{"id":"28998dac-c605-41f4-8dbc-7270ca1f7374","arxiv_id":"2412.03503","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A GAM-based reparameterization is shown to reduce the sensitivity of cosmological constraints from EFT-of-LSS analyses to the choice of nuisance parameter priors.","lead":"This paper introduces a statistical reparameterization, based on Generalized Additive Models, that separates nuisance parameters from cosmological parameters in EFT-based analyses of galaxy clustering. In 100 simulated datasets the method reduces the bias that nuisance priors induce on cosmological constraints, and iterating the procedure shrinks most biases below one sigma.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing gap is that sample-level orthogonalization of the conditional mean E[N|C] (Eq. 2.5) does not imply likelihood separability; residual C-dependent higher-order coupling can keep nuisance priors in the C marginal, and Sec 6 concedes residual coupling.","rationale":"The reader's weakest assumption correctly identifies the transfer from sample-level orthogonality to likelihood separability as load-bearing. My concern sharpens this: even a perfect GAM for E[N|C] is insufficient, because separability also requires that the C-dependent conditional variance and higher moments of N' vanish or factor. The reader emphasizes GAM fidelity, basis choice, and training-region coverage; those are real but secondary. The paper deserves credit for a careful simulation ensemble, MAP-bias statistics, coverage checks, and the iterative width test, all of which support 'reduced sensitivity' rather than the stronger 'no dependence on any independent nuisance prior.' Because the reader already conditions the verdict on exactly this gap, my reading does not change the verdict: CONDITIONAL remains appropriate. The abstract and conclusions should be softened to claim substantial reduction in prior sensitivity, with the residual coupling explicitly quantified under varied N' priors.","tokens_in":19804,"tokens_out":5900,"duration_ms":70382,"concrete_test":"Take 10 of the 100 simulated datasets and rerun the full single-iteration pipeline twice, with the current N(0,1) priors on N' and with a different independent prior on N', e.g., Uniform[-3,3] or N(1,4). If the stacked MAP values or 95% CI coverage for the cosmological parameters shift by more than the ensemble scatter between runs, the claim that the C marginal is insensitive to any independent nuisance prior fails. To isolate the mechanism, also run a synthetic likelihood y ~ Normal(C*N, sigma^2) with known C,N, use the analytic E[N|C] as the oracle GAM, and repeat with the two different N' priors; a shift in the C posterior would directly falsify the assumption that orthogonalizing the conditional mean yields likelihood separability.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that after the GAM rotation the likelihood factorizes, f(y|C,N') = f_C(y|C) f_N(y|N'), so that integrating over N' with any independent prior leaves the C posterior unchanged. What the construction actually guarantees is approximate zero conditional mean of N' given C in the preliminary posterior samples (Sec 2.1, Eq. 2.5): E[N'|C] is small. That is not conditional independence, and it does not imply likelihood factorization. Even an oracle GAM that perfectly recovers E[N|C] leaves the conditional variance and higher moments of N'|C,y potentially C-dependent. In a Gaussian likelihood with model A(C)+D(C)N, the N' quadratic term in the log-likelihood is weighted by D(C)^T Σ^{-1} D(C); if D(C) varies with C, that term does not factor, so changing the width or location of π(N') changes the C posterior. The EFT model has such C-dependence because b1 enters the kernels nonlinearly (Eqs. 3.1-3.2); counterterms and stochastic terms are linear, but the bias parameter is not. The EFT validation uses only one transformed-basis prior, N(0,1) on N', so the 'any independent nuisance prior' claim is not actually tested in the main application. The paper itself states in Sec 6 that residual coupling remains for non-additive dependence, so the strongest form of the abstract claim is already narrower than the demonstrated result. What the evidence supports is a reduction, not elimination, of nuisance-prior sensitivity.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a reparameterization method, based on Generalized Additive Models (GAMs), that aims to reduce the sensitivity of cosmological parameter posteriors to the choice of nuisance parameter priors in Effective Field Theory (EFT) analyses of large-scale structure. The method first fits a preliminary model, then fits a GAM for the conditional mean of each nuisance parameter given the cosmological parameters, and finally reparameterizes the model so that the new nuisance parameters are residuals from this fit. The authors demonstrate the approach on 100 simulated galaxy power-spectrum datasets, comparing the pre-transformed basis (PTB) with linear-model (LM) and GAM-based transformations, and report reduced biases in the maximum a posteriori (MAP) estimates for several cosmological parameters, particularly