{"id":"b4765c50-d4cc-4043-a2ff-b0a4d83e46a4","arxiv_id":"2607.18926","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A new distance-based prior for correlation matrices centers shrinkage on any target reference and can enforce graph-structured sparsity.","lead":"This paper proposes a new Bayesian prior for correlation matrices that shrinks toward any user-chosen target correlation structure, not just the identity. It also lets the prior respect conditional-independence graphs, reducing the number of parameters from one per pair of variables to one per edge.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The prior's 'Fisher arc-length distance' is actually a tangent-space Euclidean norm; exponential decay in true geodesic distance is only local, and the prior is not reparameterization-invariant as claimed.","rationale":"I read the paper in good faith. The construction in (18)-(19) is coherent: ξ is sampled as radial-exponential times uniform on the sphere, and the affine map back to θ gives a proper prior that is centered at R0. Proposition 3.1's propriety and elicitation claims are correct for the density as printed (I did not find the sign/flip error the reader mentioned; (18) has r^{-(m-1)} and integrates to 1). The GCP parametrization's bijectivity and one-parameter-per-edge claim follow from the Cholesky fill-in lemma and the row-scaling argument in Algorithm 1. The load-bearing weakness is the mismatch between the claimed distance and the actual distance used: r is the norm of the displacement in the tangent space under the Fisher metric at θ0, not the geodesic distance. On a curved manifold these differ at second order in the displacement, and the prior's level sets are not geodesic balls. This makes the abstract's 'exponential decay in Fisher arc-length' an overstatement and undermines the claimed reparameterization invariance in Sec 2.4, since different bijections (GCP vs LKJ, or different orderings) give different r-level sets in R-space. The core methodological contribution — a proper, graph-structured prior centered on any R0 — remains valid, so the verdict stays CONDITIONAL and the concern should be addressed by reframing the distance as a tangent-space approximation and removing the 'exact Fisher distance' wording.","tokens_in":21695,"tokens_out":16244,"duration_ms":159184,"concrete_test":"For p=2 (Example 3.1), compute the exact Fisher geodesic distance d_F(ρ0,ρ) = ∫_{ρ0}^{ρ} √((1+t²)/(1−t²)²) dt and compare with r(ρ)=√(1−ρ0⁴)|θ(ρ)−θ(ρ0)| where θ(ρ)=−ρ/√(1−ρ²). Take ρ0=0 and ρ=0.95. If d_F/r differs from 1 by more than 10%, the prior does not decay exponentially in true Fisher arc-length; it decays in the tangent-space norm, so the abstract's primary characterization fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central distance r in (15)-(19) is the Euclidean norm of the whitened parameter ξ = H(θ0)^{1/2}(θ−θ0). Equation (8) identifies this with Fisher arc-length only to second order near θ0. On the curved manifold R_p, level sets {θ: r ≤ c} are ellipsoids in the tangent space at R0, not geodesic balls; the abstract and Prop 3.1 claim 'mass decaying exponentially in the Fisher arc-length distance', and Sec 4.4 states 'each unit of r is one Fisher distance unit' — false for finite displacements. Because r is defined through a specific coordinate chart (GCP, or any bijection in Sec 3), the induced prior on R depends on that chart and the chosen diagonal/ordering, contradicting the reparameterization-invariance claimed in Sec 2.4 for PC priors. Properness, centering at R0, and graph sparsity survive, but the 'distance-based' interpretation is only an approximation and λ does not have the stated Fisher-distance interpretation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a prior over correlation matrices centred at an arbitrary user-specified reference matrix R0. The construction works in an unconstrained parameterisation θ: it defines a whitened coordinate ξ = H(θ0)^{1/2}(θ−θ0), where H(θ0) is the Hessian of the Kullback-Leibler divergence at R0, and assigns an exponential prior to r = ||ξ|| with a uniform direction on the sphere. The claimed motivation is a PC prior decaying exponentially in Fisher arc-length distance from R0. To handle conditional-independence structure, the paper introduces a Graphical Cholesky Parameterisation (GCP) based on the Cholesky factor of the precision matrix, reducing the number of free parameters to the edge count |E|. The paper proves propriety, contraction to R0, bijectivity of the GCP map, provides a direct sampling algorithm, and illustrates the prior on a heart-disease dataset.","tokens_in":21940,"tokens_out":16768,"duration_ms":145221,"significance":"If the construction is taken as a parameterisation-dependent prior with a local Fisher-geometric interpretation, it fills a genuine gap: LKJ-type priors only shrink toward the identity and do not incorporate