WorksMathematics
R2-GP: Variance-Calibrated Priors for Additive Gaussian Processes
Abstract
We develop R2-GP, an R2D2-type prior for Gaussian process (GP) generalised additive models (GAMs) that calibrates explained variance () for Matérn covariance kernels. The function variance for standard Matérn kernels depends jointly on the effect magnitude and length-scale. As such, smooth and rough components with identical amplitudes contribute very different amounts of variation. With independent priors on component amplitudes, this variation results in variance overestimation, especially in higher dimensions. We overcome this by taking direct control over each component's expected contribution to prior explained variance. We do this by rescaling each component based on its length scale and observed inputs, and modelling each rescaled variance as a share of the function variance induced by a global prior. This allows R2-GP to scale well to many components. To also scale to large sample sizes, we extend R2-GP to Hilbert space approximation methods, a framework for kernels exploiting their spectral decompositions for fast and efficient inference. Our simulations and real-world case studies demonstrate superior generalisation and parameter recovery over independent variance priors.
Introduction
Gaussian processes (GPs) are a class of probabilistic methods for modelling non-linear functions. They do so by assuming a multivariate Gaussian prior distribution over a finite set of these functions and are specified through a mean function and a covariance kernel (Williams and Rasmussen, [27]). Common choices of covariance kernel involve the Matérn class of functions (Rasmussen and Williams, [19]), defined using a length scale parameter and a magnitude parameter (also called GP marginal variance). The former controls the correlation decay rate between two arbitrary inputs while the latter controls prior variance of the functions at a single input.
GP priors have been specifically used in generalised additive models (GAMs; Hastie and Tibshirani, [8, 9]) as an approach to model smooth functions (Wood, [28]). In GAMs, a response is modelled using multiple covariates such that
for , where is the intercept, is the residual error and is a GP specific to the -th covariate. Additive GPs arise as a special case of the general class of models presented in Duvenaud et al. [4].
In GAMs, modelling a response with many covariates introduces a variance explosion problem familiar from high-dimensional linear regression. If each of the components receives an independent variance prior with positive mass on non-zero values, the total signal variance accumulates as grows. Consequently, the implied prior on the proportion of explained variance pathologically concentrates near one. This may lead to overfitting and contradict the modeller’s domain knowledge about , rendering the model predictively inconsistent [20].
Preventing this variance accumulation requires explicitly controlling the model’s total variance. In regression contexts, a natural approach is to place a prior on and partition it across the components, analogous to the R2D2 prior for linear models [31]. The mapping between a component’s magnitude parameter and the observed function variance one wishes to control is however blurred by the simultaneous impact of the length scale and the input design. Because of this entanglement, smooth and rough components with identical magnitudes contribute vastly different amounts to observed function variation (Figure 1).

Figure 1. GP Realisations and Normaliser . Realisations of a zero-mean GP with an SE kernel and unit marginal variance on . The shaded band is marginal standard deviation. Panels (a) to (c) hold the marginal variance fixed and vary the length scale, and report the sample standard deviation averaged over 4000 realisations. As grows, the signal variance falls even though the marginal variance is unchanged. Panel (d) shows the resulting normaliser of (5).
We overcome this structural issue with our R2-GP prior. By rescaling each component with a normaliser which is derived from its length scale and the input design, R2-GP forces the component’s magnitude to exactly match its expected variation across the data. This calibration allows a global prior which is partitioned with a simplex prior that forces components to compete for a share of the variance budget.
A drawback of this normalisation is the computational burden it adds to posterior inference. Since the normaliser depends on the length scale, it must be recomputed at every gradient evaluation when using MCMC methods like Hamiltonian Monte Carlo [16, 10]. For an exact GP, this requires , added to the standard cost of factorising the exact covariance matrix.
We solve this scalability issue by extending the R2-GP prior to Hilbert space approximation methods (HSGPs) [24, 21, 15]. HSGPs exploit the spectral decomposition of the kernel using basis functions for large sample scalability and have been extended to GAMs efficiently [25]. With the extension to HSGPs based on basis functions, computing the normaliser simplifies to a single pre-computation of and .
Code to replicate the experiments is available at https://github.com/davkoh/r2-gp.
Contributions.
We develop R2-GP by extending the R2D2 family of priors to GPs for GAMs. We derive the necessary normalising approach of the kernel such that a prior over the GP’s maps analytically to its expected sample variance. A Dirichlet decomposition of into contributions of its GAM components allows its application to high dimensions.
We derive R2-GP prior for the Hilbert space approximation and show that the normaliser admits an exact finite-sample expression requiring preparation and and scales to high dimensions.
Through simulations and real-world data, we show that the R2-GP prior outperforms independent priors for additive HSGPs in prediction and parameter recovery and sampling efficiency, and is competitive with alternate shrinkage priors.
Related Work. Priors placed on , since the original Zhang et al. [31], have been extended to multilevel models [1], generalised linear mixed models [30], autoregressions [11], spatial models [29] and grouped predictors ([2]. The present paper takes the variance allocation mechanism from that literature and addresses the question it leaves on how to extend this to GPs.
Closest to our normalisation approach is the standardisation procedure of Ferrari and Ventrucci [5], who scale variance partitioning priors in latent Gaussian models by the variance of a zero-mean-constrained effect over an assumed covariate distribution, evaluated at fixed values of unknown kernel parameters such as the range of a Matérn process. Our normaliser is instead a function of the length scale and is recomputed at every evaluation inside the sampler, which is what makes the identity of Lemma A.1 particularly efficient. Penalised complexity priors for Gaussian random fields ([22, 6]) shrink towards a simpler base model, whereas the normalisation here fixes the units in which the scale parameter is expressed, and the two could in principle be combined.
The remainder of the paper is organised as follows. § 2 derives the normalised additive prior and its properties, among them the implied prior on explained variance. § 3 presents our simulation studies and § 4 demonstrates a real-world application. Proofs and additional results are collected in Appendix.
Methods
In our framework, the -th column of the design matrix holds the observed values of covariate . We use GAM GP prior with independent processes per covariate,
where is a stationary kernel of the Matérn family, which includes the squared exponential (SE) kernel as a limiting case, with , so that has unit marginal variance. Parameter is the magnitude of the component, which equals its pointwise marginal variance, at every input.1 In contrast, the signal variance is the expected variation that the function exhibits across the inputs, . We denote by and the expectation and variance operators with respect to . In the following, we first derive the mathematical disconnect between these two variances, then provide the necessary concepts of the Hilbert space (HS) approximation to finally define an efficient finite-sample normaliser, which we ultimately use to construct the R2-GP prior.
Calibrating GPs for Explained Variance
In linear regression, placing a prior directly on the proportion of explained variance provides an interpretable way to specify regularisation on variance parameters of regression coefficients. For GAM GPs the identity is,
where a prior maps to prior , conditional on . We define Equation (3) as the marginal , since it’s marginal of the effect of the design and any other kernel parameters. The magnitude , however is not equal to the signal variance. To see this, consider the total variance of component
,
The first term is the expected variance of the function across the observed inputs, while the second term captures the variance of the average height of function realisations (See Appendix § 8.§ 1).
The second term is precisely the point of departure between a correlated GP and a linear regression. If were a linear term with mean-centred inputs, the level, , is deterministically zero, such that marginal variance equals signal variance. In contrast, the average level of a realised GP acts as a random constant shift across all input observations. Because this shift is constant with respect to , it does not contribute to the function’s variation over the inputs (the signal variance), and the level itself is completely absorbed by the intercept in (1).
The size of the level variance depends on the length scale and on the input design. Denoting by the average kernel correlation between two independent inputs and , the level variance reduces to , and the signal variance is therefore
For the Matérn family, whose correlations are positive and increase with the length scale, (unless all inputs coincide). As Figure 1 illustrates, smoother realisations (larger ) carry a larger part of their variance in the level, so that their signal variance falls well below the magnitude .
Hence, we define the realised as the modelling target:
which, whenever , implies . Furthermore, because depends on , the implied prior on explained variance depends on the length scales, and so does the share of each component in the total signal variance.
This motivates replacing (2) with a normalised prior,
Given , this is a deterministic rescaling that leaves Gaussian and its correlation structure unchanged. By (5), it makes the signal variance of equal to for every length scale, while the marginal variance becomes . Consequently, and agree, and, as in the linear case, we can place a prior on a global and partition it across components with a Dirichlet distribution,
where and parameterise the Beta prior through its mean and precision [1]. We refer to (7) together with (8) as the R2-GP prior. Its total signal variance is , and the Dirichlet weight is the share of component , for any length scale value. Since the proportions in a Dirichlet distribution have strictly negative pairwise correlation, the GAM components intrinsically compete for their share of the total explained variance.
The Hilbert Space Representation
The Hilbert space (HS) approximation methods is based on the eigenvalue decomposition of the Gram matrix for a chosen kernel using its spectral density along with eigenvalues and eigenfunctions based on the GP inputs. Following Solin and Särkkä [24], we use a compact cover containing the observed values of covariate . Based on the Dirichlet boundary conditions, we get closed form solutions for eigenvalues and eigenfunctions of the kernel such that
where is the number of basis functions. The matrix contains the eigenfunction values at the observed inputs, such that , with for its -th column. We obtain the spectral density of our chosen kernel evaluated at the -th frequency, collected in . For example, the spectral density of the SE kernel is of the form obtained by solving the Fourier integral with respect to the inputs. The spectral forms of other Matérn kernels are provided in Riutort-Mayol et al. [21].
Lemma 2.1 (Finite-design covariance). Let and . Then, conditionally on and ,
and is symmetric positive semi-definite with rank .
The matrix is the finite-basis approximation to the exact kernel matrix with . Under a sufficiently large boundary and increasing it converges to . We use this convergence as established and refer to Solin and Särkkä [24] for the underlying expansion. For practical guidance on choosing and , we follow Riutort-Mayol et al. [21]. Since , the diagonal of is approximately one, which is the marginal normalisation that § 2.1 identified as insufficient for modelling the signal variance of the function.
The Finite-sample Normaliser
In HS form, we define the normaliser of component as the expected sample variance of the unit-magnitude component of Lemma 2.1, evaluated at the observed inputs,
where is the sample variance of a vector and is the expectation over the coefficients . This is the finite-sample counterpart of (5), with the sample variance over the observed inputs in place of the variance operator and the HSGP covariance in place of the exact kernel. Evaluating (11) through the matrix would cost operations each time changes. However, because the coefficients in are independent with unit variance, the cross terms between different basis functions have expectation zero, and the normaliser simplifies to the sum of the sample variances of the basis functions, weighted by the spectral density (Lemma A.1 in Appendix § 8.§ 2),
This identity allows the R2-GP prior to scale well with . The sample variances of the basis columns depend on the design alone and not on any parameter, so they are precomputed once at cost . Inside the sampler, the normaliser then reduces to an inner product of length between the spectral density and these precomputed variances, which requires only work each time the length scale changes.2
Expanding the sample variance in (11) gives the HSGP counterpart of the decomposition (4),
where is the sample mean of the realisation . The first term aggregates the marginal variances, each approximately one, and the second is the prior variance of the realisation’s level, which the intercept absorbs. Together they equal the average marginal variance minus the average covariance between two distinct observed inputs, the sample analogue of in (5).
The Normalised Additive Prior in HS Form
We construct the normalised HSGPs, using from (8) as follows
where is the elementwise product. The square root and division are elementwise. Together with (8), (14) is the HSGP form of the R2-GP prior.
As for the exact GP in § 2.1, the signal variance of each component is then its allocated variance,
and, for independent , the signal variances add up to the total budget,
Three comments are in order that highlight the usefulness of the R2 prior for GPs. Firstly, by writing with , the proposed prior is equivalent to independent centred priors on the basis weights
Computationally, the remains standard normal and the scale is applied through a deterministic transformation, and thus, equivalent to the non-centred parameterisation ([17]) which is well known to keep the posterior geometry well conditioned when a signal variance is small.
Second, because the deterministic rescaling in (7) applies a single global scalar to all basis functions of component equally, it preserves the relative weighting of the spectral densities that govern the smoothness of the function. Furthermore, because this scalar multiplier is independent of the basis index , it does not interact with the truncation of the series. Consequently, the normalised HSGP inherits the same convergence guarantees as the standard HSGP: as the number of basis functions grows, the approximate prior converges to the exact normalised GP prior.
Third, the prior on the total signal variance does not change as components are added. By (16) and (8), it equals for every , so that its prior is the one placed on . If instead each has a prior second moment that is free of , as under independent priors, it has prior mean . Figure 2 shows the consequence for prior draws for the empirical proportion of explained variance, under different priors. Under the R2-GP prior with and (panel R2-GP), its implied prior stays close to the target for every between 1 and 100. Under independent half-normal priors with unit scale, the median rises from at to at . As competing shrinkage priors, which are further introduced in § 3, the regularised horseshoe (Piironen and Vehtari [18]) and normal-gamma (Griffin and Brown [7]) may shrink each but leave the prior on their total to shrink or grow with .