ln 10^10 A_s and w0. They also present an iterative version of the transformation and several tests intended to show that the method does not inject information or systematically underestimate uncertainties.","tokens_in":20054,"tokens_out":5002,"duration_ms":48555,"significance":"The method is practically motivated and, if the claims are supported, could provide a useful tool for EFT-based cosmological analyses, where nuisance-prior projection effects are known to be substantial. The empirical evaluation is a strength: 100 realizations with known true parameter values, frequentist MAP-bias statistics, a test involving 100 iterations to check for uncertainty shrinkage (Fig. 9), and a comparison of average posterior widths to the empirical scatter of posterior means (Table 5). These tests go beyond what is common in method papers and provide meaningful evidence that the transformation does not simply shrink posteriors. However, the central theoretical claim—that after reparameterization the likelihood becomes approximately separable and hence the marginal posterior for cosmological parameters is insensitive to any independent nuisance prior—is not established by the construction, which only enforces approximate zero conditional mean of the new nuisance parameters given the cosmological parameters. The paper's own Sec. 6 concedes residual coupling.","major_comments":[{"comment":"The claim that after reparameterization 'any independent nuisance parameter prior will not affect the marginal posterior for cosmological parameters' (Sec. 1) is not supported by the construction in Sec. 2.1. The transformation N' = N - B(C)β ensures only that the conditional mean of N' given C is approximately zero in the preliminary posterior samples. Conditional independence, which is required for likelihood separability f(y|C,N') = f_C(y|C) f_N(y|N'), is a much stronger condition; the conditional variance and higher moments of N'|C can still depend on C, and through them the prior on N' can influence the C marginal. In the EFT model, the nuisance parameter b1 enters the redshift-space kernels nonlinearly (Eqs. 3.1-3.2), so such C-dependent coupling is concrete. The paper itself acknowledges in Sec. 6 that residual coupling remains for non-additive dependence. Please reword the abstract and introduction to claim reduction rather than elimination of nuisance-prior sensitivity, and add a quantitative test of residual sensitivity, e.g., by repeating the Step 3 fit with several distinct priors on N' (varying width and location) and checking stability of the C posteriors.","section":"Sec. 1 and Sec. 2.1, Eq. (2.5)"},{"comment":"The text in Sec. 5 states that the authors 'report also relative bias, in units of the standard deviation, and 95% credible interval coverage' for the MAP estimates across the 100 datasets, but Table 3 lists only σ and Δ[σ] and no coverage values are given anywhere in the paper. Credible-interval coverage is a key frequentist diagnostic for whether the reparameterization changes the calibration of the reported uncertainties, and it is directly relevant to the paper's claim that the method does not misrepresent uncertainty. Please either add coverage rates to Table 3 (or a new table/figure) or remove the claim that coverage is reported.","section":"Sec. 5, Table 3"},{"comment":"The no-information-injection test iterates on a single dataset, and the r statistics in Table 5 are averages over the 100 realizations but do not probe sensitivity to different choices of the nuisance prior. Throughout the paper, the only prior used on the rotated nuisance parameters N' is a standard normal N(0,1). Consequently, the central assertion that the marginal posterior for cosmological parameters does not depend on 'simple priors placed on nuisance terms' is not empirically tested. To support the headline claim, the authors should run the reparameterized fit with at least two or three different priors on N' (e.g., N(0, α²) with α varied by a factor of a few, and a shifted Gaussian) and show that the cosmological posteriors remain stable. If they do not remain stable, the paper should explicitly state the range of priors for which the method is robust.","section":"Sec. 5.1, Fig. 9 and Table 5"},{"comment":"The method uses the same data to determine the transformation in Steps 1 and 3, and the prior on N' is likelihood-dependent, a fact the authors acknowledge in Sec. 6 as a violation of the strong likelihood principle. This is not inherently disqualifying, but it means the procedure is best characterized as a data-adaptive stabilization of posterior inference rather than a standard Bayesian analysis with a fixed prior. The paper should state this scope more prominently and discuss the risk of overfitting to the noise realization more thoroughly. The current checks (Fig. 9 and Table 5) are suggestive but not conclusive; for example, Fig. 9 examines only one dataset and Table 5's r statistic could remain near 1 even if the method systematically shifts posteriors in a prior-dependent way.","section":"Sec. 2.2, Step 3 and Sec. 6"}],"minor_comments":[{"comment":"After Eq. (2.2), 'orthogonal