betaFeedback

WorksMathematics

Low-rank Low Precision Training of Gaussian Processes

Abstract

Training Gaussian processes (GPs) at large scale is computationally challenging because evaluating the log-marginal likelihood and its gradients involves large, dense kernel matrices. We develop a scalable low-rank GP training approach that combines AnchorNet landmark selection with a fully independent training conditional (FITC) correction. The resulting objective and its analytic gradients can be evaluated using low-rank factors, diagonal quantities, and small dense matrices, without forming the full kernel matrix or its derivatives. The complete training procedure is implemented in single precision and evaluated on nine large-scale regression datasets. Across these problems, the method reaches the reference predictive accuracy while substantially reducing training time, although the required approximation rank varies across datasets. We also study the numerical behavior of the full-kernel predictive solve and show that single precision can become limiting for ill-conditioned problems. Finally, we investigate further reducing selected matrix operations to fp16 or TF32. These lower-precision variants provide only modest additional speedups and can become unreliable when the reduced system is poorly conditioned.

Introduction

Surrogate models can serve as accurate and computationally efficient replacements for expensive models or can represent an underlying unknown function from observed data [13]. Gaussian processes (GPs) [27] are widely used for this purpose in Bayesian optimization and machine learning because they provide both predictions at unseen inputs and a principled quantification of predictive uncertainty.

The performance of a GP depends strongly on the choice of covariance or kernel function and its associated hyperparameters [20]. A central task in GP training is therefore to estimate these hyperparameters from the available training data [19]. Both hyperparameter training and posterior prediction require matrix computations involving a large, dense covariance matrix. A direct implementation based on matrix factorization has cubic complexity in the number of training points, which quickly becomes prohibitive for large datasets. Consequently, considerable effort has been devoted to structured and low-rank approximations that reduce these computational costs [27, 1, 29, 8, 18, 25].

Further computational gains are possible by exploiting reduced-precision arithmetic on modern hardware, where single- and half-precision operations can provide substantially higher throughput and lower memory traffic than double precision. This has motivated increasing interest in low-precision algorithms for large-scale GP computations [15]. At the same time, reduced precision introduces additional numerical challenges because GP kernel matrices can become severely ill-conditioned during hyperparameter optimization.

The contributions of this paper are two-fold. First, we develop a scalable low-rank GP training approach based on landmarks selected by the AnchorNet algorithm [4] together with a fully independent training conditional (FITC) correction. We derive the corresponding objective and analytic gradients so that both the objective and its derivatives can be evaluated using low-rank factors and small dense matrices, without forming the full kernel matrix or its derivatives. Second, we study the numerical behavior of this training procedure in reduced precision on modern GPU architectures. We demonstrate that the complete low-rank training procedure can be carried out effectively in single precision and investigate the extent to which selected matrix operations can be further reduced to half precision or TF32. Our experiments also identify the conditioning regimes in which these lower-precision computations become unreliable.

The remainder of the paper is organized as follows. In § 2, we review the GP formulation and the kernel models used in this work. In § 3, we derive the low-rank FITC objective and its gradients. Section 4 describes the AnchorNet landmark-selection procedure, and § 5 summarizes the resulting training algorithm. § 6 presents the numerical experiments, including the effects of the approximation rank and the numerical behavior of the predictive solve. Finally, § 7 investigates the use and limitations of precision below fp32.

Gaussian Processes

A Gaussian process (GP) is a collection of random variables such that every finite subset has a joint Gaussian distribution [27]. It is completely characterized by its mean function

m(x):=E[f(x)]m(x) := \mathbb{E}[f(x)]

and covariance, or kernel, function

k(x,x′):=E[(f(x)−m(x))(f(x′)−m(x′))].k(x,x') := \mathbb{E}\left[(f(x)-m(x))(f(x')-m(x'))\right].

We write

f∼GP(m,k)f \sim\mathcal{GP}(m,k)

and assume a zero mean, m(x)=0m(x)=0, throughout this work.

Let Xtrain∈Rn×d\mathbf{X}_{\mathrm{train}} \in\mathbb{R}^{n \times d} contain the training inputs x1,…,xnx_1,\ldots,x_n as its rows, and define the corresponding latent function values by

ftrain=[f(x1),…,f(xn)]T.\mathbf{f}_{\mathrm{train}} = [f(x_1),\ldots,f(x_n)]^{\mathsf{T}}.

We observe

ytrain=ftrain+ε,ε∼N(0,σ2In),\mathbf{y}_{\mathrm{train}} = \mathbf{f}_{\mathrm{train}}+\boldsymbol{\varepsilon}, \qquad\boldsymbol{\varepsilon} \sim\mathcal{N}(0,\sigma^2\mathbf{I}_n),

where σ>0\sigma>0 and the observation noise is independent of the latent process. For n∗n_* test inputs Xtest∈Rn∗×d\mathbf{X}_{\mathrm{test}} \in\mathbb{R}^{n_* \times d}, define

K=k(Xtrain,Xtrain)∈Rn×n,\mathbf{K}=k(\mathbf{X}_{\mathrm{train}},\mathbf{X}_{\mathrm{train}})\in\mathbb{R}^{n\times n},
K∗=k(Xtrain,Xtest)∈Rn×n∗,K∗∗=k(Xtest,Xtest)∈Rn∗×n∗,\mathbf{K}_* = k(\mathbf{X}_{\mathrm{train}},\mathbf{X}_{\mathrm{test}})\in\mathbb{R}^{n\times n_*}, \qquad\mathbf{K}_{**}=k(\mathbf{X}_{\mathrm{test}},\mathbf{X}_{\mathrm{test}})\in\mathbb{R}^{n_*\times n_*},

where each kernel matrix is formed by evaluating kk at the corresponding pairs of inputs. Let ftest∈Rn∗\mathbf{f}_{\mathrm{test}} \in\mathbb{R}^{n_*} collect the latent function values at the test inputs. For fixed hyperparameters, the joint distribution of the training observations and test function values is

[ytrainftest]∼N(0,[K+σ2IK∗K∗TK∗∗]).\begin{bmatrix} \mathbf{y}_{\mathrm{train}} \\ \mathbf{f}_{\mathrm{test}} \end{bmatrix} \sim\mathcal{N}\left(\mathbf{0}, \begin{bmatrix} \mathbf{K}+\sigma^{2}\mathbf{I} & \mathbf{K}_{*} \\ \mathbf{K}_{*}^{T} & \mathbf{K}_{**} \end{bmatrix}\right).

Conditioning on the training observations gives

p(ftest∣Xtest,Xtrain,ytrain)=N(ftest∣μ∗,Σ∗),p(\mathbf{f}_{\mathrm{test}} \mid\mathbf{X}_{\mathrm{test}},\mathbf{X}_{\mathrm{train}},\mathbf{y}_{\mathrm{train}}) = \mathcal{N}(\mathbf{f}_{\mathrm{test}} \mid\boldsymbol{\mu}_{*},\boldsymbol{\Sigma}_{*}),
μ∗=K∗T(K+σ2I)−1ytrain,\boldsymbol{\mu}_{*} = \mathbf{K}_{*}^{T}(\mathbf{K}+\sigma^{2}\mathbf{I})^{-1}\mathbf{y}_{\mathrm{train}},
Σ∗=K∗∗−K∗T(K+σ2I)−1K∗.\boldsymbol{\Sigma}_{*} = \mathbf{K}_{**}-\mathbf{K}_{*}^{T}(\mathbf{K}+\sigma^{2}\mathbf{I})^{-1}\mathbf{K}_{*}.

The posterior mean μ∗\boldsymbol{\mu}_{*} predicts the latent function values, while Σ∗\boldsymbol{\Sigma}_{*} quantifies their uncertainty conditional on the chosen hyperparameters. For noisy test observations with independent Gaussian noise of variance σ2\sigma^{2}, the predictive covariance is Σ∗+σ2In∗\boldsymbol{\Sigma}_{*}+\sigma^{2}\mathbf{I}_{n_*}. Figure 1 illustrates the prior and posterior of a one-dimensional GP.

Plots showing a one-dimensional GP prior and posterior

Figure 1. A one-dimensional GP before and after conditioning on training data.

In the following sections, we also use the shorthand X=Xtrain\mathbf{X}=\mathbf{X}_{\mathrm{train}}, y=ytrain\mathbf{y}=\mathbf{y}_{\mathrm{train}}, and X∗=Xtest\mathbf{X}_{*}=\mathbf{X}_{\mathrm{test}}.

Kernel model. The kernel determines the covariance structure of the GP. We begin with the squared-exponential kernel

k(x,x′)=σf2exp⁡(−∥x−x′∥222ℓ2).k(x,x')=\sigma_f^{2}\exp\left(-\frac{\lVert x-x'\rVert_{2}^{2}}{2\ell^{2}}\right).

where ℓ>0\ell>0 is the length scale and σf>0\sigma_f>0 is the signal scale, so that σf2\sigma_f^{2} is the signal variance. This kernel is symmetric and positive definite.

Define the squared Euclidean distance matrix S∈Rn×n\mathbf{S}\in\mathbb{R}^{n\times n} by

Spq=∥xp−xq∥22,p,q=1,…,n.S_{pq}=\lVert x_p-x_q\rVert_{2}^{2},\qquad p,q=1,\ldots,n.

The corresponding kernel matrix is

K(ℓ,σf)=σf2exp⁡(−S2ℓ2),\mathbf{K}(\ell,\sigma_f)=\sigma_f^{2}\exp\left(-\frac{\mathbf{S}}{2\ell^{2}}\right),

where the exponential is applied entrywise.

In our experiments, we use the squared-exponential kernel with automatic relevance determination (ARD), which assigns a separate positive length scale ℓj\ell_j to each input coordinate:

k(x,x′)=σf2exp⁡(−12∑j=1d(xj−xj′)2ℓj2).(1)k(x,x') = \sigma_f^2 \exp\left(-\frac{1}{2}\sum_{j=1}^{d}\frac{(x_j-x'_j)^2}{\ell_j^2}\right). \tag*{(1)}

The hyperparameter vector is

θ=[ℓ1,…,ℓd,σf,σ]T\theta= [\ell_1,\ldots,\ell_d,\sigma_f,\sigma]^{\mathsf{T}}

and contains d+2d+2 parameters. The noise scale σ\sigma enters the observation covariance K+σ2I\mathbf{K}+\sigma^2\mathbf{I}. A large ℓj\ell_j relative to the variation in the jjth input coordinate makes the kernel less sensitive to changes along that coordinate.

For each coordinate j=1,…,dj=1,\ldots,d, define S(j)∈Rn×n\mathbf{S}^{(j)}\in\mathbb{R}^{n\times n} by

Spq(j)=(xp,j−xq,j)2,p,q=1,…,n,S_{pq}^{(j)}=(x_{p,j}-x_{q,j})^2,\qquad p,q=1,\ldots,n,

so that

S=∑j=1dS(j).\mathbf{S}=\sum_{j=1}^{d}\mathbf{S}^{(j)}.

The ARD kernel matrix can then be written as

K(θ)=σf2exp⁡(−12∑j=1dS(j)ℓj2),\mathbf{K}(\theta)=\sigma_f^2\exp\left(-\frac{1}{2}\sum_{j=1}^{d}\frac{\mathbf{S}^{(j)}}{\ell_j^2}\right),

with the exponential again applied entrywise. The isotropic kernel is recovered by setting ℓj=ℓ\ell_j=\ell for all jj. The distance matrices provide convenient notation for the kernel and its derivatives; the low-rank implementation described below forms only the required blocks.