Figure 2. Implied Prior on . Empirical CDF of for additive HSGP GAMs under four priors on , as the number of components grows. CDFs are generated from 2000 prior draws: , , and . These figures are insensitive to the prior . RHS and normal-gamma are specified as in § 3, with . The dashed curve is the target . Only R2-GP is stable in .
Notice that calibration holds for but not for the empirical proportion of explained variance, which is a nonlinear function of it. The map is strictly concave on , so Jensen’s inequality gives
with strict inequality whenever is non-degenerate. This accounts for the slight shift to the left for small in Figure 2 for the R2-GP prior. This is not a deficiency of the proposed prior, since no choice of calibrates both and for a nonlinear .
Finally, to ensure the component’s prior expected marginal variance remains finite, the prior on the length scale must be sufficiently light-tailed. As , the normaliser for the exact kernel decays such that , where for the exponential kernel and for smoother Matérn variants (Rasmussen and Williams [19]). Therefore, the exact marginal variance has a finite prior expectation if and only if . While the HSGP approximation inherently bounds this variance for finite , preserving the properties of the exact-kernel limit rules out heavy-tailed priors, such as an Inverse-Gamma distribution.
Simulation study
We assess finite-sample predictive and parameter recovery performance of the R2-GP prior, that is, {(8)} together with {(14)}, as the number of GAM components increases. The competitors are priors on the component scales that carry popular shrinkage approaches over to {(1)}. These have not yet been considered for HSGP GAMs to our knowledge, but pose straightforward extensions based off our methodology.
Setup
Data Generation. We generate data from (1) with and . Each is a draw from the normalised prior (14) with a common length scale and signal variance
for . By (16), the total signal variance is for every , while the signal variance of a single component decays at rate . We call (19) the dense data-generating process (DGP). We also consider a sparse DGP in Appendix § 9.§ 2, in which the first five components share the same total signal variance and for . We vary , and the kernel choice that includes SE, Matérn 3/2 and Matérn 5/2. For each condition, we simulate 50 replications. A data set has training and 500 test observations. Inputs are iid uniform on . Following Riutort-Mayol et al. [21], we set and with , 3.42 and 2.65 for the three kernels, so that . The data are generated with basis functions, so that the is generated close as a close representation to the desired GP. Since (16) holds in expectation over , the realised value , that is, at the generating and , varies across data sets around the nominal .
Priors. All models use the normalised components (14) and differ only in the prior on . R2-GP is (8) with , and , so that , as in § 2.4. We compare it with three alternatives, where a superscript denotes truncation to . (i) The normal prior, independently, is the analogue of ridge regression and represents independent priors on the component scales. (ii) The regularised horseshoe (RHS; Piironen and Vehtari, [18]), with , , and , represents sparsity-favouring priors. We set as in Piironen and Vehtari [18], with a prior guess of active components. While is miscalibrated for the dense DGP, is nevertheless learned from the data, and is well calibrated for the sparse DGP. (iii) The normal-gamma prior (NG; Griffin and Brown, [7]), with in shape-rate form, and , represents global-local shrinkage priors that are popular in other domains. The remaining priors are common to all models, namely , and . The comparisons presented here isolate the effect of the prior on . We show in Appendix § 9.§ 3 that the qualitative behaviour of the competing priors is the same also without the normalisation.
Expected Behaviour. The total signal variance is independent of under R2-GP, whereas its mean is under normal, and under NG, respectively. Since the generating value is throughout, expect normal and NG to overestimate the signal variance and to underestimate by a margin that grows with , in line with Figure 2. RHS is a sparsity favouring prior which for has , hence, decays faster than the generating which is expected to lead to excessive shrinkage for large .
Evaluation. Predictive performance is measured by the log predictive density summed over the test observations, which estimates the expected log predictive density (elpd; Vehtari et al., [26]), relative to R2-GP,
for prior and replication , so that negative values favour R2-GP. We report its mean over replications and, following Sivula et al. [23] and McLatchie et al. [13], regard as a meaningful difference. For we report the coverage of the central 90% posterior intervals (Table 2, Table 5 and Table 8) and the relative root-mean-squared error, that is relative to the RMSE of the R2-GP (RMSE; Table 3, Table 6 and Table 9).
| Prior | \multicolumn{5}{c}{} | \multicolumn{5}{c}{} | \multicolumn{5}{c}{} | |||||||||||||
| 5 | 10 | 30 | 50 | 100 | 5 | 10 | 30 | 50 | 100 | 5 | 10 | 30 | 50 | 100 | ||
| 0.3 | R2-GP | 97 | 100 | 100 | 99 | 100 | 100 | 100 | 100 | 100 | 100 | 92 | 82 | 92 | 94 | 98 |
| Normal | 95 | 96 | 95 | 96 | 100 | 100 | 100 | 100 | 100 | 100 | 90 | 80 | 38 | 2 | 0 | |
| NG | 91 | 97 | 80 | 77 | 99 | 100 | 100 | 100 | 100 | 100 | 92 | 80 | 72 | 68 | 2 | |
| RHS (=5) | 84 | 85 | 63 | 47 | 18 | 100 | 100 | 100 | 100 | 100 | 92 | 74 | 64 | 54 | 42 | |
| 0.5 | R2-GP | 99 | 99 | 100 | 100 | 100 | 100 | 100 | 100 | 100 | 100 | 94 | 86 | 88 | 90 | 92 |
| Normal | 95 | 95 | 94 | 94 | 99 | 100 | 100 | 100 | 100 | 100 | 92 | 84 | 60 | 8 | 0 | |
| NG | 94 | 97 | 90 | 64 | 99 | 100 | 100 | 100 | 100 | 100 | 94 | 84 | 72 | 66 | 34 | |
| RHS (=5) | 88 | 88 | 82 | 52 | 33 | 100 | 100 | 100 | 100 | 100 | 94 | 84 | 70 | 58 | 48 | |
| 0.7 | R2-GP | 98 | 99 | 100 | 100 | 99 | 100 | 100 | 100 | 100 | 100 | 92 | 90 | 88 | 88 | 96 |
| Normal | 95 | 95 | 94 | 95 | 96 | 100 | 100 | 100 | 100 | 100 | 90 | 84 | 72 | 24 | 0 | |
| NG | 98 | 98 | 93 | 68 | 88 | 100 | 100 | 100 | 100 | 100 | 92 | 86 | 74 | 72 | 58 | |
| RHS (=5) | 87 | 89 | 85 | 57 | 40 | 100 | 100 | 100 | 100 | 100 | 92 | 86 | 74 | 68 | 54 |
Table 2. Interval Coverage (%), Dense DGP, SE Kernel. Bold marks the coverage closest to 90%.
| Prior | \multicolumn{5}{c}{} | \multicolumn{5}{c}{} | \multicolumn{5}{c}{} | |||||||||||||
| 5 | 10 | 30 | 50 | 100 | 5 | 10 | 30 | 50 | 100 | 5 | 10 | 30 | 50 | 100 | ||
| 0.3 | R2-GP | 0.097 | 0.062 | 0.033 | 0.025 | 0.017 | 0.022 | 0.016 | 0.008 | 0.006 | 0.003 | 0.054 | 0.084 | 0.079 | 0.082 | 0.078 |
| Normal | 1.57 | 2.30 | 4.94 | 7.59 | 11.61 | 1.05 | 1.22 | 1.84 | 2.50 | 3.13 | 0.98 | 0.95 | 2.65 | 7.73 | 9.97 | |
| NG | 1.52 | 1.88 | 2.74 | 2.92 | 3.13 | 0.94 | 0.98 | 1.12 | 1.30 | 3.13 | 1.01 | 1.02 | 1.66 | 2.30 | 9.26 | |
| RHS (=5) | 1.64 | 1.97 | 2.64 | 2.93 | 3.46 | 0.87 | 0.76 | 0.57 | 0.54 | 0.58 | 1.11 | 1.30 | 1.72 | 1.90 | 2.22 | |
| 0.5 | R2-GP | 0.092 | 0.061 | 0.030 | 0.029 | 0.018 | 0.016 | 0.012 | 0.006 | 0.004 | 0.003 | 0.053 | 0.065 | 0.077 | 0.078 | 0.097 |
