the softest internal assumption; replacing the Gaussian detector likelihood with a distribution for averaged cross-spectra and re-deriving the marginalized likelihood would show how much of the claimed sensitivity depends on that approximation.","At higher multipoles the monopole-only covariance approximation should degrade, so the cleanest regime for this scheme is low-$\\ell$; computing the exact versus second-order inverse of eq. (3.10) as a function of $\\ell$ would identify where the scheme stops being reliable.","A successful quadrupole recovery with $r=1$ would let LISA act as another observer of the same long-wavelength perturbations seen by the CMB, making the recovered cross-spectrum a consistency test of the cosmological origin of the background."],"forward_implications":["Four years of LISA data, at the injected cosmological amplitude and with no cross-correlation, do not significantly constrain the quadrupole or hexadecapole; the posteriors are broad and consistent with zero.","With a perfectly correlated CMB tracer ($r=1$), the quadrupole $\\log_{10}\\widetilde{C}^{\\mathrm{GW}}_2$ is recovered without bias at the fiducial $-2.4$ from 4-year data, and reconstruction remains possible down to roughly $-4$.","Even a partial correlation of $r=0.85$ still recovers the fiducial quadrupole within $1\\sigma$, and the likelihood does not require a diagonalized noise covariance, so it applies to general detector arrays rather than only to time-delay-interferometry channels.","The scheme bypasses full sky-map reconstruction, and the authors report that posterior sampling completes within hours, making multipole-level Bayesian anisotropy inference computationally practical.","Because the likelihood is generic and detector-agnostic, the same pipeline transfers to ground-based interferometer networks, other space-based missions, and pulsar timing array experiments, provided foreground cleaning is applied first."],"supporting_citations":[{"why":"Supplies the LISA time-dependent response function and noise power spectral density used for the mock-data demonstration.","marker":"[43]"},{"why":"Provides the model prediction that GWB anisotropies are strongly correlated with the CMB, motivating the injected r=1 scenario.","marker":"[44]"},{"why":"Gives the conversion between GW intensity multipoles and fractional-energy-density multipoles and the LISA anisotropic-background framework.","marker":"[52]"},{"why":"The earlier CMB-style mapping likelihood that the new analytic marginalization connects to and generalizes.","marker":"[55]"},{"why":"Supplies the LISA noise model and the intensity-to-energy-density relations used in eqs. (2.2) and (2.3).","marker":"[64]"},{"why":"Documents the slow-convergence problem in sky-map reconstruction that motivates direct angular-power-spectrum inference.","marker":"[57]"},{"why":"Provides the no-cross-correlation likelihood that eq. (3.9) reduces to when the GWB-tracer cross-spectrum is set to zero.","marker":"[70]"},{"why":"Defines the forward model of space-borne detector response based on spacecraft positions, used to compute the response functions.","marker":"[74]"}],"fun_headline_variants":["Bayesian GWB anisotropy: CMB cross-correlation sharpens LISA quadrupole","Cross-correlating GWB with CMB yields LISA quadrupole in 4 years","New Bayesian method reads GWB anisotropy from LISA time series","LISA quadrupole from GWB-CMB cross-correlation: Bayesian boost","GWB anisotropy quadrupole via CMB cross-correlation: LISA in 4 years"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the detector cross-power measurements fluctuate like Gaussian variables whose spread is set almost entirely by the uniform monopole part, and that a short-cut expansion used to invert the covariance matrix is accurate; if either fails, the recovered power-spectrum estimates could be biased or look more precise than they are.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian GWB anisotropy: CMB cross-correlation sharpens LISA quadrupole","Cross-correlating GWB with CMB yields LISA quadrupole in 4 years","New Bayesian method reads GWB anisotropy from LISA time series","LISA quadrupole from GWB-CMB cross-correlation: Bayesian boost","GWB anisotropy quadrupole via CMB cross-correlation: LISA in 4 years"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000719,"raw_usage":{"total_tokens":3286,"prompt_tokens":1061,"completion_tokens":2225,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":677,"completion_tokens_details":{"reasoning_tokens":2115}},"tokens_in":677,"tokens_out":2225,"duration_ms":13337,"temperature":1.0,"reasoning_tokens":2115,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:34:05.149549+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive check would be to run many independent 4-year LISA mock realizations at the fiducial $r=1$ quadrupole and measure how often the posterior's 68% credible interval contains the injected value; if the true value lands outside that interval in substantially more than 32% of runs, the Gaussian or monopole-only covariance assumption is wrong.","supporting_citations":[],"review_version":1}