Large-scale computation. The main computational difficulty in GP regression is that K\mathbf{K} is generally dense. For σ>0\sigma>0, the observation covariance K+σ2I\mathbf{K}+\sigma^2\mathbf{I} is symmetric positive definite in exact arithmetic. A direct Cholesky factorization requires O(n3)O(n^3) operations and O(n2)O(n^2) storage, which becomes prohibitive for large datasets [27]. This has motivated approaches based on inducing points, structured kernel approximations, and matrix-free iterative methods.

Inducing-point methods use a smaller collection of inputs to construct a low-rank representation of the covariance structure; see, e.g., [18]. Examples include the subset of regressors (SOR) [21] and the fully independent training conditional (FITC) approximation [23]. FITC augments the low-rank training covariance with a diagonal correction that restores the marginal variances at the training inputs. Structured kernel interpolation, developed by Wilson and Nickisch [28], provides another way to approximate kernel matrices efficiently. Such approximations can be combined with Krylov methods for linear solves and with stochastic trace estimation for log-determinant computations [9]. These techniques are used in scalable GP software such as GPyTorch [11].

Efficient trace estimation is also important in these computations [22, 24, 16, 7, 17, 26, 30]. Matrix-free approaches evaluate the action of K\mathbf{K} without explicitly storing the dense matrix [25, 14]. For posterior-mean computations, the resulting linear systems are closely related to those arising in kernel ridge regression and can benefit from preconditioning, including low-rank preconditioners; see, e.g., [6, 3, 10, 31].

Our method uses AnchorNet to select landmarks for a low-rank kernel approximation and combines this approximation with the FITC diagonal correction. The resulting training objective and its gradients are evaluated using low-rank factors, diagonal quantities, and small dense matrices, as developed in § 3.

Training Gaussian Processes

Exact GP training estimates the hyperparameters by minimizing the negative log-marginal likelihood

f(θ)=12(nlog⁡(2π)+log⁡det⁡(K~)+yTK~−1y),K~=K+σ2I.(2)f(\theta) = \frac{1}{2}\left(n\log(2\pi) + \log\det(\widetilde{\mathbf{K}}) + \mathbf{y}^{T}\widetilde{\mathbf{K}}^{-1}\mathbf{y}\right), \qquad\widetilde{\mathbf{K}} = \mathbf{K} + \sigma^{2}\mathbf{I}. \tag*{(2)}

For large datasets, the main computational bottlenecks are the linear solve with K~\widetilde{\mathbf{K}} required by the quadratic term and the evaluation of log⁡det⁡(K~)\log\det(\widetilde{\mathbf{K}}). A direct Cholesky factorization requires O(n3)O(n^{3}) operations and O(n2)O(n^{2}) storage [27, 2], while alternative large-scale approaches typically rely on iterative solves and trace-estimation techniques [11, 8].

To reduce these costs, we construct a low-rank approximation using landmarks selected by AnchorNet [4]. The corresponding kernel blocks satisfy

K21∈Rn×r,K11∈Rr×r,K12=K21T∈Rr×n.\mathbf{K}_{21} \in\mathbb{R}^{n \times r}, \qquad\mathbf{K}_{11} \in\mathbb{R}^{r \times r}, \qquad\mathbf{K}_{12} = \mathbf{K}_{21}^{T} \in\mathbb{R}^{r \times n}.

The construction of the landmarks and kernel blocks is described in Section 4. The landmark locations are distinct and remain fixed throughout hyperparameter optimization.

In the single-precision implementation, the landmark kernel matrix is stabilized before factorization:

K‾11=K11+ϵσf2Ir.\overline{\mathbf{K}}_{11} = \mathbf{K}_{11} + \epsilon\sigma_{f}^{2}\mathbf{I}_{r}.

where ϵ≥0\epsilon\ge0 is selected as described in § 5. For notational simplicity, in the remainder of this section, K11\mathbf{K}_{11} and K11−1\mathbf{K}_{11}^{-1} denote the stabilized matrix and its inverse, respectively. With this convention, the low-rank approximation is

K≈K^=K21K11−1K12.(3)\mathbf{K} \approx\widehat{\mathbf{K}} = \mathbf{K}_{21}\mathbf{K}_{11}^{-1}\mathbf{K}_{12}. \tag*{(3)}

All derivatives below are taken with the landmark set and the accepted value of ϵ\epsilon held fixed. Changes in ϵ\epsilon caused by the stabilization procedure are not differentiated.

The FITC-corrected low-rank model

The low-rank approximation can underestimate the marginal variances at the training inputs. For nonnegative stabilization, it satisfies

0⪯K^⪯K0 \preceq\widehat{\mathbf{K}} \preceq\mathbf{K}

in exact arithmetic, and hence

[K^]ii≤k(xi,xi).[\widehat{\mathbf{K}}]_{ii} \leq k(x_i,x_i).

When ϵ=0\epsilon= 0, the standard Nyström approximation gives equality at the landmark locations. Positive stabilization generally removes this interpolation property and increases the diagonal deficit.

We use the fully independent training conditional (FITC) correction [23] to restore the marginal variances. Let Λ\boldsymbol{\Lambda} be the diagonal matrix whose entries are

Λii=k(xi,xi)−[K^]ii.\Lambda_{ii} = k(x_i,x_i) - [\widehat{\mathbf{K}}]_{ii}.

The approximate covariance of the noisy training observations is

KFITC=K^+D,D=Λ+σ2I.(4)\mathbf{K}_{\mathrm{FITC}} = \widehat{\mathbf{K}} + \mathbf{D}, \qquad\mathbf{D} = \boldsymbol{\Lambda} + \sigma^2\mathbf{I}. \tag*{(4)}

Thus, K^+Λ\widehat{\mathbf{K}} + \boldsymbol{\Lambda} approximates the latent covariance K\mathbf{K}, while KFITC\mathbf{K}_{\mathrm{FITC}} approximates K~=K+σ2I\widetilde{\mathbf{K}} = \mathbf{K} + \sigma^2\mathbf{I}. Since Λii≥0\Lambda_{ii} \ge0 and σ>0\sigma> 0,

Dii≥σ2>0,D_{ii} \ge\sigma^2 > 0,

so both D\mathbf{D} and KFITC\mathbf{K}_{\mathrm{FITC}} are positive definite in exact arithmetic.

The approximate negative log-marginal likelihood used for training is

f^(θ)=12(nlog⁡(2π)+log⁡det⁡(KFITC)+yTKFITC−1y).(5)\widehat{f}(\theta) = \frac{1}{2}\left(n\log(2\pi) + \log\det(\mathbf{K}_{\mathrm{FITC}}) + \mathbf{y}^{T}\mathbf{K}_{\mathrm{FITC}}^{-1}\mathbf{y}\right). \tag*{(5)}

The uncorrected low-rank model is recovered by setting Λ=0\boldsymbol{\Lambda} = 0.

Efficient evaluation of the objective

The inverse and log-determinant in (5) can be evaluated using low-rank factors, diagonal operations, and r×rr \times r factorizations. Define

Ksmall:=Ir+K11−1K12D−1K21.(6)\mathbf{K}_{\mathrm{small}} := \mathbf{I}_r + \mathbf{K}_{11}^{-1}\mathbf{K}_{12}\mathbf{D}^{-1}\mathbf{K}_{21}. \tag*{(6)}

Applying the Sherman–Morrison–Woodbury formula gives

α:=KFITC−1y=D−1y−D−1K21Ksmall−1K11−1K12D−1y.(7)\boldsymbol{\alpha} := \mathbf{K}_{\mathrm{FITC}}^{-1}\mathbf{y} = \mathbf{D}^{-1}\mathbf{y} - \mathbf{D}^{-1}\mathbf{K}_{21}\mathbf{K}_{\mathrm{small}}^{-1}\mathbf{K}_{11}^{-1}\mathbf{K}_{12}\mathbf{D}^{-1}\mathbf{y}. \tag*{(7)}

Similarly, the matrix determinant lemma gives

log⁡det⁡(KFITC)=∑m=1nlog⁡Dmm+log⁡det⁡(Ksmall).(8)\log\det(\mathbf{K}_{\mathrm{FITC}}) = \sum_{m=1}^{n}\log D_{mm} + \log\det(\mathbf{K}_{\mathrm{small}}). \tag*{(8)}

Although Ksmall\mathbf{K}_{\mathrm{small}} is generally nonsymmetric, it is similar to a symmetric positive definite matrix. Let

K11=LLT,W=L−1K12∈Rr×n.\mathbf{K}_{11} = \mathbf{L}\mathbf{L}^{T}, \qquad\mathbf{W} = \mathbf{L}^{-1}\mathbf{K}_{12} \in\mathbb{R}^{r \times n}.

Then

M=LTKsmallL−T=Ir+WD−1WT.(9)\mathbf{M} = \mathbf{L}^{T}\mathbf{K}_{\mathrm{small}}\mathbf{L}^{-T} = \mathbf{I}_r + \mathbf{W}\mathbf{D}^{-1}\mathbf{W}^{T}. \tag*{(9)}

Since D\mathbf{D} is positive definite, M⪰Ir\mathbf{M} \succeq\mathbf{I}_r in exact arithmetic. The similarity transformation also gives

log⁡det⁡(Ksmall)=log⁡det⁡(M).\log\det(\mathbf{K}_{\mathrm{small}}) = \log\det(\mathbf{M}).

The inverse action in (7) can equivalently be written as

α=D−1y−D−1WTM−1WD−1y.\boldsymbol{\alpha} = \mathbf{D}^{-1}\mathbf{y} - \mathbf{D}^{-1}\mathbf{W}^{T}\mathbf{M}^{-1}\mathbf{W}\mathbf{D}^{-1}\mathbf{y}.

Thus, all dense factorizations required for evaluating the objective involve only r×rr \times r matrices. The formulas can be evaluated using diagonal scaling and triangular solves, without explicitly forming the inverse matrices.

Gradient evaluation

To optimize the hyperparameters, we differentiate (5). For any hyperparameter θi\theta_i, the standard identities for the derivative of a log-determinant and a matrix inverse give

∂f^∂θi=12[tr⁡(KFITC−1∂KFITC∂θi)−αT∂KFITC∂θiα].(10)\frac{\partial\hat{f}}{\partial\theta_i} = \frac{1}{2}\left[\operatorname{tr}\left(\mathbf{K}_{\mathrm{FITC}}^{-1}\frac{\partial\mathbf{K}_{\mathrm{FITC}}}{\partial\theta_i}\right)-\boldsymbol{\alpha}^{T}\frac{\partial\mathbf{K}_{\mathrm{FITC}}}{\partial\theta_i}\boldsymbol{\alpha}\right]. \tag*{(10)}

Using

KFITC=K^+D,\mathbf{K}_{\mathrm{FITC}}=\hat{\mathbf{K}}+\mathbf{D},

we separate the gradient into low-rank and diagonal contributions:

∂f^∂θi=12[tr⁡(KFITC−1∂K^∂θi)−αT∂K^∂θiα]+12∑m=1nqm∂Dmm∂θi,\frac{\partial\hat{f}}{\partial\theta_i}=\frac{1}{2}\left[\operatorname{tr}\left(\mathbf{K}_{\mathrm{FITC}}^{-1}\frac{\partial\hat{\mathbf{K}}}{\partial\theta_i}\right)-\boldsymbol{\alpha}^{T}\frac{\partial\hat{\mathbf{K}}}{\partial\theta_i}\boldsymbol{\alpha}\right]+\frac{1}{2}\sum_{m=1}^{n}q_m\frac{\partial D_{mm}}{\partial\theta_i},

where

qm=[KFITC−1]mm−αm2.(11)q_m=\left[\mathbf{K}_{\mathrm{FITC}}^{-1}\right]_{mm}-\alpha_m^2. \tag*{(11)}

The diagonal of KFITC−1\mathbf{K}_{\mathrm{FITC}}^{-1} can be obtained without forming the full inverse. Using W=L−1K12\mathbf{W}=\mathbf{L}^{-1}\mathbf{K}_{12} and (9),

diag⁡(KFITC−1)=diag⁡(D−1)−diag⁡(D−1WTM−1WD−1).(12)\operatorname{diag}\left(\mathbf{K}_{\mathrm{FITC}}^{-1}\right)=\operatorname{diag}\left(\mathbf{D}^{-1}\right)-\operatorname{diag}\left(\mathbf{D}^{-1}\mathbf{W}^{T}\mathbf{M}^{-1}\mathbf{W}\mathbf{D}^{-1}\right). \tag*{(12)}

The second diagonal is evaluated from blocks of the low-rank factors, without forming an n×nn \times n matrix. It remains to evaluate the derivatives of K^\hat{\mathbf{K}} and the associated reduced quantities.