conditional-independence sparsity. The GCP is an elegant and potentially useful parameterisation, and the sampling algorithm is a practical contribution. The paper is also commendably explicit about the compatibility constraint between R0 and G and provides a public implementation. However, the central 'Fisher arc-length distance' claim is overstated: the distance used is a tangent-space Euclidean norm, equal to Fisher arc-length only to second order. Relatedly, the prior is not reparameterisation-invariant, despite statements in the text. These issues are fixable by revision, but they affect how the contribution should be framed.","major_comments":[{"comment":"The abstract and §4.4 claim that the prior 'assigns mass decaying exponentially in the Fisher arc-length distance' and that 'each unit of r is one Fisher distance unit'. However, r in Eqs. (15)–(19) is the Euclidean norm of ξ = H(θ0)^{1/2}(θ−θ0). Eq. (8) identifies this with Fisher arc-length only to second order near θ0. On the curved manifold R_p, the level sets are tangent-space ellipsoids, not geodesic balls; for finite displacements the exponential decay is in the whitened tangent norm, not the true Fisher arc-length. Consequently, the elicitation formula in Prop. 3.1(iii) is a tail probability for this whitened norm, not for a Fisher distance. Please revise the abstract and Prop. 3.1 wording, and either define the prior explicitly as based on the whitened norm (with Fisher interpretation as local) or derive a prior based on the exact KLD with an appropriate level-set measure.","section":"Abstract; §3 Eq. (16); §4.4"},{"comment":"The construction (18)–(19) is not invariant to the choice of parameterisation. The Hessian H(θ0) and the uniform measure on the level sets {ξ : ||ξ||=r} are coordinate-dependent. Thus the statement in §2.4 that the resulting prior is 'invariant to reparameterisation' and the claim in §3 that 'the choice is immaterial' for any smooth bijection of the form (12) are not correct for the actual prior defined. Similarly, the forward map (25) is invariant to the diagonal values d_i only in the sense that the same R can be represented by different θ; the induced prior on R depends on d_i. Since d_i is an arbitrary user choice (default p−i+1), this source of parameterisation dependence should be acknowledged, or the prior should be redefined using a coordinate-free geometric measure (e.g., Riemannian volume on geodesic spheres).","section":"§2.4; §3; §4 Remark (Choice of diagonal)"}],"minor_comments":[{"comment":"The density (18) is correct: the radial Jacobian r^{m-1} appears in the denominator, consistent with the Appendix's expression c_m λ e^{-λ r} r^{-(m-1)}. The sentence 'the hyperspherical Jacobian contributes r^{m-1}' could be misread as a multiplication; suggest 'after dividing by the hyperspherical Jacobian r^{m-1}'.","section":"§3, Eq. (18)"},{"comment":"For m ≥ 2 the prior density (18) is unbounded at θ0. This is acknowledged as harmless for prior draws, but if the log-density is evaluated on the θ-parameterisation in Stan, an interior infinite log-density can cause HMC numerical issues. Please describe how the graphpcor implementation handles this (e.g., non-centred parameterisation using r and u, or a regularized density) and whether any safeguards are used.","section":"§4.3 / §5"},{"comment":"The remark 'The forward map (25) is invariant to the choice of positive diagonal values' is only a statement about representability of the same R; the prior itself changes with d_i. This should be stated explicitly, perhaps as a limitation or as motivation for the suggested default d_i = p−i+1.","section":"§4 Remark (Choice of diagonal)"},{"comment":"The sentence 'In the dense case the choice is immaterial, however, and the derivation below applies equally to the LKJ parameterisation, or to any other smooth bijection' is misleading in light of the parameterisation dependence of the prior. Please qualify it.","section":"§3, paragraph after Eq. (12)"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about a sign error in Eq. (18) does not land: the printed density is correct and integrates to one. The substantive issues are the overstated 'Fisher arc-length' interpretation and the unqualified reparameterisation-invariance claim; both are fixable by rewriting the relevant passages and clarifying the local nature of the geometric interpretation. The graphpcor implementation and the reproducibility of the heart-data example would merit a quick check at resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper gives applied Bayesians something they've been missing: a proper prior on a correlation matrix that shrinks to any user-specified R0 (not just the identity) and can be constrained to respect a conditional-independence graph. That's a genuine step beyond LKJ and beyond the authors' own 2025 latent-variable framework, which was stuck at the identity and on fixed sign orthants. The GCP parameterization is the real contribution: one free parameter per graph edge, fill-in handled analytically, no constraints on the parameters, and a direct sampler in Algorithm 2 that makes prior predictive checks routine. The South Africa heart-disease example shows the machinery working in a realistic setting, and the code is available. Good.