parameterization' should be 'orthogonal reparameterization' for consistency with the rest of the paper.","section":"Sec. 2.1"},{"comment":"The estimation of the scaling factor α is described in one sentence; a formula or a short pseudocode block would make the procedure reproducible and easier to follow.","section":"Sec. 2.2, Step 3"},{"comment":"When describing the 100 generated datasets, please state explicitly that the noise realizations are independent draws from the same covariance matrix, and confirm that the covariance matrix is fixed (not re-estimated per realization).","section":"Sec. 4"},{"comment":"The caption says 'The red contours shows' — this should be 'show'.","section":"Appendix B, Fig. 11"},{"comment":"The column headers 'iter 1' through 'iter 6' are not explained in the caption; please define them as the iterations of the GAM iterative approach described in Sec. 4.1.","section":"Table 4"},{"comment":"The term '2Z1(µ)P11(k)' immediately before the counterterm bracket appears to be missing a multiplication symbol; please check the typesetting.","section":"Sec. 3, Eq. (3.1)"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a practically useful technique with a strong empirical evaluation, but the theoretical framing overstates what the method guarantees. The central likelihood-separability claim needs either a rigorous justification or a clear downgrade to a claim about reducing projection effects. I would recommend that the editor seek a statistician's opinion on Sec. 2.1 and the relationship between conditional-mean orthogonalization and posterior independence, as this is the load-bearing point. The missing coverage statistics in Table 3 and the lack of multi-prior tests are straightforward to fix and would substantially strengthen the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is worth taking seriously: instead of choosing a particular nuisance prior, reparameterize so that the nuisance parameters are approximately orthogonal to the cosmological ones in the posterior, using a GAM fit to preliminary posterior samples. This extends the classical linear orthogonalization of Cox and Reid and Papaspiliopoulos et al. to a data-adaptive non-linear setting, and the EFT-of-LSS application is timely. The empirical calibration is the right kind of evidence: 100 simulated realizations with known truth, MAP-bias statistics, coverage checks, and an explicit no-information-injection test with 100 iterations that does not show shrinking widths. The single-iteration ensemble results are genuinely encouraging: the bias on ln 10^10 As drops from -1.1 to -0.3 sigma, and on w0 from -1.0 to -0.5 sigma. The GAM prediction check in Appendix A is a nice sanity test.\n\nThe soft spots are real but mostly addressable. The load-bearing gap is the step from sample-level decorrelation to likelihood separability. What the construction guarantees is approximate zero conditional mean of N' given C in the preliminary posterior samples. That does not imply that the likelihood factorizes, because the conditional variance and higher moments of N'|C can still depend on C. In a Gaussian likelihood where the nuisance coefficient itself is C-dependent (as b1 is in the EFT kernels), that C-dependence does not integrate out, so independent priors on N' can still project onto the cosmological posterior. The paper itself concedes residual coupling in Sec 6, so this is acknowledged, but it means the abstract's phrasing that the marginal posterior 'does not depend' on nuisance priors is too strong.\n\nThere are also smaller issues. Table 4 shows h and omega_c with biases around 0.8-1.0 sigma across iterations, so the 'below 1 sigma' claim in the abstract and conclusions is not supported by the paper's own table for all parameters. The iterative results come from a single dataset, which is weak evidence for a general claim. No code is shipped, which makes the method harder to adopt and verify. And only one transformed-basis prior (standard normal) is tested, so the 'any independent nuisance prior' claim is not actually exercised.\n\nWho gets value from this? Anyone doing EFT-based full-shape analyses with nuisance priors, and Bayesians working on parameter orthogonalization. The paper deserves a serious referee, but the authors should be asked to soften the headline claim, add ensemble iterative results, release the pipeline, and test at least one non-additive dependence or a different N' prior. The method likely reduces projection effects; that is already a useful contribution.","headline":"Useful method for reducing nuisance-prior projection effects in EFT-of-LSS, but the 'prior-independence' claim is stronger than what the evidence supports; the paper demonstrates reduction, not elimination.","tokens_in":20719,"tokens_out":1721,"would_cite":true,"duration_ms":19757,"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 rotation of nuisance parameters makes EFT cosmological fits insensitive to nuisance priors.","keywords":["nuisance prior sensitivity","parameter reparameterization","Generalized Additive Models","EFT of large-scale structure","projection effects","Bayesian cosmology","galaxy