Length-scale derivatives. For an ARD length scale ℓj\ell_j, all three factors in

K^=K21K11−1K12\hat{\mathbf{K}}=\mathbf{K}_{21}\mathbf{K}_{11}^{-1}\mathbf{K}_{12}

depend on ℓj\ell_j. The matrix product rule [12] therefore gives

∂K^∂ℓj=∂K21∂ℓjK11−1K12+K21∂K11−1∂ℓjK12+K21K11−1∂K12∂ℓj.(13)\frac{\partial\hat{\mathbf{K}}}{\partial\ell_j}=\frac{\partial\mathbf{K}_{21}}{\partial\ell_j}\mathbf{K}_{11}^{-1}\mathbf{K}_{12}+\mathbf{K}_{21}\frac{\partial\mathbf{K}_{11}^{-1}}{\partial\ell_j}\mathbf{K}_{12}+\mathbf{K}_{21}\mathbf{K}_{11}^{-1}\frac{\partial\mathbf{K}_{12}}{\partial\ell_j}. \tag*{(13)}

Using the ARD kernel (1) and the componentwise squared-distance matrices S(j)\mathbf{S}^{(j)} defined in § 2, the factor derivatives are

∂K21∂ℓj=1ℓj3S(j)(:,nyst)⊙K21,∂K11−1∂ℓj=−1ℓj3K11−1(S(j)(nyst,nyst)⊙K11)K11−1,∂K12∂ℓj=1ℓj3S(j)(nyst,:)⊙K12,j=1,…,d.(14)\begin{aligned} \frac{\partial\mathbf{K}_{21}}{\partial\ell_j}&=\frac{1}{\ell_j^3}\mathbf{S}^{(j)}(:,\mathrm{nyst})\odot\mathbf{K}_{21},\\ \frac{\partial\mathbf{K}_{11}^{-1}}{\partial\ell_j}&=-\frac{1}{\ell_j^3}\mathbf{K}_{11}^{-1}\left(\mathbf{S}^{(j)}(\mathrm{nyst},\mathrm{nyst})\odot\mathbf{K}_{11}\right)\mathbf{K}_{11}^{-1},\\ \frac{\partial\mathbf{K}_{12}}{\partial\ell_j}&=\frac{1}{\ell_j^3}\mathbf{S}^{(j)}(\mathrm{nyst},:)\odot\mathbf{K}_{12},\qquad j=1,\ldots,d. \tag*{(14)} \end{aligned}

Here ⊙\odot denotes the Hadamard product and nyst denotes the landmark indices identified by AnchorNet. With ϵ\epsilon fixed, the stabilization term ϵσf2Ir\epsilon\sigma_f^2\mathbf{I}_r is independent of ℓj\ell_j. Moreover, the diagonal of S(j)(nyst,nyst)\mathbf{S}^{(j)}(\mathrm{nyst},\mathrm{nyst}) is zero, so its Hadamard product with the stabilization term vanishes. Thus, (14) remains valid under our convention that K11\mathbf{K}_{11} denotes the stabilized matrix.

The derivative in (3.12) admits the low-rank factorization

∂K^∂ℓj=[∂K21∂ℓj K21 K21]⏟U(j)[K11−1∂K11−1∂ℓjK11−1]⏟G(j)[K12∂K12∂ℓj]⏟V(j).(15)\frac{\partial\widehat{\mathbf{K}}}{\partial\ell_j} = \underbrace{\left[ \frac{\partial\mathbf{K}_{21}}{\partial\ell_j}\ \mathbf{K}_{21}\ \mathbf{K}_{21} \right]}_{\mathbf{U}^{(j)}} \underbrace{\left[ \begin{matrix} \mathbf{K}_{11}^{-1} & \\ \frac{\partial\mathbf{K}_{11}^{-1}}{\partial\ell_j} & \mathbf{K}_{11}^{-1} \end{matrix} \right]}_{\mathbf{G}^{(j)}} \underbrace{\left[ \begin{matrix} \mathbf{K}_{12}\\ \frac{\partial\mathbf{K}_{12}}{\partial\ell_j} \end{matrix} \right]}_{\mathbf{V}^{(j)}} . \tag*{(15)}

The factors have dimensions U(j)∈Rn×3r\mathbf{U}^{(j)} \in\mathbb{R}^{n \times3r}, G(j)∈R3r×3r\mathbf{G}^{(j)} \in\mathbb{R}^{3r \times3r}, V(j)∈R3r×n\mathbf{V}^{(j)} \in\mathbb{R}^{3r \times n}.

This block representation can be evaluated termwise without explicitly assembling the three factors. In particular,

αT∂K^∂ℓjα=αTU(j)G(j)V(j)α,(16)\alpha^{T}\frac{\partial\widehat{\mathbf{K}}}{\partial\ell_j}\alpha = \alpha^{T}\mathbf{U}^{(j)}\mathbf{G}^{(j)}\mathbf{V}^{(j)}\alpha, \tag*{(16)}

so the quadratic-form contribution requires no n×nn \times n derivative matrix.

The low-rank trace contribution can also be reduced to an r×rr \times r computation. Since the derivative of D\mathbf{D} is accounted for separately in (3.3), we differentiate Ksmall\mathbf{K}_{\mathrm{small}} with D\mathbf{D} held fixed:

dKsmall,LR(ℓj):=∂Ksmall∂ℓj∣D fixed=∂K11−1∂ℓjK12D−1K21+K11−1∂K12∂ℓjD−1K21+K11−1K12D−1∂K21∂ℓj.\begin{aligned} \mathrm{d}\mathbf{K}_{\mathrm{small},\mathrm{LR}}^{(\ell_j)} := \left.\frac{\partial\mathbf{K}_{\mathrm{small}}}{\partial\ell_j}\right|_{\mathbf{D}\ \mathrm{fixed}} \\ &= \frac{\partial\mathbf{K}_{11}^{-1}}{\partial\ell_j}\mathbf{K}_{12}\mathbf{D}^{-1}\mathbf{K}_{21} + \mathbf{K}_{11}^{-1}\frac{\partial\mathbf{K}_{12}}{\partial\ell_j}\mathbf{D}^{-1}\mathbf{K}_{21} + \mathbf{K}_{11}^{-1}\mathbf{K}_{12}\mathbf{D}^{-1}\frac{\partial\mathbf{K}_{21}}{\partial\ell_j}. \end{aligned}

Differentiating the determinant identity with D\mathbf{D} fixed gives

tr⁡(KFITC−1∂K^∂ℓj)=tr⁡(Ksmall−1dKsmall,LR(ℓj)).(17)\operatorname{tr}\left(\mathbf{K}_{\mathrm{FITC}}^{-1}\frac{\partial\widehat{\mathbf{K}}}{\partial\ell_j}\right) = \operatorname{tr}\left(\mathbf{K}_{\mathrm{small}}^{-1}\mathrm{d}\mathbf{K}_{\mathrm{small},\mathrm{LR}}^{(\ell_j)}\right). \tag*{(17)}

The diagonal contribution follows from k(xm,xm)=σf2k(x_m,x_m)=\sigma_f^2, which is independent of the length scales:

∂Dmm∂ℓj=−[∂K^∂ℓj]mm.(18)\frac{\partial D_{mm}}{\partial\ell_j} = -\left[\frac{\partial\widehat{\mathbf{K}}}{\partial\ell_j}\right]_{mm}. \tag*{(18)}

Combining these expressions yields

∂f^∂ℓj=12[tr⁡(Ksmall−1dKsmall,LR(ℓj))−αTU(j)G(j)V(j)α−∑m=1nqm[U(j)G(j)V(j)]mm],j=1,…,d.(19)\frac{\partial\widehat{f}}{\partial\ell_j} = \frac{1}{2}\left[ \operatorname{tr}\left(\mathbf{K}_{\mathrm{small}}^{-1}\mathrm{d}\mathbf{K}_{\mathrm{small},\mathrm{LR}}^{(\ell_j)}\right) - \alpha^{T}\mathbf{U}^{(j)}\mathbf{G}^{(j)}\mathbf{V}^{(j)}\alpha - \sum_{m=1}^{n}q_m\left[\mathbf{U}^{(j)}\mathbf{G}^{(j)}\mathbf{V}^{(j)}\right]_{mm} \right], \qquad j=1,\ldots,d. \tag*{(19)}

Signal-scale derivative. The signal variance σf2\sigma_f^2 scales every kernel block. Because the stabilization has the form ϵσf2Ir\epsilon\sigma_f^2\mathbf{I}_r with fixed ϵ\epsilon, the stabilized landmark matrix also scales proportionally to σf2\sigma_f^2. Consequently,

∂K21∂σf=2σfK21,∂K12∂σf=2σfK12,∂K11−1∂σf=−2σfK11−1.(20)\frac{\partial\mathbf{K}_{21}}{\partial\sigma_f} = \frac{2}{\sigma_f}\mathbf{K}_{21}, \qquad \frac{\partial\mathbf{K}_{12}}{\partial\sigma_f} = \frac{2}{\sigma_f}\mathbf{K}_{12}, \qquad \frac{\partial\mathbf{K}_{11}^{-1}}{\partial\sigma_f} = -\frac{2}{\sigma_f}\mathbf{K}_{11}^{-1}. \tag*{(20)}

The product rule then gives

∂K^∂σf=2σfK^.(21)\frac{\partial\widehat{\mathbf{K}}}{\partial\sigma_f}=\frac{2}{\sigma_f}\widehat{\mathbf{K}}. \tag*{(21)}

As in the length-scale calculation, the low-rank contribution holds D\mathbf{D} fixed:

dKsmall,LR(σf):=∂Ksmall∂σf∣D fixed=2σfK11−1K12D−1K21=2σf(Ksmall−Ir).(22)\mathrm{d}\mathbf{K}^{(\sigma_f)}_{\mathrm{small,LR}} := \left.\frac{\partial\mathbf{K}_{\mathrm{small}}}{\partial\sigma_f}\right|_{\mathbf{D}\ \mathrm{fixed}} = \frac{2}{\sigma_f}\mathbf{K}_{11}^{-1}\mathbf{K}_{12}\mathbf{D}^{-1}\mathbf{K}_{21} = \frac{2}{\sigma_f}\left(\mathbf{K}_{\mathrm{small}}-\mathbf{I}_r\right). \tag*{(22)}

Since

Λmm=σf2−[K^]mm,\Lambda_{mm}=\sigma_f^2-[\widehat{\mathbf{K}}]_{mm},

the diagonal derivative is

∂Dmm∂σf=2σfΛmm.(23)\frac{\partial D_{mm}}{\partial\sigma_f}=\frac{2}{\sigma_f}\Lambda_{mm}. \tag*{(23)}

The resulting signal-scale gradient is

∂f^∂σf=12[tr⁡(Ksmall−1dKsmall,LR(σf))−2σfαTK^α+2σf∑m=1nqmΛmm].\frac{\partial\widehat{f}}{\partial\sigma_f}=\frac{1}{2}\left[\operatorname{tr}\left(\mathbf{K}_{\mathrm{small}}^{-1}\mathrm{d}\mathbf{K}^{(\sigma_f)}_{\mathrm{small,LR}}\right)-\frac{2}{\sigma_f}\boldsymbol{\alpha}^{T}\widehat{\mathbf{K}}\boldsymbol{\alpha}+\frac{2}{\sigma_f}\sum_{m=1}^{n}q_m\Lambda_{mm}\right].