\n\nWhere it's soft: the abstract and several sections say the prior assigns mass decaying exponentially in the Fisher arc-length distance from R0. That's only true to second order near R0. The actual r is the Euclidean norm of a whitened tangent-space coordinate; on the curved manifold of correlation matrices, the level sets are ellipsoids in the tangent space, not geodesic balls, and \"each unit of r is one Fisher distance unit\" (Section 4.4) is false for finite displacements. Relatedly, the claim of reparameterization invariance in Section 2.4 doesn't hold for the construction as written—the prior depends on the chosen coordinate chart through H(θ0). This doesn't invalidate the prior; it just means the distance interpretation is an approximation. The authors should say so explicitly and drop or soften the invariance claim.\n\nOne correction to the reader's report: the apparent sign error in equation (18) is a misreading. The printed denominator r^(m-1) is exactly r^{-(m-1)}, giving the standard joint density of an Exponential radius and a uniform direction. The formula is correct, and the appendix proof confirms propriety. No problem there.\n\nMinor: the default diagonal d_i = p-i+1 is justified heuristically via a boundary Wishart argument, but it's still a convention; a short sensitivity note would help.\n\nAudience: applied Bayesians working with latent Gaussian models, graphical models, or any setting where informative correlation priors matter. It deserves serious refereeing. I'd send it out with a request to recalibrate the language: present r as a practical tangent-space proxy for Fisher distance, not the distance itself, and remove the invariance claim. With that revision, it's a solid, publishable contribution.","headline":"Useful graph-structured correlation priors with arbitrary reference R0, but the 'Fisher arc-length' framing overstates what is a tangent-space approximation; the density formula flagged by the reader is actually correct.","tokens_in":22455,"tokens_out":3670,"would_cite":true,"duration_ms":32391,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62H20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper establishes a proper, tunable prior on correlation matrices that concentrates on any specified target matrix and reduces the number of free parameters to one per edge of a conditional-independence graph.","keywords":["correlation matrix","Penalised Complexity prior","Fisher information distance","Gaussian graphical model","Cholesky factorisation","prior elicitation","Bayesian inference","conditional independence"],"falsifier":"Compute the exact Fisher geodesic distance from a reference R0 to a distant point R on the manifold by numerically integrating the geodesic equations, and compare it with the whitened norm r = ||H(theta0)^(1/2)(theta - theta0)||; if the ratio between the two varies substantially with distance or direction, the prior's purported exponential decay in Fisher arc-length is only approximate and can deviate arbitrarily far from the stated geometry.","tokens_in":1426,"feed_emoji":"🎯","tokens_out":1648,"duration_ms":57967,"temperature":0.7,"pith_summary":"Bayesian analysts often need a prior on a correlation matrix that encodes a substantive belief, such as that certain variables are strongly related, rather than the default shrinkage toward the identity matrix. This paper proposes a distance-based prior that assigns mass decaying exponentially in the Fisher arc-length distance from any user-specified reference correlation matrix, built as a Penalised Complexity prior. The key innovation is a parameterisation that constructs the correlation matrix from the Cholesky factor of its inverse subject to a user-supplied graph, so the parameter count drops from one per pair of variables to one per edge, while the prior remains proper for all rate parameters and allows both positive and negative correlations. If the construction holds, it gives practitioners a generic way to combine an arbitrary shrinkage target with conditional-independence structure, something not available from the LKJ prior or earlier latent-variable PC priors.","feed_headline":"New prior centers correlation matrices on any target","feed_subtitle":"A single rate parameter shrinks toward your chosen reference, with one free parameter per edge of a conditional-independence graph.","key_machinery":"The central object is the whitened parameter xi = H(theta0)^(1/2)(theta - theta0), where H(theta0) is the Hessian of the Kullback-Leibler divergence (the Fisher information) evaluated