| Normal | 1.96 | 2.85 | 6.65 | 8.05 | 13.26 | 1.01 | 1.07 | 1.60 | 2.50 | 3.65 | 0.98 | 1.02 | 1.83 | 4.79 | 7.92 | |
| NG | 1.66 | 2.02 | 3.05 | 2.89 | 3.35 | 0.93 | 0.95 | 1.06 | 1.13 | 2.70 | 1.00 | 1.03 | 1.49 | 1.65 | 5.26 | |
| RHS (=5) | 1.75 | 2.02 | 2.61 | 2.51 | 3.08 | 0.88 | 0.81 | 0.68 | 0.66 | 0.81 | 1.07 | 1.21 | 1.48 | 1.71 | 1.88 | |
| 0.7 | R2-GP | 0.095 | 0.061 | 0.031 | 0.027 | 0.018 | 0.011 | 0.008 | 0.004 | 0.003 | 0.002 | 0.053 | 0.065 | 0.074 | 0.075 | 0.082 |
| Normal | 2.07 | 3.11 | 7.12 | 8.98 | 15.52 | 0.94 | 1.08 | 1.45 | 1.98 | 6.32 | 0.98 | 1.00 | 1.41 | 3.23 | 9.53 | |
| NG | 1.71 | 2.05 | 3.02 | 3.04 | 3.50 | 0.92 | 0.96 | 1.01 | 1.27 | 1.82 | 1.00 | 1.03 | 1.41 | 1.52 | 3.34 | |
| RHS (=5) | 1.79 | 2.06 | 2.48 | 2.55 | 3.04 | 0.87 | 0.84 | 0.70 | 0.71 | 0.78 | 1.05 | 1.17 | 1.42 | 1.72 | 1.97 |
Table 3. Relative RMSE, Dense DGP, SE Kernel. The R2-GP rows give its RMSE in italics.
| Prior | : 5 | 10 | 30 | 50 | 100 | : 5 | 10 | 30 | 50 | 100 | : 5 | 10 | 30 | 50 | 100 | |
| 0.3 | R2-GP | 98 | 99 | 100 | 100 | 100 | 100 | 100 | 100 | 100 | 100 | 92 | 88 | 96 | 90 | 100 |
| Normal | 94 | 96 | 95 | 98 | 100 | 100 | 100 | 100 | 100 | 100 | 98 | 84 | 2 | 0 | 0 | |
| NG | 92 | 95 | 87 | 91 | 100 | 100 | 100 | 100 | 100 | 100 | 96 | 82 | 84 | 52 | 0 | |
| RHS () | 87 | 78 | 60 | 42 | 16 | 100 | 100 | 100 | 100 | 100 | 82 | 82 | 70 | 50 | 46 | |
| 0.5 | R2-GP | 98 | 99 | 98 | 100 | 98 | 100 | 100 | 100 | 100 | 100 | 88 | 94 | 94 | 88 | 92 |
| Normal | 94 | 94 | 95 | 96 | 100 | 100 | 100 | 100 | 100 | 100 | 86 | 88 | 36 | 0 | 0 | |
| NG | 93 | 95 | 82 | 77 | 99 | 100 | 100 | 100 | 100 | 100 | 86 | 86 | 80 | 66 | 14 | |
| RHS () | 89 | 81 | 63 | 52 | 29 | 100 | 100 | 100 | 100 | 100 | 86 | 80 | 68 | 44 | 54 | |
| 0.7 | R2-GP | 97 | 97 | 98 | 100 | 98 | 100 | 100 | 100 | 100 | 100 | 88 | 92 | 98 | 88 | 94 |
| Normal | 93 | 94 | 95 | 94 | 99 | 100 | 100 | 100 | 100 | 100 | 90 | 84 | 50 | 0 | 0 | |
| NG | 95 | 96 | 84 | 75 | 99 | 100 | 100 | 100 | 100 | 100 | 88 | 86 | 82 | 70 | 30 | |
| RHS () | 87 | 87 | 67 | 61 | 35 | 100 | 100 | 100 | 100 | 100 | 86 | 86 | 74 | 60 | 58 |
Table 5. Interval Coverage (%), Dense DGP, Matérn 3/2 Kernel. As in Table 2.
| Prior | : 5 | 10 | 30 | 50 | 100 | : 5 | 10 | 30 | 50 | 100 | : 5 | 10 | 30 | 50 | 100 | |
| 0.3 | R2-GP | 0.087 | 0.065 | 0.034 | 0.025 | 0.016 | 0.019 | 0.012 | 0.006 | 0.005 | 0.003 | 0.057 | 0.068 | 0.067 | 0.094 | 0.087 |
| Normal | 1.60 | 1.90 | 4.92 | 6.26 | 11.18 | 1.07 | 1.17 | 2.09 | 2.18 | 3.42 | 0.93 | 1.00 | 5.67 | 7.54 | 8.93 | |
| NG | 1.58 | 1.75 | 2.73 | 2.82 | 2.99 | 0.96 | 0.93 | 1.22 | 1.70 | 2.44 | 0.99 | 1.05 | 1.93 | 3.13 | 8.62 | |
| RHS () | 1.70 | 1.95 | 2.72 | 3.01 | 3.65 | 0.87 | 0.71 | 0.55 | 0.57 | 0.74 | 1.16 | 1.39 | 2.03 | 1.78 | 2.13 | |
| 0.5 | R2-GP | 0.096 | 0.066 | 0.035 | 0.024 | 0.016 | 0.012 | 0.008 | 0.004 | 0.003 | 0.002 | 0.065 | 0.068 | 0.064 | 0.089 | 0.098 |
| Normal | 1.77 | 2.19 | 4.85 | 8.62 | 12.48 | 0.97 | 1.03 | 1.50 | 2.60 | 4.45 | 1.02 | 0.94 | 3.40 | 6.66 | 7.85 | |
| NG | 1.60 | 1.84 | 2.61 | 3.35 | 3.38 | 0.93 | 0.91 | 1.05 | 1.39 | 2.46 | 1.02 | 1.01 | 1.76 | 2.09 | 6.61 | |
| RHS () | 1.67 | 1.93 | 2.52 | 3.07 | 3.47 | 0.88 | 0.77 | 0.63 | 0.67 | 0.96 | 1.09 | 1.31 | 1.91 | 1.98 | 1.74 | |
| 0.7 | R2-GP | 0.097 | 0.062 | 0.036 | 0.024 | 0.017 | 0.007 | 0.005 | 0.003 | 0.002 | 0.002 | 0.064 | 0.066 | 0.059 | 0.079 | 0.093 |
| Normal | 1.81 | 2.62 | 5.09 | 9.39 | 13.07 | 0.92 | 1.00 | 1.32 | 2.08 | 6.92 | 0.96 | 0.97 | 2.75 | 5.92 | 8.48 | |
| NG | 1.58 | 1.92 | 2.57 | 3.31 | 3.31 | 0.89 | 0.91 | 1.15 | 1.27 | 2.11 | 0.99 | 0.99 | 1.60 | 1.89 | 5.85 | |
| RHS () | 1.66 | 1.99 | 2.40 | 2.96 | 3.26 | 0.84 | 0.77 | 0.75 | 0.91 | 1.11 | 1.08 | 1.24 | 1.86 | 2.06 | 1.75 |
Table 6. Relative RMSE, Dense DGP, Matérn 3/2 Kernel. As in Table 3.
| Prior | 10 | 30 | 50 | 100 | 10 | 30 | 50 | 100 | 10 | 30 | 50 | 100 | ||||
| 0.3 | R2-GP | 92 | 97 | 100 | 100 | 100 | 100 | 100 | 100 | 100 | 100 | 82 | 88 | 96 | 92 | 98 |
| Normal | 93 | 97 | 95 | 97 | 100 | 100 | 100 | 100 | 100 | 100 | 88 | 86 | 20 | 0 | 0 | |
| NG | 89 | 95 | 85 | 84 | 100 | 100 | 100 | 100 | 100 | 100 | 84 | 86 | 86 | 48 | 2 | |
| RHS () | 81 | 80 | 59 | 49 | 22 | 100 | 100 | 100 | 100 | 100 | 82 | 86 | 68 | 64 | 58 | |
| 0.5 | R2-GP | 98 | 98 | 98 | 99 | 100 | 100 | 100 | 100 | 100 | 100 | 90 | 92 | 96 | 92 | 94 |
| Normal | 94 | 95 | 95 | 94 | 100 | 100 | 100 | 100 | 100 | 100 | 90 | 88 | 48 | 2 | 0 | |
| NG | 94 | 97 | 82 | 76 | 99 | 100 | 100 | 100 | 100 | 100 | 88 | 90 | 80 | 70 | 14 | |
| RHS () | 88 | 80 | 63 | 56 | 27 | 100 | 100 | 100 | 100 | 100 | 86 | 88 | 62 | 60 | 50 | |
| 0.7 | R2-GP | 98 | 98 | 98 | 100 | 100 | 100 | 100 | 100 | 100 | 100 | 88 | 94 | 90 | 94 | 94 |
| Normal | 93 | 95 | 95 | 94 | 99 | 100 | 100 | 100 | 100 | 100 | 88 | 90 | 60 | 8 | 0 | |
| NG | 95 | 97 | 84 | 77 | 95 | 100 | 100 | 100 | 100 | 100 | 88 | 92 | 78 | 76 | 44 | |
| RHS () | 90 | 85 | 66 | 61 | 38 | 100 | 100 | 100 | 100 | 100 | 86 | 88 | 66 | 66 | 54 |
Table 8. Interval Coverage (%), Dense DGP, Matérn 5/2 Kernel. As in Table 2.
| Prior | 10 | 30 | 50 | 100 | 10 | 30 | 50 | 100 | 10 | 30 | 50 | 100 | ||||
| 0.3 | R2-GP | 0.100 | 0.066 | 0.031 | 0.025 | 0.017 | 0.020 | 0.016 | 0.007 | 0.005 | 0.003 | 0.066 | 0.071 | 0.056 | 0.082 | 0.078 |
| Normal | 1.58 | 2.03 | 5.16 | 6.79 | 10.64 | 1.03 | 1.14 | 1.94 | 2.37 | 2.89 | 1.00 | 1.11 | 5.11 | 8.49 | 9.87 | |
| NG | 1.53 | 1.76 | 2.81 | 3.02 | 2.78 | 0.93 | 0.99 | 1.04 | 1.78 | 2.63 | 1.02 | 1.06 | 1.66 | 3.29 | 9.56 | |
| RHS () | 1.62 | 1.89 | 2.91 | 2.96 | 3.36 | 0.87 | 0.81 | 0.54 | 0.57 | 0.80 | 1.09 | 1.15 | 2.55 | 1.88 | 1.86 | |
| 0.5 | R2-GP | 0.099 | 0.067 | 0.034 | 0.024 | 0.018 | 0.013 | 0.009 | 0.004 | 0.003 | 0.002 | 0.054 | 0.063 | 0.070 | 0.086 | 0.087 |
| Normal | 1.81 | 2.20 | 5.18 | 8.96 | 11.94 | 0.95 | 1.09 | 1.58 | 2.44 | 4.71 | 1.02 | 1.03 | 2.31 | 6.07 | 8.97 | |
| NG | 1.59 | 1.80 | 2.67 | 3.37 | 3.25 | 0.92 | 0.93 | 0.99 | 1.33 | 3.00 | 1.02 | 1.02 | 1.58 | 1.90 | 7.14 | |
| RHS () | 1.67 | 1.89 | 2.55 | 3.01 | 3.19 | 0.85 | 0.78 | 0.64 | 0.67 | 0.85 | 1.04 | 1.13 | 1.82 | 1.73 | 1.75 | |
| 0.7 | R2-GP | 0.101 | 0.064 | 0.036 | 0.024 | 0.017 | 0.009 | 0.006 | 0.003 | 0.002 | 0.002 | 0.053 | 0.063 | 0.072 | 0.079 | 0.087 |
| Normal | 1.90 | 2.52 | 5.49 | 9.80 | 13.47 | 0.91 | 0.99 | 1.45 | 1.87 | 5.34 | 1.02 | 1.00 | 1.67 | 4.58 | 8.83 | |
| NG | 1.62 | 1.89 | 2.62 | 3.42 | 3.38 | 0.89 | 0.90 | 1.07 | 1.36 | 2.01 | 1.01 | 1.03 | 1.40 | 1.75 | 4.77 | |
| RHS () | 1.71 | 1.92 | 2.42 | 2.91 | 3.15 | 0.86 | 0.79 | 0.71 | 0.73 | 0.95 | 1.02 | 1.10 | 1.63 | 1.79 | 1.81 |
Table 9. Relative RMSE, Dense DGP, Matérn 5/2 Kernel. As in Table 3.
Values above one indicate worse recovery than under R2-GP. Lastly, we compare the posterior of with . All models are implemented in Stan (Carpenter et al. [3]) and fitted with four chains of 2000 iterations, with the half as warm-up at a target acceptance rate of 0.95. Figure 3 shows the predictive performance and the recovery of for the SE kernel with . We show figures for all other conditions in Appendix § 9. See Table 4, Table 7 and Table 10 for sampling diagnostics.