Noise derivative. The low-rank approximation K^\widehat{\mathbf{K}} and the diagonal correction Λ\mathbf{\Lambda} are independent of the noise scale σ\sigma, while

∂D∂σ=2σIn.\frac{\partial\mathbf{D}}{\partial\sigma}=2\sigma\mathbf{I}_n.

Thus, only the diagonal contribution in (4) remains:

∂f^∂σ=σ∑m=1nqm=σ(tr⁡KFITC−1−αTα).(24)\frac{\partial\widehat{f}}{\partial\sigma}=\sigma\sum_{m=1}^{n}q_m=\sigma\left(\operatorname{tr}\mathbf{K}_{\mathrm{FITC}}^{-1}-\boldsymbol{\alpha}^{T}\boldsymbol{\alpha}\right). \tag*{(24)}

All quantities in (19), (4), and (3.23) can be evaluated using low-rank factors, diagonal quantities, and reduced r×rr \times r matrices. Neither the full kernel matrix nor any n×nn \times n derivative matrix needs to be formed. The componentwise squared-distance matrices S(j)\mathbf{S}^{(j)} are likewise not stored explicitly; the required Hadamard products are generated blockwise together with the corresponding kernel blocks.

Positive parameterization

To enforce positivity without constrained optimization, we introduce unconstrained parameters

ζ=(ζℓ1,…,ζℓd,ζσf,ζσ)T∈Rd+2\boldsymbol{\zeta}=(\zeta_{\ell_1},\ldots,\zeta_{\ell_d},\zeta_{\sigma_f},\zeta_{\sigma})^{T}\in\mathbb{R}^{d+2}

and parameterize

ℓj=softplus⁡(ζℓj),σf=softplus⁡(ζσf),\ell_j=\operatorname{softplus}(\zeta_{\ell_j}),\qquad\sigma_f=\operatorname{softplus}(\zeta_{\sigma_f}),

and

σ=10−4+softplus⁡(ζσ),softplus⁡(ζ)=log⁡(1+eζ).(25)\sigma=10^{-4}+\operatorname{softplus}(\zeta_{\sigma}),\qquad\operatorname{softplus}(\zeta)=\log(1+e^{\zeta}). \tag*{(25)}

Since

softplus⁡′(ζ)=(1+e−ζ)−1,\operatorname{softplus}'(\zeta)=(1+e^{-\zeta})^{-1},

the derivatives with respect to the unconstrained optimization variables follow from the chain rule. Writing ζi\zeta_i for the component corresponding to θi\theta_i, we obtain

∂f^∂ζi=∂f^∂θi−1,i=1,…,d+2.(26)\frac{\partial\hat{f}}{\partial\zeta_i}=\frac{\partial\hat{f}}{\partial\theta_i}^{-1},\qquad i=1,\ldots,d+2. \tag*{(26)}

This formula also applies to the noise parameter because its additional floor is constant.

AnchorNet Low-Rank Approximation

We use AnchorNet [4] to select landmark points for the low-rank kernel approximation. Given a set of landmarks SS, the standard Nyström approximation is

K≈K^=K21K11−1K12,\mathbf{K}\approx\widehat{\mathbf{K}}=\mathbf{K}_{21}\mathbf{K}_{11}^{-1}\mathbf{K}_{12},

where K21∈Rn×r\mathbf{K}_{21}\in\mathbb{R}^{n\times r}, K11∈Rr×r\mathbf{K}_{11}\in\mathbb{R}^{r\times r}, and K12=K21T∈Rr×n\mathbf{K}_{12}=\mathbf{K}_{21}^{\mathsf{T}}\in\mathbb{R}^{r\times n}. Here r=#Sr=\#S is the number of distinct landmarks.

Let

X={x1,…,xn}⊂Rd\mathcal{X}=\{x_1,\ldots,x_n\}\subset\mathbb{R}^d

denote the training locations stored in the rows of Xtrain\mathbf{X}_{\mathrm{train}}. Let mm be a positive integer specifying the target net size. The construction has two stages: it first generates anchor points adapted to the data geometry and then maps them to training points. All nearest-point ties are resolved by a fixed rule.

Anchor construction.

  1. Compute the smallest axis-aligned bounding box B0B_0 containing X\mathcal{X} and generate an initial low-discrepancy set

T={t1,…,ts}⊂B0,s=O(m).\mathcal{T}=\{t_1,\ldots,t_s\}\subset B_0,\qquad s=O(m).

Possible constructions include Halton points and the adaptive tensor grids described in [4].

  1. Assign each training point xjx_j to one nearest point in T\mathcal{T} with respect to the ℓ∞\ell_\infty norm. Specifically, choose

a(j)∈arg min⁡1≤i≤s∥xj−ti∥∞a(j)\in\operatorname*{arg\,min}_{1\le i\le s}\lVert x_j-t_i\rVert_\infty

according to the fixed tie-breaking rule and define

Gi={xj:a(j)=i},i=1,…,s.G_i=\{x_j:a(j)=i\},\qquad i=1,\ldots,s.

Discard empty groups and relabel the remaining groups as G1,…,GQG_1,\ldots,G_Q.

  1. For each nonempty group GiG_i, compute its smallest axis-aligned bounding box BiB_i. Let λ(Bi)\lambda(B_i) denote its dd-dimensional Lebesgue measure. For the volume-based allocation, assume

∑k=1Qλ(Bk)>0\sum_{k=1}^{Q}\lambda(B_k)>0

and set

mi=⌊mλ(Bi)∑k=1Qλ(Bk)⌋.m_i=\left\lfloor m\frac{\lambda(B_i)}{\sum_{k=1}^{Q}\lambda(B_k)}\right\rfloor.

When mi>0m_i>0, generate a low-discrepancy set A(i)⊂BiA^{(i)}\subset B_i with mim_i points. If mi=0m_i=0, set A(i)=∅A^{(i)}=\varnothing.

Form the anchor net

AX=⋃i=1QA(i).\mathcal{A}_X = \bigcup_{i=1}^{Q} \mathcal{A}^{(i)}.

The ceiling allocation gives

#AX≤∑i=1Qmi≤m+Q=O(m),\#\mathcal{A}_X \le\sum_{i=1}^{Q} m_i \le m + Q = O(m),

because Q≤s=O(m)Q \le s = O(m). Thus, the number of distinct anchors need not equal the target size mm.

From anchors to landmarks. For each anchor y∈AXy \in\mathcal{A}_X, select a nearest training point,

s(y)∈argmin⁡x∈X∥x−y∥∞,s(y) \in\operatorname*{argmin}_{x \in\mathcal{X}} \lVert x-y\rVert_{\infty},

using the fixed tie-breaking rule. The landmark set is

S={s(y):y∈AX}⊂X,S = \{s(y):y \in\mathcal{A}_X\} \subset\mathcal{X},

with repeated selections removed. Consequently,

r=#S≤#AX≤m+Q=O(m),r = \#S \le\#\mathcal{A}_X \le m + Q = O(m),

and r≤nr \le n.

The volume-based allocation is designed to distribute anchors across the data geometry. Mapping these anchors to the data produces landmarks intended to improve the approximation quality and numerical behavior of the Nyström approximation [4]. The landmark-selection stage costs O(mdn)O(mdn) and depends only on the input locations, requiring neither kernel matrix entries nor kernel matrix–vector products [4]. The selected landmarks define the low-rank approximation described in § 3, with stabilization implemented as described in § 5.

Algorithm

5. Algorithm. Algorithm 5.1 summarizes the training and prediction procedure. The landmarks are selected once and remain fixed during optimization, with r=#Sr = \#S denoting the resulting approximation rank. Initialization and optimizer settings are given in Appendix § 9.

Training is performed in fp32. The stabilization routine factors

K‾11=K11+ϵσf2Ir=LLT,\overline{K}_{11} = K_{11} + \epsilon\sigma_f^2 I_r = LL^{\mathsf{T}},

starting with ϵ=10−4\epsilon= 10^{-4} and increasing it by a factor of ten after an unsuccessful attempt, up to seven attempts. Failure is reported if no acceptable factor is obtained; this did not occur in the reported experiments. The gradient formulas in § 3 use K‾11\overline{K}_{11} and its inverse, with the accepted ϵ\epsilon held fixed. Final factor reconstruction repeats the same stabilization procedure. The operator clip⁡10\operatorname{clip}_{10} clips the Euclidean norm of the unconstrained gradient to 10.

The factor computations are blocked by default. We select the unblocked implementation only when an initial timing test shows that it is at least 20% faster and sufficient memory is available. Kernel products during prediction can also be evaluated in blocks.