power spectrum"],"falsifier":"Run the pipeline on the 100 mocks with a deliberately shifted or much wider nuisance prior, e.g. a uniform prior of width 100 on every EFT parameter instead of the standard priors in Table 2, and measure the shift in the stacked MAP for $\\ln 10^{10} A_s$ and $w_0$ in the GAM-transformed basis: if the shift remains comparable to the un-transformed bias, the claimed independence has failed.","tokens_in":19479,"feed_emoji":"📐","tokens_out":3368,"duration_ms":27023,"temperature":0.7,"pith_summary":"This paper proposes a reparameterization technique that removes the correlation between cosmological parameters and nuisance parameters in the likelihood of EFT-based large-scale structure analyses. The method fits a flexible regression of each nuisance parameter on the cosmological parameters using Generalized Additive Models, then redefines the nuisance parameters as the residuals of that fit. After the transformation, the likelihood becomes approximately separable, so the marginal posterior for cosmological parameters no longer depends on the choice of prior for the nuisance parameters. The authors demonstrate on 100 simulated galaxy power-spectrum datasets that this removes prior-driven biases, especially on the scalar amplitude and dark-energy equation of state, without broadening the constraints.","feed_headline":"Rotating nuisance parameters frees EFT cosmology from prior choice","feed_subtitle":"A GAM-based reparameterization removes correlations that let nuisance priors bias cosmological constraints.","key_machinery":"The load-bearing mechanism is the Generalized Additive Model conditional-expectation fit, $E[N_j|C] \\approx \\sum_l f_{lj}(C_l)$, with each $f_{lj}$ expanded in 20 penalized cubic B-splines and fitted by penalized least squares to the posterior samples of a preliminary run. The transformed nuisance parameter is the residual $N'_j = N_j - E[N_j|C]$, which is approximately uncorrelated with every smooth function in the model; when the true dependence is additive, this removes the dependence of the likelihood on $N'_j$ from the marginal cosmological posterior.","core_discovery":"The central claim is that an additive, data-adaptive fit of the conditional mean $E[N_j | C]$ of each nuisance parameter on the cosmological parameters provides a reparameterization $N'_j = N_j - \\sum_l f_{lj}(C_l)$ in which the nuisance parameters are approximately independent of the cosmological parameters in the likelihood. Once this holds, the joint posterior factorizes as a product of a cosmological and a nuisance part, so any independent prior on the rotated nuisance parameters drops out of the marginal posterior for the cosmological parameters. The paper argues that the reduction of the prior projection effect is obtained without discarding information: the transformation merely separates the part of the nuisance variation explained by cosmology from the part that is not. In the EFTofLSS application, both a linear and a GAM-based transformation reduce the biases on $\\ln 10^{10} A_s$ and $w_0$ relative to the un-transformed fit, and iterating the transformation a few times brings the residual biases below $1\\sigma$.","pith_inferences":["The same residualization idea could be applied to any Bayesian analysis with a nuisance-dominated likelihood, e.g. weak-lensing shear calibration or galaxy-cluster mass-observable relations, whenever a preliminary posterior sample is available.","The method's reliance on sampling the prior-influenced posterior region implies that if the true likelihood support lies far outside the preliminary sample volume, the GAM fit will not capture the dependence there; a two-stage design with wider initial priors would test this.","The paper's separability result is framed as approximate for non-additive couplings; a natural stronger test is to check whether the variance of the marginal cosmological posterior, averaged over many simulated data realizations, changes when the nuisance prior is moved by several sigma.","The linear-model variant performing nearly as well as the GAM on these simulations suggests that, for the current EFTofLSS likelihood, most of the cosmological-nuisance coupling is captured by the leading linear term; non-additive couplings may matter more at higher $k_{\\max}$ or for higher-order statistics such as the bispectrum."],"forward_implications":["Applying the transformation to EFT-based full-shape analyses of BOSS, eBOSS, or DESI data would make the reported cosmological constraints insensitive to the widths, locations, and shapes of the priors placed on galaxy bias, counterterm, and stochastic parameters.","The method removes the need to choose physically motivated nuisance priors (e.g. HOD-informed priors) as a protection against projection effects, while still allowing informative priors on the cosmological parameters.","Because the rotated nuisance parameters are standardized to unit variance, the method supplies a natural default prior choice for nuisance parameters in future survey analyses.","The iterative variant of the method reduces the remaining non-linear residual correlation, so