Figure 3. Dense DGP, SE Kernel, . Left: mean over 50 simulations (solid) with Monte Carlo standard error (dotted). Dashed lines mark and negative values favour R2-GP. Right: posterior means (points) and central 90% intervals (bars) of for each simulation. Crosses mark the realised .
| Prior | \multicolumn{5}{c}{} | \multicolumn{5}{c}{ESS} | \multicolumn{5}{c}{Divergences} | \multicolumn{5}{c}{Time (s)} | |||||||||||||||||
| 5 | 10 | 30 | 50 | 100 | 5 | 10 | 30 | 50 | 100 | 5 | 10 | 30 | 50 | 100 | 5 | 10 | 30 | 50 | 100 | ||
| 0.3 | R2-GP | 1.00 | 1.00 | 1.01 | 1.01 | 1.01 | 1078 | 998 | 814 | 661 | 498 | 0.0 | 0.0 | 0.2 | 0.6 | 1.9 | 5.7 | 9.1 | 26 | 51 | 129 |
| Normal | 1.00 | 1.01 | 1.01 | 1.26 | 1.37 | 951 | 868 | 480 | 43 | 18 | 1.2 | 1.4 | 5.8 | 283 | 235 | 7.7 | 16 | 76 | 458 | 1634 | |
| NG | 1.02 | 1.04 | 1.17 | 1.40 | 1.12 | 569 | 222 | 39 | 25 | 62 | 14 | 10 | 116 | 116 | 78 | 12 | 32 | 204 | 350 | 903 | |
| RHS (=5) | 1.00 | 1.01 | 1.01 | 1.01 | 1.01 | 896 | 770 | 551 | 578 | 708 | 2.3 | 3.6 | 7.2 | 15 | 12 | 8.4 | 14 | 36 | 63 | 134 | |
| 0.5 | R2-GP | 1.00 | 1.00 | 1.01 | 1.01 | 1.01 | 1176 | 1148 | 1019 | 946 | 607 | 0.0 | 0.0 | 0.0 | 0.3 | 1.4 | 7.2 | 12 | 32 | 56 | 191 |
| Normal | 1.00 | 1.00 | 1.01 | 1.05 | 1.37 | 1043 | 973 | 756 | 217 | 19 | 2.0 | 1.6 | 1.1 | 80 | 233 | 9.3 | 20 | 97 | 349 | 1292 | |
| NG | 1.01 | 1.03 | 1.15 | 1.25 | 1.43 | 605 | 328 | 51 | 24 | 32 | 8.9 | 18 | 69 | 199 | 49 | 13 | 31 | 175 | 395 | 961 | |
| RHS (=5) | 1.00 | 1.01 | 1.01 | 1.01 | 1.01 | 1035 | 911 | 578 | 624 | 568 | 4.1 | 4.6 | 17 | 16 | 32 | 11 | 16 | 45 | 69 | 171 | |
| 0.7 | R2-GP | 1.00 | 1.00 | 1.00 | 1.01 | 1.01 | 1289 | 1266 | 1099 | 999 | 684 | 0.7 | 0.4 | 0.0 | 0.0 | 1.7 | 8.1 | 14 | 38 | 65 | 196 |
| Normal | 1.00 | 1.00 | 1.01 | 1.01 | 1.48 | 1111 | 1078 | 919 | 531 | 14 | 2.9 | 1.9 | 0.8 | 8.2 | 247 | 13 | 26 | 128 | 324 | 1457 | |
| NG | 1.01 | 1.03 | 1.15 | 1.23 | 1.61 | 609 | 334 | 54 | 22 | 20 | 28 | 34 | 89 | 193 | 58 | 17 | 41 | 205 | 439 | 1125 | |
| RHS (=5) | 1.00 | 1.01 | 1.01 | 1.01 | 1.01 | 1093 | 930 | 632 | 636 | 587 | 6.6 | 10 | 12 | 18 | 21 | 13 | 21 | 57 | 82 | 199 |
Table 4. Sampling Diagnostics, Dense DGP, SE Kernel. Time is the sampling time in seconds.
| Prior | : 5 | 10 | 30 | 50 | 100 | ESS: 5 | 10 | 30 | 50 | 100 | Divergences: 5 | 10 | 30 | 50 | 100 | Time (s): 5 | 10 | 30 | 50 | 100 | |
| 0.3 | R2-GP | 1.00 | 1.00 | 1.01 | 1.01 | 1.01 | 1053 | 990 | 716 | 581 | 475 | 0.0 | 0.0 | 0.1 | 2.4 | 6.1 | 7.7 | 13 | 38 | 72 | 215 |
| Normal | 1.00 | 1.01 | 1.07 | 1.25 | 1.37 | 925 | 807 | 171 | 23 | 16 | 0.6 | 0.4 | 114 | 243 | 247 | 9.1 | 20 | 177 | 738 | 2401 | |
| NG | 1.02 | 1.05 | 1.19 | 1.39 | 1.08 | 512 | 197 | 37 | 28 | 58 | 2.3 | 12 | 69 | 66 | 53 | 14 | 49 | 272 | 450 | 1678 | |
| RHS () | 1.01 | 1.01 | 1.01 | 1.01 | 1.02 | 861 | 748 | 541 | 579 | 678 | 1.2 | 2.3 | 6.9 | 18 | 28 | 11 | 19 | 54 | 99 | 273 | |
| 0.5 | R2-GP | 1.00 | 1.00 | 1.01 | 1.01 | 1.01 | 1179 | 1120 | 966 | 775 | 540 | 0.0 | 0.1 | 0.1 | 0.4 | 7.2 | 8.7 | 15 | 42 | 73 | 265 |
| Normal | 1.00 | 1.01 | 1.01 | 1.20 | 1.39 | 1016 | 988 | 545 | 42 | 17 | 1.7 | 0.7 | 6.5 | 239 | 273 | 12 | 25 | 138 | 690 | 2543 | |
| NG | 1.03 | 1.04 | 1.18 | 1.33 | 1.21 | 650 | 276 | 38 | 23 | 50 | 6.1 | 28 | 110 | 118 | 73 | 16 | 47 | 294 | 483 | 1760 | |
| RHS () | 1.00 | 1.01 | 1.01 | 1.01 | 1.02 | 940 | 833 | 614 | 559 | 634 | 2.2 | 8.4 | 8.1 | 17 | 40 | 14 | 22 | 59 | 112 | 317 | |
| 0.7 | R2-GP | 1.00 | 1.00 | 1.01 | 1.01 | 1.01 | 1228 | 1179 | 1099 | 856 | 610 | 1.0 | 0.0 | 0.1 | 0.4 | 2.0 | 11 | 17 | 49 | 90 | 282 |
| Normal | 1.00 | 1.00 | 1.01 | 1.09 | 1.45 | 1057 | 1026 | 754 | 112 | 16 | 2.1 | 1.0 | 2.3 | 130 | 238 | 14 | 30 | 148 | 574 | 2364 | |
| NG | 1.01 | 1.03 | 1.15 | 1.28 | 1.37 | 679 | 317 | 41 | 21 | 31 | 18 | 24 | 122 | 118 | 59 | 20 | 49 | 294 | 568 | 1924 | |
| RHS () | 1.00 | 1.01 | 1.01 | 1.02 | 1.02 | 1018 | 921 | 643 | 530 | 561 | 4.2 | 5.5 | 14 | 33 | 37 | 16 | 27 | 72 | 127 | 349 |
Table 7. Sampling Diagnostics, Dense DGP, Matérn 3/2 Kernel. As in Table 4.
| Prior | 5 | 10 | 30 | 50 | 100 | ESS 5 | 10 | 30 | 50 | 100 | Divergences 5 | 10 | 30 | 50 | 100 | Time (s) 5 | 10 | 30 | 50 | 100 | |
| 0.3 | R2-GP | 1.00 | 1.00 | 1.01 | 1.01 | 1.01 | 1053 | 984 | 765 | 635 | 448 | 0.0 | 0.0 | 0.2 | 0.4 | 4.6 | 7.2 | 12 | 33 | 59 | 178 |
| Normal | 1.00 | 1.01 | 1.02 | 1.28 | 1.34 | 932 | 839 | 320 | 30 | 17 | 0.8 | 0.6 | 38 | 233 | 263 | 9.2 | 18 | 135 | 662 | 2012 | |
| NG | 1.01 | 1.03 | 1.21 | 1.32 | 1.10 | 547 | 265 | 27 | 29 | 58 | 15 | 15 | 104 | 102 | 65 | 16 | 40 | 275 | 397 | 1400 | |
| RHS () | 1.01 | 1.01 | 1.01 | 1.01 | 1.03 | 860 | 736 | 584 | 558 | 575 | 2.1 | 2.7 | 7.3 | 12 | 47 | 11 | 18 | 49 | 84 | 242 | |
| 0.5 | R2-GP | 1.00 | 1.00 | 1.01 | 1.01 | 1.01 | 1169 | 1158 | 996 | 909 | 618 | 0.0 | 0.0 | 0.1 | 0.7 | 1.6 | 9.0 | 14 | 39 | 69 | 207 |
| Normal | 1.00 | 1.01 | 1.01 | 1.14 | 1.38 | 990 | 975 | 677 | 84 | 17 | 1.8 | 2.0 | 1.9 | 136 | 212 | 12 | 23 | 128 | 529 | 2034 | |
| NG | 1.01 | 1.03 | 1.17 | 1.28 | 1.22 | 644 | 266 | 38 | 23 | 48 | 7.2 | 28 | 106 | 104 | 84 | 16 | 44 | 283 | 525 | 1315 | |
| RHS () | 1.00 | 1.01 | 1.01 | 1.01 | 1.01 | 971 | 936 | 654 | 583 | 657 | 2.3 | 4.0 | 15 | 22 | 29 | 13 | 21 | 57 | 99 | 238 | |
| 0.7 | R2-GP | 1.00 | 1.00 | 1.01 | 1.01 | 1.01 | 1241 | 1231 | 1114 | 902 | 659 | 0.2 | 0.1 | 0.1 | 0.6 | 2.5 | 9.9 | 17 | 48 | 85 | 239 |
| Normal | 1.00 | 1.00 | 1.01 | 1.03 | 1.36 | 1102 | 1069 | 862 | 279 | 15 | 2.2 | 1.4 | 2.1 | 53 | 228 | 15 | 31 | 150 | 500 | 2106 | |
| NG | 1.01 | 1.03 | 1.15 | 1.25 | 1.50 | 666 | 330 | 37 | 19 | 29 | 12 | 35 | 119 | 122 | 69 | 18 | 51 | 281 | 542 | 1488 | |
| RHS () | 1.01 | 1.01 | 1.01 | 1.01 | 1.01 | 1010 | 964 | 717 | 565 | 593 | 10 | 3.9 | 14 | 16 | 28 | 15 | 24 | 68 | 128 | 310 |
Table 10. Sampling Diagnostics, Dense DGP, Matérn 5/2 Kernel. As in Table 4.
Results
Predictive Performance. R2-GP attains the highest mean elpd in each of the 45 conditions, against every competitor. At the priors are close, however normal and NG deteriorate as grows. This is as a consequence of the lack of joint shrinkage of the GAM components, which leads to overestimation of . RHS stays closest to R2-GP at higher , with mean for all . As expected, with RHS overshrinks as increases, which can be seen as underestimation of in Figure 3. Due to the small true variance shares as increases, this results in less of a penalty in predictive performance than the overestimation of seen for normal and NG.
Recovery. Recovery mirrors the predictive results for the dense DGP. R2-GP has the lowest or close to the lowest RMSE for in all conditions. Interval coverage is close to or above the nominal level, with its advantage widening with . (Table 2 and Table 3 for the SE kernel, Appendix § 9 for the Matérn kernels).
Sampling Efficiency. R2-GP also improves the sampling efficiency markedly. On average over replications, its largest is at most 1.01 and its smallest bulk effective sample size is at least 448 in every condition. It has the fewest divergent transitions in all conditions and the shortest sampling time in all but one. At , by contrast, the normal prior and NG fail to converge. The results of the normal and NG priors at large therefore reflect the prior and the sampler jointly. The RHS prior converges, with , but needs up to 1.6 times the sampling time.
Sparse Allocation. Under the sparse DGP (Appendix § 9.§ 2), RHS with the calibration attains a higher mean elpd than R2-GP for , however by small margins. The normal prior underperforms as in the dense DGP, and NG does so at . Relative to RHS, R2-GP’s RMSE is slightly lower for the non-zero in all conditions, however higher for the true zero . This is as expected, since for the allocation is uniform on the simplex, whereas concentration near zero requires a small ([31, 1, 11]).
Summary. For dense signals R2-GP performs at least as well as every competing prior in terms of prediction and recovers component scales most accurately. In particular, our method recovers the component scales most accurately in all conditions, with a margin over the normal and NG priors that grows with . Sampling wise, our prior is the most reliable prior to sample. The patterns are almost identical across the three kernels and the three length scales.
Real-world Case Study
Data. In addition to our simulation scenarios, we demonstrate our proposed methods in analysing the BloodBrain data [14] from the R package caret [12]. The dataset describes 208 chemical compounds using 134 molecular descriptors based on the compounds structure. The response is the logarithm of the ratio of the compound’s concentration in the brain to the blood, a standard procedure to measure how readily a chemical compound crosses the blood-brain barrier (with as response). For simplicity, we consider only continuous covariates. The data has samples and covariates.
Model Fit. We estimate the same set of models as in § 3, except for the RHS prior for which . We consider the SE and Matérn 3/2 kernels. We set boundary conditions with and compute the number of basis functions accordingly. To compare predictive performance, the data is split into 80/20 train/test ratio with 10 random replications. Each fit uses 4 chains with 1000 warm-up and 1000 sample draws.
Results. The covariate clogp describing lipophilicity is a biologically relevant covariate for modelling [14] in the BloodBrain data. Additionally, we also investigate here lumo describing the energy component. The effects of these two covariates on response for all fitted models are shown in Figure 4. Across both kernels and all models, increases with an increase in clogp, indicating lipophilic compounds cross the blood-brain-barrier readily. For lumo, we see a decreasing effect on . The shrinkage priors agree while maintaining tight posterior credible intervals while the Normal model remains unstable. See Appendix § 10 for the effect of more covariates. Our proposed R2-GP prior provides superior predictive performance as well as sampling speed and diagnostics (Appendix Table 20) compared to the Normal model. RHS and NG provide comparable predictive performance R2-GPs, but worse MCMC convergence, the latter also reflected in the simulation study. Overall, using R2-GP is computationally and inferentially more robust for this data while providing the benefit of a prior based off of first principles.