The low-rank predictive mean uses only reduced solves and kernel products with the landmarks. For the full-kernel mean, datasets with fewer than 60,000 training Algorithm 5.1 FITC Gaussian process training with AnchorNet landmarks

Input: Training data (X,y)(\mathbf{X},\mathbf{y}); test inputs X∗\mathbf{X}_*; target net size mm; iteration budget T≥2T \ge2; initial learning rate η0\eta_0; low-rank predictive-mean flag bLRb_{\mathrm{LR}}

Output: Hyperparameters θ\boldsymbol{\theta} and predictive mean μ∗\boldsymbol{\mu}_*

  1. S←AnchorNet⁡(X,m)\mathcal{S} \leftarrow\operatorname{AnchorNet}(\mathbf{X},m); r←#Sr \leftarrow\#\mathcal{S}

  2. θ0←Initialize⁡(X,y)\boldsymbol{\theta}_0 \leftarrow\operatorname{Initialize}(\mathbf{X},\mathbf{y})

  3. ζ←InversePositiveMap⁡(θ0)\boldsymbol{\zeta} \leftarrow\operatorname{InversePositiveMap}(\boldsymbol{\theta}_0)

  4. initialize the Adam optimizer state

  5. for t=0,…,T−1t = 0,\ldots,T-1 do

  6.   set ηt\eta_t using (A.2)

  7.   θ←PositiveMap⁡(ζ)\boldsymbol{\theta} \leftarrow\operatorname{PositiveMap}(\boldsymbol{\zeta})

  8.   form K11=k(S,S)\mathbf{K}_{11}=k(\mathcal{S},\mathcal{S}), K21=k(X,S)\mathbf{K}_{21}=k(\mathbf{X},\mathcal{S}), and K12=K21T\mathbf{K}_{12}=\mathbf{K}_{21}^{T}

  9.   (L,ϵ)←StabilizedCholesky⁡(K11,σf)(\mathbf{L},\epsilon) \leftarrow\operatorname{StabilizedCholesky}(\mathbf{K}_{11},\sigma_f)

  10.   compute W,D,M\mathbf{W},\mathbf{D},\mathbf{M} as in Section 3 and factor M\mathbf{M}

  11.   evaluate f^\hat{f} and ∇θf^\nabla_{\boldsymbol{\theta}}\hat{f} using the stabilized factors, with ϵ\epsilon fixed

  12.   compute g=∇ζf^\mathbf{g}=\nabla_{\boldsymbol{\zeta}}\hat{f} using (26)

  13.   ζ←Adam⁡(ζ,clip⁡10(g),ηt)\boldsymbol{\zeta}\leftarrow\operatorname{Adam}(\boldsymbol{\zeta},\operatorname{clip}_{10}(\mathbf{g}),\eta_t)

  14.   θ←PositiveMap⁡(ζ)\boldsymbol{\theta}\leftarrow\operatorname{PositiveMap}(\boldsymbol{\zeta})

  15.   if bLR=Trueb_{\mathrm{LR}}=\mathrm{True} then

  16.     recompute L,W,D,M\mathbf{L},\mathbf{W},\mathbf{D},\mathbf{M} at the final θ\boldsymbol{\theta}

  17.     factor M\mathbf{M} and solve Mu=WD−1y\mathbf{M}\mathbf{u}=\mathbf{W}\mathbf{D}^{-1}\mathbf{y}

  18.     solve LTc=u\mathbf{L}^{T}\mathbf{c}=\mathbf{u}

  19.     evaluate μ∗←k(X∗,S)c\boldsymbol{\mu}_*\leftarrow k(\mathbf{X}_*,\mathcal{S})\mathbf{c}

  20.   else

  21.     solve (k(X,X)+σ2In)αfull=y\left(k(\mathbf{X},\mathbf{X})+\sigma^2\mathbf{I}_n\right)\boldsymbol{\alpha}_{\mathrm{full}}=\mathbf{y} at the final θ\boldsymbol{\theta}

  22.     evaluate μ∗←k(X∗,X)αfull\boldsymbol{\mu}_*\leftarrow k(\mathbf{X}_*,\mathbf{X})\boldsymbol{\alpha}_{\mathrm{full}}

points use a dense factorization; larger datasets use preconditioned CG with KeOps matrix–vector products. The default fp32 preconditioner uses the first min⁡(r,200)\min(r,200) landmarks. Additional fp64 solves establish the full-kernel reference predictions on 3DRoad, Buzz, and HouseElectric, where fp32 does not attain sufficient accuracy; see § 6 and appendix § 10.

Numerical experiments

We evaluate the low-rank model described in Algorithm 5.1, which we refer to as ANGP (AnchorNet Gaussian process),1 on the nine UCI regression datasets used by Maddox et al. [15]. The datasets are obtained through the public bayesian_benchmarks interface and range from n=13,500n=13{,}500 to n=1,844,352n=1{,}844{,}352 training points, with input dimension dd ranging from 3 to 90, as summarized in Table 1. We use the standard preprocessing provided by bayesian_benchmarks; in particular, the input features and targets are standardized.

dataset(n,d)(n, d)rrRMSE ANGPRMSE Maddoxtime [s] ANGPtime [s] Maddoxspeed-up
PoleTele(13.5K, 26)1000.1075 ± 0.00850.117 ± 0.0042.25826.6×
Elevators(14.9K, 18)4000.3607 ± 0.00370.364 ± 0.0043.45817.4×
Bike(15.6K, 17)1000.0707 ± 0.00710.074 ± 0.0032.46226.4×
Kin40K(36K, 8)16000.0981 ± 0.00180.099 ± 0.00117.4663.8×
Protein(41.2K, 9)16000.5406 ± 0.00680.550 ± 0.00620.5713.5×
3DRoad(391.4K, 3)30000.1915 ± 0.0040∗^{*}0.194 ± 0.0103731,2603.4×
Song(463.8K, 90)30000.7605 ± 0.00310.761 ± 0.0041,40424,35717.4×
Buzz(524.9K, 77)500.2672 ± 0.0056∗^{*}0.300 ± 0.01022.725,4361,120.3×
HouseElectric(1844.4K, 11)8000.0518 ± 0.0003∗^{*}0.052 ± 0.00243142,75199.1×

Table 1. Fit time comparison with Maddox [15 Table 1].

Throughout the experiments, we use the same kernel, hyperparameter initialization, optimizer, and optimization budget for all datasets. The only model parameter varied across datasets is the number of AnchorNet landmarks, which determines the approximation rank rr. We use the squared-exponential kernel with automatic relevance determination (ARD) in (1), as in [15], resulting in d+2d+2 hyperparameters.

The hyperparameters are initialized using the data-driven procedure described in Appendix § 9: the length scales are initialized using the median pairwise distance heuristic, the signal scale using the standard deviation of the target, and the noise scale using the mode of its prior. We optimize the FITC negative log-marginal likelihood using Adam for a fixed budget of 100 iterations with a cosine-decayed learning rate. We use neither a validation split nor early stopping and report the final iterate for every run. With the initialization and learning-rate schedule used here, we did not observe divergence, collapse of a length scale, or activation of the noise floor in the reported experiments.

Each evaluation of the FITC objective and its gradient has cost O(nr(r+d))O(nr(r+d)) in our implementation. The unblocked implementation stores O(nr)O(nr) intermediate quantities, whereas the blocked implementation processes the training data in row blocks of size BB and requires O(rB+r2)O(rB+r^2) working memory. At the beginning of each fit, we time one objective-and-gradient evaluation using both implementations. The blocked implementation is used by default; the unblocked variant is selected only when it is at least 20% faster and fits in available memory. Additional implementation details are given in Appendix § 9.

For prediction, datasets with fewer than 60,000 training points use a dense full-kernel factorization. For the larger datasets, the full-kernel predictive mean is computed using preconditioned conjugate gradients (CG), with matrix–vector products evaluated through KeOps [5]. We also evaluate the predictive mean obtained directly from the fitted low-rank model, which avoids the full-kernel linear solve.

Each configuration is evaluated on five data splits, and we report the mean and standard deviation. For each dataset, we test several landmark counts and report every tested configuration in Table 2. When we subsequently identify the cheapest rank that reaches a reference RMSE, this should be interpreted as a retrospective cost–accuracy comparison rather than as a rank-selection procedure used during training.

rrPoleTele (0.117)Elevators (0.364)Bike (0.074)Kin40K (0.099)Protein (0.550)3DRoad* (0.194)Song (0.761)Buzz* (0.300)HouseElectric* (0.052)
500.12510.38060.09240.26630.75350.63190.77400.26720.0551
1000.10750.37710.07070.21350.70860.53890.77100.26300.0549
2000.10150.36670.04970.17640.67650.49440.76850.26550.0537
2500.10000.36490.04800.17430.65390.48280.76770.26420.0534
4000.10110.36070.04060.13710.61510.39910.76610.26480.0526
5000.10160.35920.03890.13060.59520.38050.76520.26380.0522
8000.10110.35650.03490.11060.56490.30150.76400.26220.0518
10000.10340.35550.03340.10510.55240.28710.76340.26280.0516
16000.10430.35450.03030.09810.54060.25020.76200.0506
20000.10550.35420.09310.53430.23160.76130.0490
30000.10760.35390.08350.52820.19150.76050.0467
40000.07590.52470.17530.7598
60000.06830.52160.1511
80000.06490.5207
speed-up26.6×17.4×26.4×3.8×3.5×3.4×17.4×1120.3×99.1×

Table 2. Test RMSE from the posterior mean solve against rank. The last row reports the speedup for hyperparameter fitting at the smallest rank that achieves at least the same accuracy as Maddox.

All ANGP timings are measured on an NVIDIA A40 GPU with 48 GB of memory. We compare against the V100 (32 GB) timings reported by Maddox et al. [15], denoted by Maddox in the tables. Following their experimental setup, the reported fitting times include hyperparameter optimization but exclude the predictive solve.

The two sets of timings are obtained on different hardware. The A40 has roughly a 2.4× advantage in nominal single-precision throughput over the V100, so part of the observed speedup is due to hardware rather than the algorithm alone. We therefore report the raw timings and interpret a speedup divided by 2.4 as an approximate hardware-normalized comparison. Under this normalization, ANGP remains faster than the reference on all datasets, although the advantage on 3DRoad becomes small.

The remainder of this section examines three aspects of the method: the cost–accuracy tradeoff as the approximation rank varies, the difference between the low-rank and full-kernel predictive means, and the numerical behavior of the full-kernel predictive solve in single precision.

Comparison with the reference

Table 1 reports, for each dataset, the cheapest tested rank whose mean test RMSE reaches the reference value reported in [15]. As noted above, this selection is performed retrospectively from the tested ranks and is intended to summarize the cost required to attain the reference accuracy.

Whenever the fp32 full-kernel solve converges, we compute the reported RMSE using that predictive mean. For 3DRoad, Buzz, and HouseElectric, the fp32 CG solve does not attain sufficient accuracy because of finite-precision effects. For these datasets, we instead report the RMSE obtained from the corresponding fp64 full-kernel solve; these entries are marked by * in Table 1.

The Protein RMSE reported in [15 Table 1] is 0.055. This value is inconsistent with both the half-precision result (0.635) in the same table and the values 0.481–0.509 reported in the appendix of that paper. We therefore interpret 0.055 as a misplaced decimal point and use 0.55 as the reference value throughout.

Table 2 shows that ANGP reaches the reference RMSE on all nine datasets for at least one of the tested ranks. Increasing rr generally improves the quality of the low-rank model but also increases the fitting cost, which grows quadratically with rr in the dominant low-rank operations. The rank required to reach the reference accuracy varies substantially across datasets.

For example, 3DRoad requires r=3000r = 3000 to reach the reference accuracy. At r=6000r = 6000, ANGP is already slower than Maddox even before accounting for the hardware difference; see Table 4. Across the datasets, the raw fitting speedups range from 3.4× on 3DRoad to 1120× on Buzz. On 3DRoad, the observed speedup is only somewhat larger than the approximately 2.4× hardware advantage of the A40, leaving about 1.4× after this approximate hardware normalization. One contributing factor is that 3DRoad requires a relatively large approximation rank while having only d=3d = 3 input dimensions, making the kernel matrix–vector products used by the reference method comparatively inexpensive.

rreff.fit [s]×\timesreflow-rank mean RMSElow-rank mean time [s]float32 CG RMSEfloat32 CG time [s]float64 CG RMSEfloat64 CG time [s]
50494.10307.3×\times0.6931 ±\pm 0.00510.010.6319 ±\pm 0.0211330.6319 ±\pm 0.021147.6
1001006.26201.1×\times0.6320 ±\pm 0.00690.010.5389 ±\pm 0.0110330.5389 ±\pm 0.011094.1
20019810.98114.7×\times0.5771 ±\pm 0.00300.030.4944 ±\pm 0.0115330.4944 ±\pm 0.0115110
25024913.3294.6×\times0.5641 ±\pm 0.00180.030.4828 ±\pm 0.0112330.4828 ±\pm 0.0113128
40039722.5455.9×\times0.5232 ±\pm 0.00620.060.3992 ±\pm 0.0194330.3991 ±\pm 0.0195487
50049627.4545.9×\times0.5026 ±\pm 0.00330.070.3805 ±\pm 0.0102330.3805 ±\pm 0.0101553
80079550.9624.7×\times0.4479 ±\pm 0.00730.150.3015 ±\pm 0.0112330.3015 ±\pm 0.01131,420
100098967.2318.7×\times0.4241 ±\pm 0.00510.190.2876 ±\pm 0.0064330.2871 ±\pm 0.00621,588
16001573135.559.3×\times0.3811 ±\pm 0.00560.390.2557 ±\pm 0.0077330.2502 ±\pm 0.00652,091
20001959192.696.5×\times0.3627 ±\pm 0.00720.510.2371 ±\pm 0.0050330.2316 ±\pm 0.00662,469
30002921372.893.4×\times0.3302 ±\pm 0.00380.900.2041 ±\pm 0.0052330.1915 ±\pm 0.00402,950
40003869743.441.7×\times0.3062 ±\pm 0.00301.420.1953 ±\pm 0.0069330.1753 ±\pm 0.00193,221
600057321,419.640.9×\times0.2716 ±\pm 0.00402.910.1822 ±\pm 0.0057330.1511 ±\pm 0.00203,556

Table 4. Table 4: Test RMSE and prediction time of the three predictive means on 3DRoad. The reference RMSE is 0.194.