at the reference matrix. The prior is a multivariate Laplace on xi: the radius r = ||xi|| follows an exponential distribution and the direction is uniform on the unit sphere, giving a Cartesian density proportional to lambda exp(-lambda r) / r^(m-1). This is pushed forward to the correlation manifold through a smooth bijection theta maps to R(theta). The Graphical Cholesky Parameterisation constructs R by Cholesky-factorising the inverse correlation matrix Q = LL^T: the diagonal is fixed, edge entries are free,","core_discovery":"The paper claims that the prior defined by equation (19), a multivariate Laplace on the whitened parameter xi = H(theta0)^(1/2)(theta - theta0) with H(theta0) the Fisher information at the reference, is a proper prior on correlation matrices that concentrates on any user-specified reference R0. Combined with the Graphical Cholesky Parameterisation, which factors the inverse correlation matrix and leaves one free parameter per edge of a user-supplied graph, it is supported exactly on the submanifold of correlation matrices whose precision respects the graph. The prior is proper for every positive rate lambda, accommodates correlations of either sign, reduces to the dense prior when the graph","pith_inferences":["The local quadratic approximation linking the whitened norm to true Fisher arc-length may fail far from the reference on the curved manifold; one testable extension is to compute exact geodesic distances numerically and compare them with the whitened norm to assess how well the exponential decay holds globally.","The GCP's dependence on vertex labelling, since fill-in count varies with ordering, suggests that graph reordering heuristics could be used as a tuning step to improve the accuracy of the whitening approximation, an idea the paper notes but does not turn into an algorithm.","The prior's density vanishes as correlations approach plus or minus 1 for all lambda, unlike the polynomial boundary decay of LKJ; this could affect sparse-data posteriors that seek to allow near-boundary correlations, and a direct posterior comparison would be informative."],"forward_implications":["Users can specify an informative prior centred on any substantively motivated correlation matrix, not just the identity, with a single rate parameter controlling shrinkage strength in interpretable distance units.","The graph-based parameterisation imposes conditional-independence structure with one parameter per edge instead of p(p-1)/2, making problems up to roughly p = 150 tractable in standard MCMC.","The prior is proper for every positive lambda, so posterior inference is well-defined for any shrinkage rate.","Exact sampling is available via a simple algorithm, exponential radius, uniform direction, affine whitening, enabling prior predictive checks and sensitivity analysis before data are observed.","When the graph is complete the prior reduces to the unstructured dense case, and when R0 = I it provides a shrinkage-toward-independence default comparable to the LKJ prior but with a dimension-stable rate."],"fun_headline_variants":["Distance-based prior centers on any correlation target","Prior shrinks correlations toward a chosen reference matrix","Graph-aware prior for correlation matrices","Proper prior for correlations with any sign and graph","Sampling algorithm for graph-structured correlation priors"],"cache_read_input_tokens":23808,"weakest_assumption_plain":"The identification of the Euclidean norm in whitened coordinates with Fisher arc-length relies on a local quadratic approximation to the Kullback-Leibler divergence, which may break down far from the reference matrix on the curved correlation manifold.","fun_headline_variants_meta":{"raw":{"variants":["Distance-based prior centers on any correlation target","Prior shrinks correlations toward a chosen reference matrix","Graph-aware prior for correlation matrices","Proper prior for correlations with any sign and graph","Sampling algorithm for graph-structured correlation priors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000471,"raw_usage":{"total_tokens":2179,"prompt_tokens":746,"completion_tokens":1433,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":1380}},"tokens_in":490,"tokens_out":1433,"duration_ms":9838,"temperature":1.0,"reasoning_tokens":1380,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T13:55:26.019837+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the exact Fisher geodesic distance from a reference R0 to a distant point R on the manifold by numerically integrating the geodesic equations, and compare it with the whitened norm r = ||H(theta0)^(1/2)(theta - theta0)||; if the ratio between the two varies substantially with distance or direction, the prior's purported exponential decay in Fisher arc-length is only approximate and can deviate arbitrarily far from the stated geometry.","supporting_citations":[],"review_version":1}