Figure 4. Covariate Effects, BloodBrain Data. Top 2 covariates are shown based on their contribution to explained variance.
Discussion
We introduce the R2-GP prior for GAM GPs under which the magnitude of a component equals its signal variance, the expected variation across the inputs. We model magnitudes based off of an interpretable prior on the model’s proportion of explained variance and a Dirichlet partition of that variance across components. Thus, components compete for explained variance, preventing variance explosion with many components. In HS form, the normaliser costs per gradient evaluation, so the approach also scales well for large sample sizes. Our simulations and real-data case study demonstrate superior predictive performance of R2-GP over independent normal priors. Although competitive in predictive performance with other shrinkage priors, R2-GP outperforms in terms of sampling efficiency, especially when many components are active. Future works should include richer kernel structures, such as interactions between covariates and as well as applications in other GP extensions.
AI Use Statement
In this work, we used generative AI tools for code prototyping of parts of the code (Claude Opus 5 High from Anthropic). Further, we used the ChatGPT 5.6 model from OpenAI for minor language edits, aiming to enhance readability. LLM-generated code and text was verified, modified and tested for correctness by the authors. We take responsibility for the final content of this work, including text, claims, or artifacts produced with the aid of generative AI.
References
Overview. This supplement has four parts. Appendix A gives the derivations and proofs for § 2. Appendix B gives the full results of the simulation study of § 3. Appendix C gives further results for the case study of § 4. Appendix D describes the computing infrastructure. Table 1 lists the content of each part and the section of the main text that it supports.
| Appendix | Content | Figures | Tables | Section |
| A.1 | Variance decomposition (4) and normaliser (5) for the exact GP | 2.1 | ||
| A.2 | Proofs of Lemmas 2.1 and A.1, of the decomposition (13) and of the calibration results (15) and (16) | 2.2–2.4 | ||
| B.1 | Dense DGP: and posterior of (figures), interval coverage, relative RMSE and sampling diagnostics (tables), for all kernels and length scales | 5–10 | 2–10 | 3.2 |
| B.2 | Sparse DGP: the same figures and tables | 11–16 | 11–19 | 3.2 |
| B.3 | Competing priors without the normalisation: for both DGPs | 17–22 | 3.1 | |
| C | BloodBrain data: effects of the 12 covariates with the largest share of explained variance, predictive performance and sampling diagnostics | 23–24 | 20 | 4 |
| D | Hardware, software versions and run times | 3, 4 |
Table 1. Overview of the Supplementary Material
Proofs
Derivations for Section 2.1
Throughout, and are the expectation and variance operators with respect to random variable , and and are independent inputs.
Decomposition (4). Apply the law of total variance to under the joint distribution of and . Conditioning on gives
since under (2) the prior mean is zero and at every input. Conditioning on instead gives , and the two expressions together are (4).
Linear Term. The same two steps hold for any zero-mean prior on a function of , with on the left-hand side. For with centred inputs, , the level is for every . Its variance is therefore zero, and the marginal variance, averaged over the inputs, equals the signal variance, .
Variance of the Level and . Under (2), the level has mean zero, and , and are independent, so that
which is and, with (4), gives (5). A Matérn kernel has the form , where is positive and strictly decreasing with . Hence for , and increases with . Unless all inputs coincide, it follows that increases with , so that decreases with . For a general stationary kernel, is still a variance that is at most , so that without any assumption on the sign of the correlations.
Finite-sample Counterpart. Section 2.3 replaces by the sample variance over the observed inputs . The proof of (26) holds for any covariance matrix. For the exact kernel matrix in place of , with , it gives
under (2), so that the expectation over two independent inputs in becomes the average over pairs of distinct observed inputs.
Proofs for Sections 2.2 to 2.4
Throughout, denotes expectation over the coefficient vectors only, holding , and fixed. Two elementary identities are used repeatedly and are stated here once. Let be the centring matrix, which is symmetric and idempotent with and . Subtracting the sample mean is an orthogonal projection, so for any
and for a random vector with mean zero and covariance and a fixed matrix , writing the scalar as and using linearity of the trace gives
Proof of Lemma 2.1. Write with the fixed matrix . A linear image of a Gaussian vector is Gaussian, with and
using , valid since the matrices are diagonal and . Symmetry is immediate, and gives positive semi-definiteness. The rank bound follows from . □
Lemma A.1 (Equivalent expressions for the normaliser). Let and be as in Lemma 2.1. Then the normaliser (11) satisfies
Moreover, whenever and the observed inputs of covariate lie in and are not all equal. Proof. By (22), . Conditionally on and the matrix is fixed and has mean zero with covariance by Lemma 2.1, so (23) gives (24).
For (12), substitute and use cyclicity together with the fact that for any square :
Dividing by and applying (22) to each column turns into . The off-diagonal entries of , which are the sample covariances between different basis columns, do not enter, because the coefficients in are uncorrelated.
For (26), expand inside the trace and use .
For strict positivity, and , so that (12) is zero only if every basis column is constant across the observed inputs. Let , which lies in by assumption, so that by (9). The first column is constant only if every equals or for some . If both values occur and , the second column takes the two distinct values and is not constant. Otherwise all inputs are equal. The condition cannot be dropped: for , a design with two distinct input values that are symmetric about the centre of the interval gives .
Proof of the decomposition (13). Write with . By Lemma 2.1, and have mean zero, with and . Taking the expectation term by term gives (13), in agreement with (26). Splitting into its diagonal and off-diagonal sums gives the equivalent form
the average marginal variance minus the average covariance between two distinct observed inputs. When the diagonal of is one, if and only if this average covariance is positive. Without a condition on the sign of the covariances, positive semi-definiteness of gives only , which is for a unit diagonal.
Proof of the component calibration (15). Conditionally on , and , the factor is a deterministic scalar, and is homogeneous of degree two, since . Hence
by the definition (11). Strict positivity of holds under the conditions of Lemma A.1.
Proof of the additive calibration (16). Let , which is bilinear and symmetric with . Bilinearity gives the exact decomposition
For , cyclicity of the trace as in (23) gives
since given , and the vectors and are functions of the independent and alone and each has mean zero, so . Applying (15) to each diagonal term gives , which equals because lies in the simplex.
Additional Simulation Results
This appendix gives the full results of the simulation study of § 3. The settings are those of § 3.1. In the figures and tables, in a panel title or in the first column of a table is the generating length scale , and RHS without a qualifier and RHS () are the calibration with .
Tables. For each DGP and kernel, three tables give the coverage of the central 90% posterior intervals, the RMSE of the posterior mean relative to R2-GP, and the sampling diagnostics. All entries are means over 50 replications, with one column for each . For and , coverage and RMSE are taken across the components. A relative RMSE above one indicates worse recovery than under R2-GP. For each fit, is the largest value and ESS is the smallest bulk effective sample size over all , , and the log density, and divergences are counted among the 4000 post-warm-up draws. Bold marks the best value in each column for each .
Dense DGP
Figures 5 to 7 show the mean and Figures 8 to 10 the posterior of , for the three kernels and the three length scales. Figure 3 is the condition with the SE kernel and . Tables 2 to 10 give the parameter recovery and the sampling diagnostics. Averaged over replications, the posterior mean of minus lies in under R2-GP in all conditions. At , it lies in under the normal prior, in under NG and in under RHS (). The recovery of is governed by its informative prior: its coverage is 100% after rounding and its RMSE is below 0.03 under all priors.