The effect of rank

Table 2 reports the test RMSE for every rank tested. Entries that match or improve upon the reference value in [15 Table 1] are shown in bold. The rank required to reach the reference varies by roughly two orders of magnitude and does not exhibit a simple dependence on either the number of training points nn or the input dimension dd.

As in the previous subsection, the full-kernel predictive mean is computed in fp64 for 3DRoad, Buzz, and HouseElectric when the fp32 CG solve does not attain sufficient accuracy. On Buzz, increasing the model rank does not lead to a corresponding improvement in the full-kernel predictive RMSE, and some of the best results occur at relatively small ranks.

The reference RMSE is reached at rank 100 for PoleTele and Bike, at rank 1600 for Kin40K, and at rank 3000 for both Song and 3DRoad. The variation cannot be explained by nn or dd alone. For example, Song, with n=464,000n = 464{,}000, requires rank 3000, whereas HouseElectric, with n=1.84Mn = 1.84\mathrm{M}, reaches the reference at rank 800. Similarly, 3DRoad has only d=3d = 3 input dimensions but requires a larger rank than PoleTele, which has d=26d = 26. The required rank is instead strongly influenced by how rapidly the fitted kernel spectrum decays, together with the geometry of the particular dataset; these properties are not known before fitting.

rrPoleTele (0.117)Elevators (0.364)Bike (0.074)Kin40K (0.099)Protein (0.550)3DRoad (0.194)Song (0.761)Buzz (0.300)HouseElectric (0.052)
500.35950.39830.18120.67520.79120.69310.83330.42720.1012
1000.31620.39030.15660.57430.76230.63200.82270.37960.0708
2000.27440.37720.11140.41300.73250.57710.81450.37010.0596
2500.26320.37420.11070.37570.72300.56410.81110.36190.0586
4000.24380.36680.09320.32220.70040.52320.80690.35470.0566
5000.23080.36460.08640.29650.68250.50260.80400.34420.0560
8000.20100.36060.06460.25120.65650.44790.79880.33230.0542
10000.18820.35850.06870.22500.64370.42410.79660.32920.0536
16000.16980.35720.05170.18050.61800.38110.79220.0534
20000.15490.35610.16130.60940.36270.79040.0526
30000.14160.35490.13440.58920.33020.78620.0522
40000.11980.11980.58050.30620.7838
60000.10130.10130.56920.2716
80000.09150.5598

Table 3. Test RMSE from the low-rank predictive mean for the fits of Table 2.

The rank also exposes a clear accuracy–cost tradeoff. On Bike, for example, the RMSE decreases from 0.0707 at rank 100 to 0.0303 at rank 1600, while the fitting time increases from 2.4 s to 10.2 s. Rank 100 is sufficient to match the reference, but a larger rank can substantially improve predictive accuracy at additional cost. A similar effect occurs on Kin40K, where the RMSE decreases to 0.0649 at rank 8000, compared with the reference value 0.099, although rank 1600 is already sufficient to reach the reference. Thus, the cheapest rank that matches the reference should not be interpreted as an accuracy-optimal rank.

Table 3 reports the same trained models but uses the predictive mean of the low-rank FITC model instead of solving the full-kernel system. This predictor requires no iterative full-kernel solve, can be evaluated in closed form using the low-rank factors, and therefore cannot suffer from iterative convergence failure. Across all reported configurations, its RMSE has a median relative increase of approximately 27% compared with the full-kernel predictive mean. It reaches the reference accuracy on Elevators, Bike, and Kin40K. The gap generally decreases as rr increases, but at markedly different rates across datasets. On HouseElectric, for example, the low-rank mean is within approximately 4% of the full-kernel result at r=800r = 800, whereas on Buzz it remains at least 24% worse over the tested ranks.

rreff.fit [s]×\timesreflow-rank mean RMSElow-rank mean time [s]float32 CG RMSEfloat32 CG time [s]float64 CG RMSEfloat64 CG time [s]
504622.711120.3×\times0.4272 ±\pm 0.02360.010.2783 ±\pm 0.01821,3430.2672 ±\pm 0.005617,142
1008339.68641.0×\times0.3796 ±\pm 0.01680.020.2709 ±\pm 0.01581,3410.2630 ±\pm 0.002919,022
20015174.94339.4×\times0.3701 ±\pm 0.01390.040.3021 ±\pm 0.08331,3450.2655 ±\pm 0.004520,687
25017593.16273.0×\times0.3619 ±\pm 0.01000.050.2723 ±\pm 0.00571,3370.2642 ±\pm 0.004720,396
400244149.77169.8×\times0.3547 ±\pm 0.01600.080.3047 ±\pm 0.02651,3440.2648 ±\pm 0.003920,945
500281186.65136.3×\times0.3442 ±\pm 0.00710.100.3132 ±\pm 0.02821,3410.2638 ±\pm 0.004020,949
800393308.3582.5×\times0.3323 ±\pm 0.00800.180.3563 ±\pm 0.05821,3440.2622 ±\pm 0.003222,623
1000461389.4465.3×\times0.3292 ±\pm 0.00330.250.3414 ±\pm 0.02711,3360.2628 ±\pm 0.003022,060

Table 5. Table 5: Test RMSE and prediction time of the three predictive means on Buzz. The reference RMSE is 0.300.

Tables 4–6 compare the low-rank predictive mean with the fp32 and fp64 full-kernel predictive solves on the three largest datasets for which these comparisons are most informative. Results are averaged over five splits. The “effective rank” (eff.) denotes the numerical rank of K11K_{11}; when it is substantially smaller than the nominal landmark count, additional landmarks contribute little independent numerical information. The CG columns report predictive-solve times, whereas the column ×\timesref reports the fitting-time speedup relative to Maddox. Entries that reach the Maddox reference RMSE are shown in bold.

The low-rank predictive mean requires at most 4.1 s in these experiments, compared with 33 s to 25 min for the fp32 full-kernel CG solve and up to 24 h for the fp64 solve. Consequently, for the low-rank predictor, fitting dominates prediction time, whereas the full-kernel solve can itself become a dominant component of the total cost.

HouseElectric illustrates the advantage of avoiding the full-kernel solve. The fp32

CG solve does not attain the required accuracy at any tested rank, while the low-rank mean reaches an RMSE of 0.0536 at r=1000r = 1000, within the standard deviation of the reference (0.052 ± 0.002). Its prediction requires 0.9 s following a 544 s fit, and the fitting stage is 78.5× faster than the reported Maddox reference. The complete low-rank pipeline in this case remains in single precision. On 3DRoad and Buzz, by contrast, the low-rank predictive mean is inexpensive but does not reach the accuracy of the full-kernel predictor, so the latter remains necessary when the highest predictive accuracy is required.

The predictive-mean solve in single precision

For datasets with n<60,000n < 60,000, we compute the full-kernel predictive mean using a dense factorization. For the four larger datasets, we instead use preconditioned CG in fp32, with kernel matrix–vector products evaluated using KeOps [5]. The preconditioner is constructed from the low-rank AnchorNet approximation using at most 200 landmarks.

The behavior of the fp32 solve varies substantially across datasets. It converges on Song, stagnates at a relative residual of approximately 10−210^{-2} on 3DRoad, varies strongly across splits on Buzz, and does not attain the required accuracy on House-Electric; see Table 4–Table 6.

rreff.fit [s]×\timesreflow-rank mean RMSElow-rank mean time [s]float32 CG RMSEfloat32 CG time [s]float64 CG RMSEfloat64 CG time [s]
505024.171768.7×0.1012±0.02250.1012 \pm0.02250.031.0219±0.37121.0219 \pm0.37121,5220.0551±0.00060.0551 \pm0.000612,353
1009044.45961.8×0.0708±0.00200.0708 \pm0.00200.061.4842±1.14771.4842 \pm1.14778340.0549±0.00060.0549 \pm0.000611,040
20015786.92491.8×0.0596±0.00020.0596 \pm0.00020.147.0671±1.39027.0671 \pm1.39022230.0537±0.00010.0537 \pm0.000113,651
250181108.54393.9×0.0586±0.00070.0586 \pm0.00070.173.8081±2.29433.8081 \pm2.29433360.0534±0.00060.0534 \pm0.000613,651
400252191.95222.7×0.0566±0.00040.0566 \pm0.00040.334.1274±0.27444.1274 \pm0.27442720.0526±0.00070.0526 \pm0.000717,570
500286238.73179.1×0.0560±0.00060.0560 \pm0.00060.403.9374±0.57043.9374 \pm0.57042890.0522±0.00020.0522 \pm0.000219,528
800393431.2099.1×0.0542±0.00040.0542 \pm0.00040.744.6517±1.12494.6517 \pm1.12493050.0518±0.00030.0518 \pm0.000324,752
1000449544.3778.5×0.0536±0.00030.0536 \pm0.00030.914.2526±2.41264.2526 \pm2.41262720.0516±0.00020.0516 \pm0.000226,128
16007781,022.8141.8×0.0534±0.00030.0534 \pm0.00031.731.3297±0.07911.3297 \pm0.07915350.0506±0.00030.0506 \pm0.000342,254
200010111,368.8331.2×0.0526±0.00030.0526 \pm0.00032.261.0387±0.26551.0387 \pm0.26551,8120.0490±0.00020.0490 \pm0.000256,746
300017232,459.8917.4×0.0522±0.00030.0522 \pm0.00034.130.9021±0.05150.9021 \pm0.05151,8110.0467±0.00030.0467 \pm0.000386,133

Table 6. Test RMSE and prediction time of the three predictive means on House-Electric. The reference RMSE is 0.052.

This behavior is closely related to conditioning. Since K\mathbf{K} is positive semidefinite,

κ2(K+σ2I)=λmax⁡(K)+σ2λmin⁡(K)+σ2.\kappa_2(\mathbf{K} + \sigma^2\mathbf{I}) = \frac{\lambda_{\max}(\mathbf{K}) + \sigma^2}{\lambda_{\min}(\mathbf{K}) + \sigma^2}.