the gain in robustness grows with the number of iterations at the cost of increased MCMC runtime.","A differentiable version of the GAM would allow Hamiltonian Monte Carlo sampling, making the method scalable to the much larger nuisance-parameter space of multi-redshift-bin surveys like DESI."],"supporting_citations":[{"why":"Supplies the penalized B-spline regression machinery and the quadratic smoothness penalty used in the GAM orthogonalization step.","marker":"[41]"},{"why":"Defines the frequentist notion of parameter orthogonality through a diagonal information matrix that the proposed transformation aims to approximate.","marker":"[43]"},{"why":"Introduces the linear posterior-orthogonalization transformation $\\alpha' = \\alpha - \\nu \\beta$ that the paper generalizes to non-linear additive functions.","marker":"[45]"},{"why":"Recommends the family $\\alpha' = \\alpha - \\psi(\\beta)$ of reparameterizations for weakly identified posteriors, which is the direct prior art the GAM transformation extends.","marker":"[47]"},{"why":"The DESI full-shape analysis whose prior-volume and prior-weight projection effects motivate the problem and whose data the method is aimed at.","marker":"[11]"},{"why":"Investigates the sensitivity of EFT-based BOSS power-spectrum analyses to nuisance priors, providing the baseline problem the method addresses.","marker":"[22]"},{"why":"Provides the PyBird EFTofLSS model and notation for the galaxy power spectrum used in the simulation pipeline.","marker":"[49]"},{"why":"Supplies the perturbation-theory covariance matrix used to generate the 100 noisy mock power-spectrum datasets.","marker":"[55]"},{"why":"The Effort.jl differentiable emulator used to compute the EFT power spectra in the MCMC runs.","marker":"[58]"},{"why":"Provides the preconditioned Monte Carlo sampler used for all preliminary and reparameterized posterior fits.","marker":"[60, 61]"}],"fun_headline_variants":["GAM reparametrization removes nuisance-prior bias in EFT-LSS fits","Rotating nuisance params frees cosmology from prior sensitivity","Non-linear rotation kills nuisance-prior effects in LSS cosmology","GAM-based trick strips nuisance priors from EFT large-scale fits","Redefining nuisance axes decouples EFT cosmology from priors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The GAM fitted to the preliminary posterior samples captures all statistically relevant dependence of each nuisance parameter on the cosmological parameters, so that the residuals are truly independent of cosmology within the volume that the data and priors actually probe.","fun_headline_variants_meta":{"raw":{"variants":["GAM reparametrization removes nuisance-prior bias in EFT-LSS fits","Rotating nuisance params frees cosmology from prior sensitivity","Non-linear rotation kills nuisance-prior effects in LSS cosmology","GAM-based trick strips nuisance priors from EFT large-scale fits","Redefining nuisance axes decouples EFT cosmology from priors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000368,"raw_usage":{"total_tokens":1981,"prompt_tokens":960,"completion_tokens":1021,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":927}},"tokens_in":576,"tokens_out":1021,"duration_ms":8020,"temperature":1.0,"reasoning_tokens":927,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:19:39.661720+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline on the 100 mocks with a deliberately shifted or much wider nuisance prior, e.g. a uniform prior of width 100 on every EFT parameter instead of the standard priors in Table 2, and measure the shift in the stacked MAP for $\\ln 10^{10} A_s$ and $w_0$ in the GAM-transformed basis: if the shift remains comparable to the un-transformed bias, the claimed independence has failed.","supporting_citations":[{"cited_title":"Wood, Generalized Additive Models: An Introduction with R , Chapman & Hall/CRC Texts in Statistical Science, Taylor & Francis (2006)","cited_arxiv_id":null,"evidence_quote":"Supplies the penalized B-spline regression machinery and the quadratic smoothness penalty used in the GAM orthogonalization step."},{"cited_title":"Cox and N","cited_arxiv_id":null,"evidence_quote":"Defines the frequentist notion of parameter orthogonality through a diagonal information matrix that the proposed transformation aims to approximate."},{"cited_title":"Albert, Nuisance parameters and the use of exploratory graphical methods in a bayesian analysis, The American Statistician 43 (1989) 191","cited_arxiv_id":null,"evidence_quote":"Introduces the linear posterior-orthogonalization transformation $\\alpha' = \\alpha - \\nu \\beta$ that the paper generalizes to non-linear additive functions."},{"cited_title":"Papaspiliopoulos, G.O","cited_arxiv_id":null,"evidence_quote":"Recommends the family $\\alpha' = \\alpha - \\psi(\\beta)$ of reparameterizations for weakly identified posteriors, which is the direct prior art the GAM transformation extends."},{"cited_title":"D’Amico, L","cited_arxiv_id":null,"evidence_quote":"Provides the PyBird EFTofLSS model and notation for the galaxy power spectrum used in the simulation pipeline."}],"review_version":1}