Figure 6. Predictive Performance, Dense DGP, Matérn 3/2 Kernel. As in Figure 5.

Figure 5. Predictive Performance, Dense DGP, SE Kernel. Mean over 50 replications (solid) with Monte Carlo standard error (dotted), for each generating length scale (panels). Dashed lines mark and negative values favour R2-GP.

Figure 7. Predictive Performance, Dense DGP, Matérn 5/2 Kernel. As in Figure 5.
Figure 9. Explained Variance, Dense DGP, Matérn 3/2 Kernel. As in Figure 8.
Figure 8. Explained Variance, Dense DGP, SE Kernel. Posterior means (points) and central 90% intervals (bars) of for each replication, for each generating length scale (rows). Crosses mark .
Figure 10. Explained Variance, Dense DGP, Matérn 5/2 Kernel. As in Figure 8.
Sparse DGP
In the sparse DGP only the first five components are active, with for and otherwise, so that the total signal variance equals that of the dense DGP (19). All other settings are those of § 3.1. The priors are R2-GP, normal, NG and RHS with , which equals the number of active components for . Figure 11 to Figure 16 and Table 11 to Table 19 give the results. The tables report the RMSE of separately for the active and the inactive components. For , the intervals of R2-GP contain the generating value of an active in at most 27% of cases, and those of RHS in at most 49%. This agrees with the uniform allocation for that § 3.2 discusses. The recovery of and of under R2-GP stays accurate.
Figure 11. Predictive Performance, Sparse DGP, SE Kernel. As in Figure 5.
Figure 16. Explained Variance, Sparse DGP, Matérn Kernel. As in Figure 8.
| Prior | 5 | 10 | 30 | 50 | 100 | 5 | 10 | 30 | 50 | 100 | 5 | 10 | 30 | 50 | 100 | ||||
| 0.3 | R2-GP | 97 | 77 | 23 | 11 | 1 | 100 | 100 | 100 | 100 | 100 | 92 | 82 | 84 | 94 | 94 | |||
| Normal | 95 | 96 | 95 | 95 | 100 | 100 | 100 | 100 | 100 | 100 | 90 | 78 | 30 | 2 | 0 | ||||
| NG | 92 | 85 | 64 | 67 | 60 | 100 | 100 | 100 | 100 | 100 | 92 | 80 | 80 | 80 | 8 | ||||
| RHS () | 85 | 66 | 44 | 41 | 23 | 100 | 100 | 100 | 100 | 100 | 92 | 78 | 80 | 90 | 84 | ||||
| 0.5 | R2-GP | 99 | 84 | 21 | 9 | 0 | 100 | 100 | 100 | 100 | 100 | 94 | 84 | 92 | 90 | 90 | |||
| Normal | 95 | 95 | 97 | 96 | 98 | 100 | 100 | 100 | 100 | 100 | 92 | 78 | 52 | 8 | 0 | ||||
| NG | 94 | 90 | 66 | 63 | 64 | 100 | 100 | 100 | 100 | 100 | 94 | 80 | 88 | 80 | 46 | ||||
| RHS () | 88 | 75 | 47 | 40 | 25 | 100 | 100 | 100 | 100 | 100 | 94 | 84 | 94 | 82 | 94 | ||||
| 0.7 | R2-GP | 98 | 82 | 18 | 6 | 1 | 100 | 100 | 100 | 100 | 100 | 92 | 88 | 82 | 90 | 90 | |||
| Normal | 94 | 95 | 97 | 94 | 99 | 100 | 100 | 100 | 100 | 100 | 92 | 82 | 68 | 18 | 0 | ||||
| NG | 98 | 91 | 66 | 56 | 55 | 100 | 100 | 100 | 100 | 100 | 92 | 82 | 86 | 90 | 76 | ||||
| RHS () | 88 | 76 | 45 | 38 | 26 | 100 | 100 | 100 | 100 | 100 | 92 | 84 | 88 | 86 | 90 |
Table 11. Interval Coverage (%), Sparse DGP, SE Kernel. As in Table 2, for the active components. The interval of a positive parameter cannot contain the zero of an inactive component.
| Prior | 5 | 10 | 30 | 50 | 100 | ESS 5 | ESS 10 | ESS 30 | ESS 50 | ESS 100 | Divergences 5 | Divergences 10 | Divergences 30 | Divergences 50 | Divergences 100 | Time (s) 5 | Time (s) 10 | Time (s) 30 | Time (s) 50 | Time (s) 100 | |
| 0.3 | R2-GP | 1.00 | 1.00 | 1.01 | 1.01 | 1.02 | 1033 | 1032 | 745 | 582 | 386 | 0.0 | 0.0 | 0.1 | 2.1 | 20 | 7.6 | 12 | 37 | 75 | 234 |
| Normal | 1.00 | 1.01 | 1.02 | 1.22 | 1.34 | 930 | 867 | 368 | 30 | 18 | 1.0 | 0.5 | 32 | 228 | 231 | 7.9 | 20 | 130 | 768 | 2202 | |
| NG | 1.01 | 1.04 | 1.14 | 1.31 | 1.13 | 536 | 239 | 42 | 23 | 46 | 10 | 13 | 65 | 81 | 43 | 20 | 47 | 304 | 484 | 1628 | |
| RHS () | 1.01 | 1.01 | 1.01 | 1.02 | 1.02 | 879 | 768 | 482 | 432 | 376 | 2.2 | 1.8 | 9.7 | 14 | 26 | 11 | 18 | 48 | 95 | 295 | |
| 0.5 | R2-GP | 1.00 | 1.00 | 1.01 | 1.01 | 1.01 | 1163 | 1171 | 1036 | 812 | 570 | 0.2 | 0.0 | 0.1 | 0.3 | 12 | 8.9 | 14 | 41 | 81 | 236 |
| Normal | 1.00 | 1.01 | 1.01 | 1.16 | 1.38 | 1004 | 1026 | 706 | 77 | 17 | 1.8 | 1.3 | 2.7 | 129 | 189 | 13 | 25 | 131 | 606 | 2350 | |
| NG | 1.01 | 1.04 | 1.14 | 1.22 | 1.39 | 623 | 233 | 41 | 20 | 25 | 6.7 | 43 | 137 | 134 | 74 | 17 | 53 | 364 | 590 | 1465 | |
| RHS () | 1.00 | 1.01 | 1.01 | 1.01 | 1.02 | 977 | 912 | 695 | 548 | 425 | 2.6 | 5.5 | 7.9 | 15 | 31 | 13 | 19 | 56 | 110 | 323 | |
| 0.7 | R2-GP | 1.00 | 1.00 | 1.01 | 1.01 | 1.01 | 1234 | 1262 | 1168 | 1006 | 701 | 0.2 | 0.0 | 0.1 | 0.1 | 1.1 | 11 | 18 | 51 | 95 | 311 |
| Normal | 1.00 | 1.01 | 1.01 | 1.05 | 1.56 | 1069 | 1083 | 898 | 298 | 12 | 5.0 | 2.1 | 1.4 | 48 | 185 | 15 | 30 | 161 | 573 | 2116 | |
| NG | 1.01 | 1.05 | 1.11 | 1.19 | 1.65 | 653 | 270 | 44 | 25 | 14 | 13 | 51 | 164 | 146 | 56 | 19 | 67 | 380 | 716 | 1724 | |
| RHS () | 1.00 | 1.01 | 1.01 | 1.01 | 1.02 | 1010 | 1018 | 827 | 666 | 496 | 7.0 | 7.9 | 9.2 | 14 | 30 | 15 | 25 | 64 | 131 | 374 |
Table 19. Sampling Diagnostics, Sparse DGP, Matérn 5/2 Kernel. As in Table Table 4.
Figure 12. Predictive Performance, Sparse DGP, Matérn 3/2 Kernel. As in Figure 5.
Figure 13. Predictive Performance, Sparse DGP, Matérn 5/2 Kernel. As in Figure 5.
Figure 14. Explained Variance, Sparse DGP, SE Kernel. As in Figure 8.
Figure 15. Explained Variance, Sparse DGP, Matérn Kernel. As in Figure 8.
| Prior | 5 | 10 | 30 | 50 | 100 | ESS | 5 | 10 | 30 | 50 | 100 | Divergences | 5 | 10 | 30 | 50 | 100 | Time (s) | 5 | 10 | 30 | 50 | 100 | ||
| 0.3 | R2-GP | 1.00 | 1.00 | 1.01 | 1.01 | 1.01 | 1040 | 1003 | 843 | 645 | 463 | 0.0 | 0.0 | 0.0 | 0.5 | 11 | 6.2 | 9.1 | 26 | 49 | 124 | ||||
| Normal | 1.00 | 1.01 | 1.01 | 1.18 | 1.29 | 946 | 901 | 503 | 51 | 20 | 1.4 | 1.7 | 3.9 | 251 | 253 | 8.0 | 15 | 75 | 437 | 1377 | |||||
| NG | 1.01 | 1.03 | 1.15 | 1.32 | 1.24 | 524 | 243 | 34 | 17 | 43 | 5.7 | 28 | 95 | 110 | 81 | 12 | 38 | 212 | 356 | 904 | |||||
| RHS () | 1.00 | 1.01 | 1.01 | 1.02 | 1.02 | 904 | 832 | 627 | 417 | 487 | 3.0 | 3.4 | 5.0 | 12 | 21 | 8.6 | 13 | 35 | 65 | 156 | |||||
| 0.5 | R2-GP | 1.00 | 1.00 | 1.00 | 1.01 | 1.01 | 1191 | 1186 | 1087 | 960 | 648 | 0.0 | 0.0 | 0.0 | 0.1 | 1.9 | 7.3 | 11 | 32 | 53 | 137 | ||||
| Normal | 1.00 | 1.00 | 1.01 | 1.02 | 1.51 | 1054 | 1047 | 836 | 250 | 12 | 2.4 | 1.5 | 2.3 | 70 | 222 | 10 | 19 | 87 | 303 | 1331 | |||||
| NG | 1.03 | 1.02 | 1.11 | 1.20 | 1.59 | 594 | 293 | 42 | 23 | 19 | 11 | 20 | 121 | 146 | 92 | 13 | 33 | 224 | 435 | 867 | |||||
| RHS () | 1.00 | 1.01 | 1.01 | 1.01 | 1.03 | 1010 | 925 | 689 | 625 | 547 | 6.5 | 4.2 | 11 | 16 | 40 | 10 | 16 | 40 | 72 | 169 | |||||
| 0.7 | R2-GP | 1.00 | 1.00 | 1.01 | 1.01 | 1.01 | 1290 | 1266 | 1204 | 1115 | 811 | 0.7 | 0.1 | 0.1 | 0.1 | 1.4 | 8.3 | 13 | 37 | 64 | 167 | ||||
| Normal | 1.00 | 1.00 | 1.01 | 1.01 | 1.50 | 1094 | 1108 | 949 | 562 | 11 | 4.0 | 1.7 | 2.7 | 9.2 | 181 | 12 | 26 | 123 | 308 | 1303 | |||||
| NG | 1.01 | 1.03 | 1.11 | 1.22 | 1.59 | 674 | 292 | 46 | 22 | 9 | 25 | 53 | 161 | 175 | 89 | 16 | 44 | 260 | 467 | 1047 | |||||
| RHS () | 1.00 | 1.01 | 1.01 | 1.01 | 1.01 | 1067 | 1027 | 828 | 763 | 612 | 9.1 | 5.8 | 16 | 11 | 21 | 12 | 18 | 50 | 85 | 199 |
Table 13. Sampling Diagnostics, Sparse DGP, SE Kernel. As in Table 4.
| Prior | \multicolumn{5}{c}{} | \multicolumn{5}{c}{} | \multicolumn{5}{c}{} | |||||||||||||
| 5 | 10 | 30 | 50 | 100 | 5 | 10 | 30 | 50 | 100 | 5 | 10 | 30 | 50 | 100 | ||
| 0.3 | R2-GP | 98 | 80 | 27 | 11 | 0 | 100 | 100 | 100 | 100 | 100 | 94 | 88 | 98 | 88 | 94 |
| Normal | 95 | 95 | 94 | 97 | 100 | 100 | 100 | 100 | 100 | 100 | 98 | 86 | 2 | 0 | 0 | |
| NG | 95 | 84 | 70 | 69 | 64 | 100 | 100 | 100 | 100 | 100 | 98 | 86 | 86 | 64 | 0 | |
| RHS () | 88 | 68 | 49 | 33 | 24 | 100 | 100 | 100 | 100 | 100 | 82 | 86 | 90 | 66 | 80 | |
| 0.5 | R2-GP | 97 | 80 | 18 | 6 | 0 | 100 | 100 | 100 | 100 | 100 | 88 | 90 | 88 | 92 | 90 |
| Normal | 93 | 93 | 96 | 98 | 100 | 100 | 100 | 100 | 100 | 100 | 84 | 86 | 26 | 0 | 0 | |
| NG | 94 | 85 | 60 | 59 | 62 | 100 | 100 | 100 | 100 | 100 | 84 | 86 | 92 | 78 | 18 | |
| RHS () | 88 | 70 | 38 | 33 | 30 | 100 | 100 | 100 | 100 | 100 | 86 | 88 | 94 | 82 | 82 | |
| 0.7 | R2-GP | 98 | 80 | 19 | 7 | 0 | 100 | 100 | 100 | 100 | 100 | 88 | 86 | 96 | 96 | 86 |
| Normal | 94 | 93 | 96 | 96 | 100 | 100 | 100 | 100 | 100 | 100 | 88 | 82 | 40 | 0 | 0 | |
| NG | 94 | 88 | 59 | 58 | 64 | 100 | 100 | 100 | 100 | 100 | 88 | 84 | 96 | 90 | 22 | |
| RHS () | 87 | 71 | 40 | 34 | 32 | 100 | 100 | 100 | 100 | 100 | 84 | 84 | 96 | 80 | 82 |
Table 14. Interval Coverage (%), Sparse DGP, Matérn 3/2 Kernel. As in Table 11.
| Prior | \multicolumn{5}{c}{, active} | \multicolumn{5}{c}{, inactive} | \multicolumn{5}{c}{} | \multicolumn{5}{c}{} | |||||||||||||||||
| 5 | 10 | 30 | 50 | 100 | 5 | 10 | 30 | 50 | 100 | 5 | 10 | 30 | 50 | 100 | 5 | 10 | 30 | 50 | 100 | ||
| 0.3 | R2-GP | 0.087 | 0.121 | 0.176 | 0.207 | 0.244 | — | 0.163 | 0.109 | 0.085 | 0.060 | 0.019 | 0.016 | 0.015 | 0.012 | 0.010 | 0.056 | 0.066 | 0.079 | 0.106 | 0.103 |
| Normal | 1.61 | 1.14 | 0.93 | 0.66 | 0.27 | — | 1.35 | 2.21 | 2.72 | 4.15 | 1.07 | 1.15 | 1.36 | 1.38 | 1.15 | 0.94 | 1.03 | 4.79 | 6.91 | 7.46 | |
| NG | 1.56 | 1.34 | 1.11 | 0.94 | 0.70 | — | 0.77 | 0.72 | 0.93 | 1.62 | 0.97 | 0.98 | 1.09 | 1.19 | 1.68 | 0.97 | 1.05 | 1.49 | 2.45 | 7.20 | |
| RHS () | 1.70 | 1.51 | 1.24 | 1.17 | 1.10 | — | 0.55 | 0.48 | 0.43 | 0.36 | 0.87 | 0.83 | 0.82 | 0.72 | 0.83 | 1.17 | 1.20 | 1.08 | 1.21 | 1.36 | |