When λmin⁡(K)\lambda_{\min}(\mathbf{K}) is small, reducing σ2\sigma^2 can therefore increase the condition number approximately in proportion to 1/σ21/\sigma^2. In finite precision, the attainable accuracy of an iterative solve generally deteriorates as the product of the unit roundoff and the condition number grows. Thus, the small fitted noise levels are an important contributor to the observed loss of accuracy in fp32. For example, the fitted values are approximately σ2=6×10−2\sigma^2 = 6 \times10^{-2} on Buzz and σ2=2.5×10−3\sigma^2 = 2.5 \times10^{-3} on HouseElectric.

The preconditioner can introduce an additional finite-precision difficulty. Let

rp=min⁡(r,200)r_p = \min(r, 200)

and let Wp∈Rrp×n\mathbf{W}_p \in\mathbb{R}^{r_p \times n} denote the low-rank factor constructed from the first rpr_p landmarks. We compute a thin QR factorization

\begin{bmatrix} \mathbf{W}_p^{\mathsf{T}} \\ \sigma\mathbf{I}_{r_p} \end{bmatrix} = \mathbf{Q}\mathbf{R}, \qquad\mathbf{Q}_1 = \mathbf{Q}(:, 1:n,:). \tag*{(27)}\end{bmatrix}

The preconditioner is then applied in the form

P−1r=σ−2(r−Q1Q1Tr).(28)P^{-1}\mathbf{r}=\sigma^{-2}\left(\mathbf{r}-\mathbf{Q}_1\mathbf{Q}_1^{T}\mathbf{r}\right). \tag*{(28)}

When a substantial component of r\mathbf{r} lies in the span of Q1\mathbf{Q}_1, the two vectors r\mathbf{r} and Q1Q1Tr\mathbf{Q}_1\mathbf{Q}_1^{T}\mathbf{r} can be close, so their subtraction is susceptible to cancellation in single precision. The resulting rounding error is subsequently multiplied by σ−2\sigma^{-2}. This amplification factor is approximately 16 on Buzz and 400 on HouseElectric. The observation is consistent with Figure 3, where increasing the preconditioner rank on HouseElectric can worsen the true residual.

True residual of CG for different preconditioner ranks for HouseElectric in fp32, at model rank 500 and 3000

Figure 3. True residual of CG for different preconditioner ranks for HouseElectric in fp32, at model rank 500 (left) and 3000 (right).

Repeating the full-kernel solve in fp64 removes this accuracy barrier in the cases considered here. However, the A40 has much lower fp64 than fp32 throughput, so these reference solves can require many hours. We therefore use fp64 only when needed to establish the predictive accuracy attainable by the full-kernel model. A more detailed analysis of these solves is provided in Appendix § 10.

Mixed-precision experiments

The training algorithm described above is implemented primarily in fp32. Profiling shows that approximately two-thirds of the fitting time is spent in dense matrix multiplications (GEMMs) and triangular solves (TRSMs), making these operations natural candidates for lower-precision arithmetic. We therefore investigate whether selected operations can be performed with fp16 inputs while retaining fp32 accumulation. Our experiments show that the attainable speedup is modest and, more importantly, that the resulting computation is not reliable for all datasets.

In the implementation considered here, the triangular solves associated with the Cholesky factors remain in fp32. Consequently, only four large GEMMs in each objective-and-gradient evaluation are candidates for reduced precision. One of these forms the matrix

M=Ir+WD−1WT,M = I_r + WD^{-1}W^T,

which is similar to KsmallK_{\mathrm{small}} and is the matrix factored by Cholesky. In exact arithmetic,

λmin⁡(M)≥1.\lambda_{\min}(M) \ge1.

The remaining three GEMMs appear only in the gradient and therefore do not directly affect either the objective value or the Cholesky factorization.

We consider two mixed-precision variants, both using fp16 inputs with fp32 accumulation. The first performs the GEMM forming MM and one gradient GEMM in reduced precision; we refer to this as the two-GEMM variant. The second performs all four large GEMMs in reduced precision. For the four-GEMM variant, the GEMM inputs are scaled by powers of two to avoid overflow; without this scaling, one gradient block exceeds the dynamic range of fp16 near the fitted solution and produces NaN length-scale gradients. The two-GEMM variant is useful for isolating the stability of the GEMM that forms MM, whereas the four-GEMM variant represents the largest precision reduction considered without changing the triangular solves.

Table 7 compares single fits on 3DLoad with r=3000r=3000 using fp32 and the two mixed-precision variants. All runs use split 73, the ARD kernel, and 100 Adam iterations with the same cosine learning-rate schedule. The two-GEMM variant produces nearly the same fitted hyperparameters and test RMSE as the fp32 run: 0.3261 compared with 0.3258, while reducing the fit time from 465 s to 449 s, a speedup of only 1.04×. Using all four GEMMs in reduced precision lowers the fit time to 424 s (1.10× speedup), with a test RMSE of 0.3277.

fp322 GEMMs in fp164 GEMMs in fp16
Fit time (s)465449424
Final train NLL0.23650.23730.2406
Median length scale0.039060.039060.03911
σf\sigma_f0.59300.59070.5899
σ\sigma0.23080.23080.2303
Effective rank292129192931
Test RMSE0.32580.32610.3277
Reduced-precision M used–74 of 10079 of 100

Table 7. Single fits on 3DRoad with r=3000r = 3000 using fp32 and two mixed-precision variants.

Because these results are based on a single fit for each precision configuration, the small differences in objective value and RMSE should not be interpreted as systematic accuracy degradation. At the fp32 fitted hyperparameters, however, we can compare the gradients directly. The relative gradient error is 1.65% for the two-GEMM variant and 1.35% for the four-GEMM variant. Thus, for this particular problem, reducing the precision of these GEMMs introduces only a modest perturbation to an individual gradient evaluation but provides correspondingly modest performance gains.

The limited speedup is consistent with the operation mix of the algorithm. The triangular solves, kernel evaluations, and several other operations remain in fp32. For the two-GEMM variant, the speedup of a single objective-and-gradient evaluation grows with rank, from no measurable improvement at r=250r = 250 to approximately 1.13× at r=6000r = 6000 on 3DRoad; see Table 8. At r=3000r = 3000, using all four GEMMs in reduced precision gives an evaluation speedup of approximately 1.15×. At the level of a complete fit, however, the observed speedup remains close to 1.1×.

Datasetrrκ2(M^)\kappa_2(\widehat{\mathbf{M}})FactorizationRel. gradient errorTime fp32 / mixed (s)
3DRoad2501.3⋅1041.3 \cdot10^4succeeds0.0290.16 / 0.17
3DRoad10002.1⋅1042.1 \cdot10^4succeeds0.0340.95 / 0.91
3DRoad30002.5⋅1042.5 \cdot10^4succeeds0.0164.58 / 4.20
3DRoad60002.5⋅1042.5 \cdot10^4succeeds0.01213.6 / 12.0
Buzz50 to 10002.4⋅1062.4 \cdot10^6 to 5.4⋅1065.4 \cdot10^6fails (7/8)4.2–
HouseElectric50 to 30001.3⋅1061.3 \cdot10^6 to 1.2⋅1071.2 \cdot10^7fails (9/11)3.0 and 29.6–

Table 8. Single objective-and-gradient evaluations with the two-GEMM mixed-precision variant. The condition number is evaluated from the fp32 baseline matrix.

Stability of the reduced-precision factorization

The main numerical difficulty is the GEMM used to form M. To distinguish exact-arithmetic properties from finite-precision effects, let

M=Ir+WD−1WTM = I_r + W D^{-1} W^T

denote the exact matrix, let M~\widetilde{M} denote the matrix obtained using the fp32 baseline computation, and let M~\widetilde{M} denote the corresponding matrix when the selected GEMM is evaluated with fp16 inputs and fp32 accumulation.

Although λmin⁡(M)≥1\lambda_{\min}(M) \ge1 in exact arithmetic, rounding in the matrix products can perturb the computed matrices. We write

M^=M+ΔM.\widehat{M} = M + \Delta M.

As a rough error model, the norm of ΔM\Delta M can grow proportionally to the fp16 unit roundoff and the scale of the matrix product. By Weyl’s inequality, positive definiteness of the computed matrix is guaranteed whenever

∥ΔM∥2<λmin⁡(M).\lVert\Delta M\rVert_2 < \lambda_{\min}(M).

This suggests the heuristic condition

ufp16 κ2(M)≲1,u_{\mathrm{fp16}} \,\kappa_2(M) \lesssim1,

although the actual behavior also depends on the structure and scaling of the matrix multiplication.

Table 8 reports one mixed-precision objective-and-gradient evaluation per rank at hyperparameters obtained from the fp32 fit on split 73. On 3DRoad, the reduced-precision factorization succeeds at every tested rank between 50 and 6000, even though ufp16κ2(M)≈12u_{\mathrm{fp16}}\kappa_2(\mathbf{M}) \approx12 at r=3000r = 3000. Thus, the heuristic condition is conservative for this problem.

The behavior is substantially worse on Buzz and HouseElectric. Their estimated condition numbers lie between approximately 1.3⋅1061.3 \cdot10^6 and 1.2⋅1071.2 \cdot10^7. The reduced-precision factorization fails for 7 of the 8 tested ranks on Buzz and for 9 of the 11 tested ranks on HouseElectric. At the few ranks where the factorization succeeds, the relative gradient errors are already large. We therefore do not attempt complete reduced-precision fits on these two datasets.

These experiments also show that conditioning at the final fitted hyperparameters does not characterize the entire optimization trajectory. Table 9 follows the 3DRoad two-GEMM fit through selected Adam iterations. The exact-arithmetic matrix M\mathbf{M} remains positive definite by construction, but the computed matrices can lose this property because of rounding.

Iterationσf\sigma_fσ\sigmaλmax⁡(Mf)\lambda_{\max}(M_f)λmin⁡(Mf)\lambda_{\min}(M_f)λmin⁡(Mc)\lambda_{\min}(M_c)Factorization
01.001.4141.2⋅1051.2 \cdot10^51.0000.997succeeds
72.070.6111.5⋅1061.5 \cdot10^61.0000.975succeeds
8 to 102.25 to 2.610.55 to 0.571.7⋅1061.7 \cdot10^6 to 1.8⋅1061.8 \cdot10^60.999Infoverflow
163.640.6345.9⋅1055.9 \cdot10^50.9360.222succeeds
18 to 404.07 to 1.440.60 to 0.302.2⋅1052.2 \cdot10^5 to 1.2⋅1061.2 \cdot10^60.87 to 0.99−4.2-4.2 to −0.15-0.15fails
441.010.2918.6⋅1048.6 \cdot10^40.9970.535succeeds
Final0.590.2312.5⋅1042.5 \cdot10^41.00–succeeds

Table 9. Computed spectra along the 3DRoad mixed-precision fit. MfM_f denotes the fp32 baseline matrix and McM_c the matrix obtained from the reduced-precision GEMM.

Between iterations 18 and 40, the signal-to-noise ratio σf2/σ2\sigma_f^2/\sigma^2 ranges from approximately 25 to 120, compared with 6.6 near the final iterate. Over this part of the trajectory, the largest eigenvalue of the fp32 baseline matrix grows to between 2.2⋅1052.2 \cdot10^5 and 1.2⋅1061.2 \cdot10^6, while the reduced-precision matrix M^\widehat{\mathbf{M}} becomes indefinite, with its smallest eigenvalue between −4.2-4.2 and −0.15-0.15. At iterations 8–10, the reduced-precision GEMM instead overflows.