| 0.5 | R2-GP | 0.096 | 0.116 | 0.187 | 0.213 | 0.235 | — | 0.163 | 0.100 | 0.081 | 0.064 | 0.012 | 0.011 | 0.009 | 0.007 | 0.008 | 0.065 | 0.067 | 0.066 | 0.079 | 0.123 |
| Normal | 1.78 | 1.43 | 0.79 | 0.70 | 0.38 | — | 1.36 | 2.45 | 3.35 | 4.13 | 0.98 | 0.94 | 1.16 | 1.28 | 1.25 | 1.02 | 1.00 | 3.54 | 7.64 | 6.29 | |
| NG | 1.60 | 1.44 | 1.14 | 1.06 | 0.73 | — | 0.75 | 0.63 | 0.63 | 1.54 | 0.94 | 0.90 | 0.93 | 1.05 | 1.40 | 1.02 | 0.99 | 1.30 | 1.48 | 5.17 | |
| RHS () | 1.68 | 1.55 | 1.20 | 1.13 | 1.07 | — | 0.55 | 0.49 | 0.43 | 0.42 | 0.89 | 0.80 | 0.78 | 0.79 | 0.89 | 1.08 | 1.13 | 1.10 | 1.42 | 1.02 | |
| 0.7 | R2-GP | 0.097 | 0.117 | 0.187 | 0.213 | 0.235 | — | 0.162 | 0.099 | 0.079 | 0.063 | 0.007 | 0.007 | 0.006 | 0.005 | 0.005 | 0.063 | 0.066 | 0.054 | 0.068 | 0.114 |
| Normal | 1.82 | 1.53 | 0.86 | 0.81 | 0.48 | — | 1.45 | 2.57 | 3.67 | 4.49 | 0.90 | 0.95 | 1.12 | 1.28 | 1.79 | 0.97 | 0.99 | 3.30 | 6.83 | 6.83 | |
| NG | 1.59 | 1.45 | 1.15 | 1.11 | 0.74 | — | 0.78 | 0.60 | 0.55 | 1.49 | 0.87 | 0.91 | 1.00 | 0.98 | 1.47 | 1.00 | 0.97 | 1.13 | 1.24 | 4.51 | |
| RHS () | 1.66 | 1.53 | 1.18 | 1.11 | 1.06 | — | 0.55 | 0.49 | 0.44 | 0.43 | 0.84 | 0.82 | 0.83 | 0.83 | 0.88 | 1.09 | 1.11 | 1.00 | 1.37 | 0.96 |
Table 15. Relative RMSE, Sparse DGP, Matérn 3/2 Kernel. As in Table 12.
| Prior | , active | 5 | 10 | 30 | 50 | 100 | , inactive | 5 | 10 | 30 | 50 | 100 | 5 | 10 | 30 | 50 | 100 | 5 | 10 | 30 | 50 | 100 | |||
| 0.3 | R2-GP | 0.097 | 0.118 | 0.179 | 0.203 | 0.242 | — | 0.158 | 0.103 | 0.086 | 0.060 | 0.022 | 0.022 | 0.017 | 0.016 | 0.013 | 0.053 | 0.082 | 0.084 | 0.077 | 0.102 | ||||
| Normal | 1.57 | 1.30 | 0.95 | 0.80 | 0.35 | — | 1.40 | 2.34 | 2.98 | 4.35 | 1.06 | 1.17 | 1.39 | 1.42 | 0.95 | 0.98 | 0.93 | 2.66 | 7.94 | 7.56 | |||||
| NG | 1.52 | 1.39 | 1.21 | 1.01 | 0.72 | — | 0.74 | 0.56 | 0.70 | 1.65 | 0.93 | 1.00 | 1.07 | 1.18 | 1.67 | 1.00 | 1.05 | 1.24 | 1.97 | 6.88 | |||||
| RHS () | 1.64 | 1.49 | 1.23 | 1.14 | 1.08 | — | 0.54 | 0.47 | 0.44 | 0.39 | 0.88 | 0.87 | 0.82 | 0.78 | 0.83 | 1.11 | 1.15 | 1.18 | 1.23 | 1.15 | |||||
| 0.5 | R2-GP | 0.092 | 0.112 | 0.180 | 0.205 | 0.243 | — | 0.163 | 0.103 | 0.083 | 0.058 | 0.016 | 0.014 | 0.012 | 0.011 | 0.008 | 0.053 | 0.067 | 0.073 | 0.075 | 0.081 | ||||
| Normal | 1.96 | 1.64 | 1.00 | 0.90 | 0.52 | — | 1.49 | 2.60 | 3.55 | 5.09 | 0.99 | 1.02 | 1.11 | 1.16 | 2.08 | 0.97 | 1.03 | 2.03 | 4.92 | 9.85 | |||||
| NG | 1.65 | 1.41 | 1.17 | 1.13 | 0.77 | — | 0.82 | 0.62 | 0.48 | 1.44 | 0.92 | 0.91 | 0.92 | 0.95 | 1.70 | 1.00 | 1.04 | 1.00 | 1.04 | 5.30 | |||||
| RHS () | 1.76 | 1.50 | 1.17 | 1.11 | 1.06 | — | 0.58 | 0.52 | 0.46 | 0.42 | 0.87 | 0.81 | 0.83 | 0.81 | 0.92 | 1.07 | 1.11 | 1.01 | 1.02 | 1.04 | |||||
| 0.7 | R2-GP | 0.095 | 0.118 | 0.183 | 0.208 | 0.243 | — | 0.160 | 0.100 | 0.080 | 0.057 | 0.011 | 0.010 | 0.009 | 0.008 | 0.006 | 0.053 | 0.066 | 0.069 | 0.069 | 0.074 | ||||
| Normal | 2.07 | 1.64 | 1.07 | 0.97 | 0.69 | — | 1.67 | 2.89 | 3.90 | 5.74 | 0.95 | 0.98 | 1.05 | 1.04 | 2.52 | 0.98 | 1.00 | 1.64 | 3.56 | 10.69 | |||||
| NG | 1.70 | 1.34 | 1.18 | 1.18 | 0.94 | — | 0.85 | 0.60 | 0.46 | 0.95 | 0.91 | 0.91 | 0.89 | 0.93 | 1.36 | 1.00 | 1.01 | 1.02 | 0.91 | 2.52 | |||||
| RHS () | 1.79 | 1.47 | 1.14 | 1.09 | 1.05 | — | 0.59 | 0.53 | 0.47 | 0.44 | 0.86 | 0.80 | 0.88 | 0.81 | 0.89 | 1.05 | 1.11 | 0.97 | 0.98 | 0.95 |
Table 12. Relative RMSE, Sparse DGP, SE Kernel. As in Table 3, with the RMSE of taken separately across the active and the inactive components, and the RMSE of across the active components. At all components are active.
| Prior | \multicolumn{5}{c}{} | \multicolumn{5}{c}{ESS} | \multicolumn{5}{c}{Divergences} | \multicolumn{5}{c}{Time (s)} | |||||||||||||||||
| 5 | 10 | 30 | 50 | 100 | 5 | 10 | 30 | 50 | 100 | 5 | 10 | 30 | 50 | 100 | 5 | 10 | 30 | 50 | 100 | ||
| 0.3 | R2-GP | 1.00 | 1.00 | 1.01 | 1.01 | 1.01 | 1049 | 1029 | 759 | 555 | 426 | 0.0 | 0.0 | 0.1 | 4.1 | 6.2 | 7.9 | 14 | 45 | 86 | 282 |
| Normal | 1.00 | 1.01 | 1.06 | 1.28 | 1.37 | 925 | 840 | 209 | 25 | 14 | 0.5 | 0.6 | 59 | 248 | 233 | 9.8 | 21 | 188 | 847 | 3109 | |
| NG | 1.01 | 1.03 | 1.15 | 1.36 | 1.15 | 518 | 240 | 40 | 23 | 41 | 2.8 | 29 | 61 | 95 | 77 | 16 | 53 | 312 | 559 | 1934 | |
| RHS () | 1.01 | 1.01 | 1.01 | 1.01 | 1.03 | 849 | 750 | 458 | 475 | 416 | 2.0 | 0.9 | 16 | 11 | 41 | 11 | 18 | 58 | 107 | 345 | |
| 0.5 | R2-GP | 1.00 | 1.00 | 1.01 | 1.01 | 1.01 | 1181 | 1156 | 1020 | 750 | 483 | 0.1 | 0.0 | 0.1 | 0.1 | 6.0 | 9.1 | 15 | 48 | 89 | 324 |
| Normal | 1.00 | 1.00 | 1.01 | 1.19 | 1.33 | 1041 | 1000 | 576 | 40 | 18 | 0.8 | 1.4 | 9.2 | 201 | 223 | 12 | 25 | 146 | 738 | 2664 | |
| NG | 1.01 | 1.04 | 1.14 | 1.24 | 1.32 | 599 | 271 | 47 | 22 | 35 | 5.6 | 35 | 114 | 151 | 88 | 18 | 53 | 333 | 654 | 1844 | |
| RHS () | 1.00 | 1.01 | 1.01 | 1.02 | 1.02 | 984 | 888 | 704 | 542 | 419 | 2.3 | 3.5 | 9.1 | 28 | 39 | 13 | 23 | 60 | 107 | 323 | |
| 0.7 | R2-GP | 1.00 | 1.00 | 1.01 | 1.01 | 1.01 | 1239 | 1239 | 1121 | 904 | 581 | 0.5 | 0.0 | 0.1 | 0.4 | 3.4 | 11 | 18 | 53 | 101 | 372 |
| Normal | 1.00 | 1.01 | 1.01 | 1.07 | 1.37 | 1069 | 1057 | 782 | 112 | 17 | 2.1 | 1.0 | 2.4 | 170 | 204 | 15 | 32 | 162 | 666 | 2537 | |
| NG | 1.01 | 1.02 | 1.12 | 1.20 | 1.41 | 693 | 310 | 43 | 25 | 21 | 17 | 40 | 106 | 165 | 108 | 22 | 65 | 382 | 720 | 1967 | |
| RHS () | 1.00 | 1.01 | 1.01 | 1.01 | 1.03 | 1014 | 964 | 748 | 633 | 463 | 5.1 | 5.7 | 8.2 | 15 | 43 | 15 | 25 | 68 | 130 | 466 |
Table 16. Sampling Diagnostics, Sparse DGP, Matérn 3/2 Kernel. As in Table 4.
| Prior | 5 | 10 | 30 | 50 | 100 | 5 | 10 | 30 | 50 | 100 | 5 | 10 | 30 | 50 | 100 | |
| 0.3 | R2-GP | 92 | 82 | 23 | 14 | 1 | 100 | 100 | 100 | 100 | 100 | 82 | 92 | 94 | 98 | 92 |
| Normal | 92 | 96 | 96 | 96 | 100 | 100 | 100 | 100 | 100 | 100 | 88 | 86 | 12 | 0 | 0 | |
| NG | 90 | 86 | 66 | 68 | 65 | 100 | 100 | 100 | 100 | 100 | 84 | 84 | 84 | 66 | 2 | |
| RHS () | 81 | 67 | 42 | 40 | 26 | 100 | 100 | 100 | 100 | 100 | 82 | 86 | 90 | 92 | 74 | |
| 0.5 | R2-GP | 98 | 74 | 18 | 13 | 0 | 100 | 100 | 100 | 100 | 100 | 90 | 92 | 94 | 90 | 86 |
| Normal | 94 | 93 | 97 | 92 | 100 | 100 | 100 | 100 | 100 | 100 | 88 | 88 | 46 | 4 | 0 | |
| NG | 94 | 84 | 63 | 64 | 64 | 100 | 100 | 100 | 100 | 100 | 86 | 92 | 90 | 86 | 10 | |
| RHS () | 88 | 66 | 43 | 44 | 28 | 100 | 100 | 100 | 100 | 100 | 88 | 88 | 80 | 94 | 78 | |
| 0.7 | R2-GP | 98 | 76 | 19 | 10 | 0 | 100 | 100 | 100 | 100 | 100 | 88 | 94 | 94 | 90 | 86 |
| Normal | 93 | 94 | 95 | 93 | 99 | 100 | 100 | 100 | 100 | 100 | 90 | 90 | 52 | 10 | 0 | |
| NG | 94 | 87 | 62 | 61 | 70 | 100 | 100 | 100 | 100 | 100 | 88 | 90 | 88 | 86 | 34 | |
| RHS () | 89 | 70 | 42 | 45 | 32 | 100 | 100 | 100 | 100 | 100 | 86 | 92 | 86 | 90 | 88 |
Table 17. Interval Coverage (%), Sparse DGP, Matérn 5/2 Kernel. As in Table Table 11.
| Prior | , active 5 | , active 10 | , active 30 | , active 50 | , active 100 | , inactive 5 | , inactive 10 | , inactive 30 | , inactive 50 | , inactive 100 | 5 | 10 | 30 | 50 | 100 | 5 | 10 | 30 | 50 | 100 | |
| 0.3 | R2-GP | 0.100 | 0.114 | 0.183 | 0.201 | 0.240 | — | 0.162 | 0.103 | 0.087 | 0.062 | 0.020 | 0.018 | 0.015 | 0.014 | 0.011 | 0.067 | 0.070 | 0.069 | 0.085 | 0.124 |
| Normal | 1.57 | 1.33 | 0.84 | 0.79 | 0.30 | — | 1.32 | 2.31 | 2.76 | 4.04 | 1.01 | 1.17 | 1.26 | 1.41 | 1.05 | 1.00 | 1.12 | 4.31 | 8.10 | 6.19 | |
| NG | 1.54 | 1.39 | 1.12 | 0.99 | 0.69 | — | 0.77 | 0.67 | 0.83 | 1.57 | 0.92 | 1.01 | 0.98 | 1.33 | 1.77 | 1.02 | 1.05 | 1.43 | 2.47 | 5.98 | |
| RHS () | 1.62 | 1.51 | 1.22 | 1.16 | 1.09 | — | 0.57 | 0.46 | 0.44 | 0.40 | 0.88 | 0.91 | 0.78 | 0.85 | 0.80 | 1.09 | 1.08 | 1.25 | 1.21 | 1.03 | |
| 0.5 | R2-GP | 0.099 | 0.125 | 0.186 | 0.195 | 0.237 | — | 0.158 | 0.098 | 0.089 | 0.061 | 0.013 | 0.012 | 0.010 | 0.010 | 0.008 | 0.055 | 0.060 | 0.064 | 0.077 | 0.114 |
| Normal | 1.83 | 1.28 | 0.84 | 0.99 | 0.39 | — | 1.46 | 2.55 | 3.13 | 4.42 | 0.94 | 1.02 | 1.17 | 1.32 | 1.43 | 1.02 | 1.06 | 2.68 | 6.89 | 6.87 | |
| NG | 1.60 | 1.34 | 1.15 | 1.16 | 0.72 | — | 0.77 | 0.52 | 0.62 | 1.59 | 0.90 | 0.90 | 0.93 | 1.08 | 1.48 | 1.01 | 1.03 | 1.19 | 1.40 | 5.28 | |
| RHS () | 1.68 | 1.47 | 1.19 | 1.11 | 1.07 | — | 0.54 | 0.46 | 0.51 | 0.42 | 0.85 | 0.81 | 0.74 | 0.87 | 0.85 | 1.04 | 1.09 | 1.23 | 0.97 | 0.99 | |
| 0.7 | R2-GP | 0.100 | 0.126 | 0.188 | 0.197 | 0.234 | — | 0.156 | 0.096 | 0.087 | 0.062 | 0.009 | 0.008 | 0.007 | 0.007 | 0.006 | 0.053 | 0.061 | 0.063 | 0.074 | 0.102 |
| Normal | 1.92 | 1.34 | 0.92 | 1.05 | 0.55 | — | 1.56 | 2.76 | 3.36 | 4.74 | 0.90 | 0.94 | 1.06 | 0.94 | 2.26 | 1.01 | 1.00 | 2.07 | 4.93 | 7.88 | |
| NG | 1.62 | 1.32 | 1.16 | 1.20 | 0.80 | — | 0.80 | 0.52 | 0.56 | 1.27 | 0.89 | 0.89 | 0.92 | 0.87 | 1.16 | 1.00 | 1.01 | 1.17 | 1.15 | 3.79 | |
| RHS () | 1.72 | 1.45 | 1.17 | 1.10 | 1.04 | — | 0.55 | 0.47 | 0.51 | 0.44 | 0.86 | 0.83 | 0.76 | 0.81 | 0.86 | 1.01 | 1.05 | 1.20 | 0.92 | 0.98 |
Table 18. Relative RMSE, Sparse DGP, Matérn 5/2 Kernel. As in Table Table 12.
Non-normalised Competing Priors
All comparisons above apply the competing priors to the normalised components (14). Here we place the normal, NG and RHS priors on in the usual HSGP parameterisation, , in which is the marginal variance. R2-GP remains the reference, and all other settings are unchanged. Since is then on a different scale from the generating value, we compare predictive performance. Figure 17 to Figure 22 show that the qualitative behaviour of the competing priors is the same as with the normalisation.

Figure 17. Non-normalised Competing Priors, Dense DGP, SE Kernel. As in Figure 5, with the normal, NG and RHS priors on in .

Figure 22. Non-normalised Competing Priors, Sparse DGP, Matérn 5/2 Kernel. As in Figure 17.
In the dense DGP, R2-GP attains the highest mean elpd in each condition against every non-normalised prior. At the mean elpd lies between and for the normal prior, between and for NG, and between and for RHS. Since for the kernels considered, a given prior on implies less signal variance without the normalisation. This lessens the overestimation under the normal and NG priors, whose mean elpd improves by up to 37 and 24, but does not remove it. For RHS the mean elpd is lower by at most 1.4 in both DGPs. In the sparse DGP, RHS again attains a higher mean elpd than R2-GP for , by at most 12.1, and the normal prior again underperforms, with a mean elpd between and at . At the bias in has the same sign and a similar range for each prior as with the normalisation, and the normal and NG priors again fail to converge (not shown).

Figure 18. Non-normalised Competing Priors, Dense DGP, Matérn 3/2 Kernel. As in Figure 17.

Figure 19. Non-normalised Competing Priors, Dense DGP, Matérn 5/2 Kernel. As in Figure 17.

Figure 20. Non-normalised Competing Priors, Sparse DGP, SE Kernel. As in Figure 17.

Figure 21. Non-normalised Competing Priors, Sparse DGP, Matérn 3/2 Kernel. As in Figure 17.
Additional Case Study Results
Figures 23 and 24 extend Figure 4 to 12 covariates, for the SE and the Matérn 3/2 kernel. The covariates are ranked by their posterior mean share of explained variance under R2-GP with the SE kernel, and both figures show the same covariates. For these figures, the models of § 4 are fitted to all 208 compounds. Table 20 gives the predictive performance and the sampling diagnostics over the 10 train/test splits.
| Prior | SE | Matérn 3/2 | ||||||||||||
| ELPD | ESS | Div. | Time | ELPD | ESS | Div. | Time | |||||||
| R2-GP | −40.9 | 0.73 | 0.56 | 1.008 | 529 | 0.0 | 51 | −40.2 | 0.77 | 0.56 | 1.016 | 479 | 0.0 | 126 |
| RHS | −41.4 | 0.72 | 0.55 | 1.023 | 178 | 0.2 | 72 | −41.0 | 0.76 | 0.55 | 1.023 | 150 | 0.0 | 176 |
| NG | −41.3 | 0.70 | 0.55 | 1.232 | 14 | 0.0 | 61 | −40.7 | 0.74 | 0.55 | 1.037 | 71 | 0.0 | 126 |
| Normal | −46.0 | 0.95 | 0.56 | 2.117 | 5 | 2.5 | 100 | −48.9 | 1.00 | 0.55 | 3.462 | 4 | 23.8 | 257 |
Table 20. Case Study Predictive and Sampling Diagnostics

Figure 24. Covariate Effects, BloodBrain Data, Matérn 3/2 Kernel. As in Figure 23, for the same covariates.

Figure 23. Covariate Effects, BloodBrain Data, SE Kernel. Posterior means (lines) and pointwise 90% intervals (bands) of the centred effects for the 12 covariates with the largest share of explained variance. Ticks mark the observed covariate values.
Note: Predictive performance and sampling behaviour of fitted models on the BloodBrain data over 10 random train/test split replications. ELPD and posterior (in-sample and out-of-sample) are reported as means over the replications. and ESS are reported as the worst case over all model parameters across replications; Div. is the mean number of divergent transitions per fit and time the median sampling time per fit in minutes.
Computing Infrastructure
All computations ran on CPUs; no GPUs were used. The simulation study was run on an institutional high-performance computing cluster managed by Slurm, on nodes with Intel Xeon (Haswell to Cascade Lake) and AMD EPYC (Milan) processors. Each fit was allocated four cores, one per chain, and 3 GB of memory, and the workflow was orchestrated with Snakemake 8.18. Models were fitted with CmdStan 2.35.0 [3] through the R package cmdstanr 0.9.0 in R 4.5.3, compiled with GCC 12.3.
The simulation comprises 57 750 model fits, all of which completed. Sampling took 1.3 minutes per fit at the median and 5.9 minutes on average; 99% of the fits finished within about an hour and the slowest took 4.7 hours. In total this is about 5 700 hours of four-core time, or 23 000 core-hours. The case study adds 80 fits on the data splits, which took between 27 minutes and 4.5 hours each (median 1.7 hours), about 650 core-hours in total, and eight fits to the full data.
References
- [1]Javier Enrique Aguilar and Paul-Christian BürKner. Intuitive joint priors for Bayesian linear multilevel models: The R2D2M2 prior. Electronic Journal of Statistics, 17(1):1711–1767, 2023. doi: 10.1214/23-EJS2136.DOI
- [2]Javier Enrique Aguilar, David Kohns, Aki Vehtari, and Paul-Christian BürKner. R2 priors for grouped variance decomposition in high-dimensional regression. arXiv preprint arXiv:2507.11833, 2025.arxiv.org/abs/2507.11833
- [3]Bob Carpenter, Andrew Gelman, Matthew D. Hoffman, Daniel Lee, Ben Goodrich, Michael Betancourt, Marcus Brubaker, Jiqiang Guo, Peter Li, and Allen Riddell. Stan: A probabilistic programming language. Journal of Statistical Software, 76(1):1–32, 2017. doi: 10.18637/jss.v076.i01.DOI
- [4]David Duvenaud, Hannes Nickisch, and Carl Edward Rasmussen. Additive Gaussian processes. In Advances in Neural Information Processing Systems, volume 24, 2011. Page range to be confirmed.arxiv.org/abs/1112.4394
- [5]Luisa Ferrari and Massimo Ventrucci. A standardization procedure to incorporate variance partitioning based priors in latent Gaussian models. arXiv preprint arXiv:2501.16057, 2025.arxiv.org/abs/2501.16057
- [6]Geir-Arne Fuglstad, Daniel Simpson, Finn Lindgren, and Håvard Rue. Constructing priors that penalize the complexity of Gaussian random fields. Journal of the American Statistical Association, 114(525):445–452, 2019. doi: 10.1080/01621459.2017.1415907.DOI
- [7]Jim E. Griffin and Philip J. Brown. Inference with normal-gamma prior distributions in regression problems. Bayesian Analysis, 5(1):171–188, 2010. doi: 10.1214/10-BA507.DOI
- [8]Trevor Hastie and Robert Tibshirani. Generalized additive models. Statistical Science, 1(3):297–310, 1986.
- [9]Trevor J. Hastie and Robert J. Tibshirani. Generalized Additive Models. Chapman and Hall, London, 1990.
- [10]Matthew D. Hoffman and Andrew Gelman. The No-U-Turn sampler: Adaptively setting path lengths in Hamiltonian Monte Carlo. Journal of Machine Learning Research, 15:1593–1623, 2014.DOI
- [11]David Kohns, Noa Kallioinen, Yann McLatchie, and Aki Vehtari. The ARR2 prior: Flexible predictive prior definition for Bayesian auto-regressions. Bayesian Analysis, 2025. doi: 10.1214/25-BA1512. Advance publication.DOI
- [12]Max Kuhn. Building predictive models in R using the caret package. Journal of Statistical Software, 28(5):1–26, 2008.
- [13]Yann McLatchie, Sölvi Rögnvaldsson, Frank Weber, and Aki Vehtari. Advances in projection predictive inference. Statistical Science, 40(1):128–147, 2025. doi: 10.1214/24-STS949.DOI
- [14]Scot R. Mente and Franco Lombardo. A recursive-partitioning model for blood–brain barrier permeation. Journal of Computer-Aided Molecular Design, 19(7):465–481, 2005.DOI
- [15]Soham Mukherjee, Manfred Claassen, and Paul-Christian Bürkner. Hilbert space methods for approximating multi-output latent variable gaussian processes. Statistics and Computing, 36:110, 2026. doi: https://doi.org/10.1007/s11222-026-10869-x.
- [16]Radford M. Neal. MCMC using Hamiltonian dynamics. In Steve Brooks, Andrew Gelman, Galin L. Jones, and Xiao-Li Meng, editors, Handbook of Markov Chain Monte Carlo, pages 113–162. Chapman and Hall/CRC, 2011.DOI
- [17]Omiros Papaspiliopoulos, Gareth O. Roberts, and Martin Sköld. A general framework for the parametrization of hierarchical models. Statistical Science, 22(1):59–73, 2007. doi: 10.1214/088342307000000014.DOI
- [18]Juho Piironen and Aki Vehtari. Sparsity information and regularization in the horseshoe and other shrinkage priors. Electronic Journal of Statistics, 11(2):5018–5051, 2017. doi: 10.1214/17-EJS1337SI.DOI
- [19]Carl Edward Rasmussen and Christopher K. I. Williams. Gaussian Processes for Machine Learning. MIT Press, Cambridge, MA, 2006.DOI
- [20]Anna Elisabeth Riha, Leevi Lindgren, David Kohns, Paul-Christian Bürkner, and Aki Vehtari. To select or not to select: predictively consistent priors instead of model selection. arXiv preprint arXiv:2606.22850, 2026.arxiv.org/abs/2606.22850
- [21]Gabriel Riutort-Mayol, Paul-Christian Bürkner, Michael R. Andersen, Arno Solin, and Aki Vehtari. Practical Hilbert space approximate Bayesian Gaussian processes for probabilistic programming. Statistics and Computing, 33(1):17, 2023. doi: 10.1007/s11222-022-10167-2.
- [22]Daniel Simpson, Håvard Rue, Andrea Riebler, Thiago G. Martins, and Sigrunn H. Sørbye. Penalising model component complexity: A principled, practical approach to constructing priors. Statistical Science, 32(1):1–28, 2017. doi: 10.1214/16-STS576.DOI
- [23]Tuomas Sivula, Måns Magnusson, Asael Alonzo Matamoros, and Aki Vehtari. Uncertainty in Bayesian leave-one-out cross-validation based model comparison. Bayesian Analysis, 2025. doi: 10.1214/25-BA1569. Advance publication.DOI
- [24]Arno Solin and Simo Särkkä. Hilbert space methods for reduced-rank Gaussian process regression. Statistics and Computing, 30(2):419–446, 2020. doi: 10.1007/s11222-019-09886-w.DOI
- [25]Juho Timonen and Harri Lähdesmäki. Scalable Mixed-Domain Gaussian Process Modeling and Model Reduction for Longitudinal Data. Bayesian Analysis, pages 1 – 28, 2025. doi: 10.1214/25-BA1539. URL https://doi.org/10.1214/25-BA1539.
- [26]Aki Vehtari, Andrew Gelman, and Jonah Gabry. Practical Bayesian model evaluation using leave-one-out cross-validation and WAIC. Statistics and Computing, 27(5):1413–1432, 2017. doi: 10.1007/s11222-016-9696-4.DOI
- [27]Christopher Williams and Carl Rasmussen. Gaussian processes for regression. In D. Touretzky, M. C. Mozer, and M. Hasselmo, editors, Advances in Neural Information Processing Systems, volume 8. MIT Press, 1995.
- [28]Simon N. Wood. Generalized Additive Models, An Introduction with R, Second Edition. Chapman and Hall/CRC, 2017.
- [29]Eric Yanchenko, Howard D Bondell, and Brian J Reich. Spatial regression modeling via the r2d2 framework. Environmetrics, 35(2):e2829, 2024.arxiv.org/abs/2301.09951
- [30]Eric Yanchenko, Howard D Bondell, and Brian J Reich. The r2d2 prior for generalized linear mixed models. The American Statistician, 79(1):40–49, 2025.DOI
- [31]Yan Dora Zhang, Brian P. Naughton, Howard D. Bondell, and Brian J. Reich. Bayesian regression using a prior on the model fit: The R2-D2 shrinkage prior. Journal of the American Statistical Association, 117(538):862–874, 2022. doi: 10.1080/01621459.2020.1825449.