Because the reduced-precision matrix still admits a Cholesky factorization at iteration 7, despite a large spectral range, the condition number is best interpreted as an indicator of numerical risk rather than a deterministic breakdown criterion.

To allow the mixed-precision fit to complete, we recompute the GEMM forming M\mathbf{M} in fp32 whenever the Cholesky factorization of M^\widehat{\mathbf{M}} fails. Without this fallback, the 3DRoad fit terminates at iteration 8. With the fallback, the reduced-precision matrix is accepted in 74 of the 100 Adam iterations for the two-GEMM variant and in 79 of 100 iterations for the four-GEMM variant. The power-of-two scaling used in the latter also prevents the observed overflow. The fallback, however, detects only factorization failure; it does not guarantee that a successful reduced-precision factorization produces sufficiently accurate gradients. Instead of recomputing M^\widehat{\mathbf{M}} in fp32, the failed factorization can be repaired by raising the eigenvalues of M^\widehat{\mathbf{M}} below 1 to 1, which uses λmin⁡(M)≥1\lambda_{\min}(\mathbf{M}) \ge1. On 3DRoad, this keeps all four GEMMs in fp16 in every iteration and gives a test RMSE of 0.3259, compared with 0.3258 in fp32. Such a modified factorization is therefore a possible fix when the loss of positive definiteness is the primary failure, but not on Buzz and HouseElectric, where the relative gradient error remains between 5 and 650 after the repair.

We also tested TF32, which retains a larger dynamic range but has reduced mantissa precision, and observed behavior qualitatively similar to fp16. This indicates that overflow is not the only limitation: even when scaling prevents overflow, insufficient precision can still perturb M enough to compromise its factorization or substantially degrade the gradient.

Buzz and HouseElectric are also the datasets for which K11K_{11} is already close to numerically rank deficient in fp32, with observed numerical ranks of 443 out of 1000 and 1724 out of 3000, respectively. This is consistent with their greater sensitivity to further precision reduction.

Overall, these experiments show that selected GEMMs in the AnchorNet–FITC training procedure can be executed below fp32 for moderately conditioned problems, but the resulting performance gain is small and the approach is not robust for ill-conditioned cases. In particular, a Cholesky-based training method requires the computed reduced matrix to remain positive definite, which cannot in general be guaranteed after aggressive precision reduction. This observation is consistent with the use of iterative methods rather than Cholesky factorization for half-precision GP training in [15].

Conclusion

We developed a scalable Gaussian process training approach that combines AnchorNet landmark selection with an FITC correction and evaluates the resulting objective and analytic gradients using low-rank factors and small dense matrices. Across nine regression datasets, using the same kernel, initialization, optimizer, and optimization budget while varying the approximation rank, ANGP reaches the reference predictive accuracy on all nine datasets. The corresponding fitting speedups range from 3.4×3.4\times to 1120×1120\times, although these timings are obtained on an A40 GPU with roughly a 2.4×2.4\times single-precision throughput advantage over the V100 used for the reference results.

Our experiments also highlight an important distinction between training and prediction in reduced precision. The low-rank training procedure can be carried out effectively in single precision, whereas the full-kernel predictive solve becomes numerically challenging on 3DRoad, Buzz, and HouseElectric. On these datasets, the fp32 iterative solve does not attain sufficient accuracy and an fp64 solve is required to establish the full-kernel reference prediction. The low-rank predictive mean avoids this solve entirely and requires only a few seconds. It reaches the reference accuracy on Elevators, Bike, and Kin40K, and on HouseElectric it reaches the reference within its reported standard deviation while keeping the complete training-and-prediction pipeline in single precision.

We further investigated whether selected dense matrix operations in the training procedure can be reduced below fp32. Using fp16 or TF32 for selected GEMMs provides only modest additional speedups and can become unreliable when the reduced system is ill-conditioned. In particular, rounding can destroy the positive definiteness required by the Cholesky factorization even though the corresponding matrix is positive definite in exact arithmetic. These results suggest that fp32 provides a useful balance between performance and numerical robustness for the proposed training procedure, while more aggressive precision reduction requires additional stabilization.

The low-rank predictive mean is therefore an attractive first choice when prediction cost is important. When its accuracy is insufficient, the full-kernel predictive mean can be computed as a second stage, using higher precision when required by the conditioning of the problem. This suggests a practical adaptive strategy in which the low-rank prediction is computed first and its accuracy is assessed on validation data before deciding whether the more expensive full-kernel solve is necessary. Developing a reliable criterion for this decision, together with automatic selection of the approximation rank, is an interesting direction for future work.

Acknowledgments. The authors used Claude to assist with coding, including parameter initialization, parameter sweeps, and plotting and table formatting, and ChatGPT to assist with manuscript editing and revision. All AI-assisted code and manuscript content were checked and verified by the authors. The authors assume responsibility for all content.

Experimental setup

Initialization. All datasets use the same initialization. The length scales start at the median pairwise distance of a random subsample SS of 2,000 training inputs,

ℓ0=median⁡{∥xi−xj∥2:i<j, xi,xj∈S},(29)\ell_{0} = \operatorname{median}\left\{\lVert x_{i} - x_{j} \rVert_{2} : i < j,\ x_{i}, x_{j} \in S\right\}, \tag*{(29)}

with the same value for all dd length scales. The signal scale starts at σf,0=std⁡(y)=1\sigma_{f,0} = \operatorname{std}(y) = 1, since the targets are standardized, and the noise at the mode of the Gamma⁡(1.1,0.05)\operatorname{Gamma}(1.1, 0.05) prior on σ2\sigma^2, i.e. σ0=2≈1.414\sigma_0 = \sqrt{2} \approx1.414. A large initial noise keeps KsmallK_{\mathrm{small}} well conditioned at the start; a small one multiplies the rounding errors by σ−2\sigma^{-2} and was the main cause of failed fits in single precision. Each split is a single run from this starting point, without restarts.

Optimization. We run Adam for T=100T = 100 iterations with gradients clipped to norm 10 and the learning rate

ηt=η0(0.02+0.98⋅12(1+cos⁡(πtT−1)))(30)\eta_t = \eta_0\left(0.02 + 0.98 \cdot\frac{1}{2}\left(1 + \cos\left(\frac{\pi t}{T - 1}\right)\right)\right) \tag*{(30)}

and η0=0.2\eta_0 = 0.2.

The decay allows one learning rate for all datasets: 3DRoad has to move 4.1 in log⁡ℓ\log\ell from its starting point, PoleTele only 0.12. There is no validation split and no early stopping; we report the final iterate. In none of the fits reported here did a length scale collapse or the noise reach its floor.

Gradient path. The blocked and unblocked gradient implementations agree to 3×10−53 \times10^{-5} in test RMSE. The unblocked path holds several r×nr \times n arrays in memory, while the blocked path processes the data in blocks of 20,000 rows. We therefore use the blocked path by default and switch to the unblocked one only if a single timed evaluation shows it to be at least 20% faster (§ 6). This margin does not guarantee that the unblocked path fits in memory; in the runs reported here it did.

The predictive-mean solve

The exact predictive mean requires the solution of

K~α=y(31)\widetilde{\mathbf K}\boldsymbol{\alpha} = \mathbf y \tag*{(31)}

at the full training size. We use preconditioned CG with KeOps matrix–vector products, at most 200 iterations, and an AnchorNet preconditioner built from the first min⁡(r,200)\min(r, 200) landmarks of the model. Figure 2 shows a fp32 solve on Buzz at rank 250 over 500 iterations. Up to about iteration 250, the CG recurrence residual, the preconditioned residual norm, and the true residual ∥y−K~αk∥/∥y∥\lVert\mathbf y - \widetilde{\mathbf K}\boldsymbol{\alpha}_k\rVert/\lVert\mathbf y\rVert agree. After that, the first two keep decreasing to 3.9×10−53.9 \times10^{-5} and 4.4×10−74.4 \times10^{-7}, while the true residual stays at 3.0×10−23.0 \times10^{-2}. A stopping criterion based on either cheap quantity would therefore report convergence that has not happened. The preconditioned norm is particularly misleading, since P−1P^{-1} removes the residual component in the span of the landmarks, which matters most for prediction. The test RMSE is flat from iteration 200 onward (0.2659 at iteration 200, 0.2675 at 500), so more iterations do not help: the remaining error is a floor set by single precision.

Residuals and test RMSE of fp32 CG with a rank-200 AnchorNet preconditioner for Buzz

Figure 2. Residuals (top) and test RMSE (bottom) of fp32 CG with a rank-200 AnchorNet preconditioner for Buzz with rank 250.

On HouseElectric, the fp32 solve does not converge for any preconditioner rank in {0,5,50,200,500,1000}\{0, 5, 50, 200, 500, 1000\} below the model rank, at r=500r = 500 and r=3000r = 3000 (Figure 3). The preconditioner here is built from the landmarks of the model, so its rank is at most the model rank. The true residual worsens as the preconditioner rank grows, and no preconditioner is the best setting. This is consistent with the factor 1/σ2≈4001/\sigma^2 \approx400 in P−1P^{-1}, which amplifies its own rounding error.

In fp64, with a rank-1000 preconditioner, the solve converges on both datasets (Figure 4). On HouseElectric, the true residual reaches 6×10−86 \times10^{-8} at rank 500. These solves provide the reference values in Table 4–Table 6. On Buzz, the fp32 solve at rank 250 gives 0.2723, only 2.4% above the fp64 value of 0.2659, while the low-rank mean gives 0.3619. There, the loss from single precision is much smaller than the loss from the low-rank predictive mean. On HouseElectric, the opposite holds: fp32 fails, and the low-rank mean is within 0.002 to 0.006 of the fp64 result from rank 200 upward.

True residual and test RMSE of the fp64 solves on HouseElectric and Buzz

Figure 4. True residual (top) and test RMSE (bottom) of the fp64 solves on HouseElectric (left) and Buzz (right). Dotted lines show the low-rank predictive mean at the same rank, in the same color.

References

  1. [1]N. Alder and R. Herbrich, Energy-efficient Gaussian processes using low-precision arithmetic, in Forty-first International Conference on Machine Learning, 2024.
  2. [2]C. M. Bishop and N. M. Nasrabadi, Pattern recognition and machine learning, vol. 4, Springer, 2006.
  3. [3]D. Cai, E. Chow, and Y. Xi, Posterior covariance structures in gaussian processes, SIAM Journal on Matrix Analysis and Applications, 46 (2025), pp. 1640–1673, https://doi.org/10.1137/24M1684918.
  4. [4]D. Cai, J. Nagy, and Y. Xi, Fast deterministic approximation of symmetric indefinite kernel matrices with high dimensional datasets, SIAM J. Matrix Anal. Appl., 43 (2022), pp. 1003–1028, https://doi.org/10.1137/21M142627.
  5. [5]B. Charlier, J. Feydy, J. A. Glaunès, F.-D. Collin, and G. Durif, Kernel operations on the GPU, with Autodiff, without memory overflows, J. Mach. Learn. Res., 22 (2021), pp. 1–6.
  6. [6]Y. Chen, E. N. Epperly, J. A. Tropp, and R. J. Webber, Randomly pivoted Cholesky: Practical approximation of a kernel matrix with few entry evaluations, Comm. Pure Appl. Math., 78 (2025), pp. 995–1041.arxiv.org/abs/2207.06503
  7. [7]A. Cortinovis and D. Kressner, On randomized trace estimates for indefinite matrices with an application to determinants, Found. Comput. Math., 22 (2022), pp. 875–903, https://doi.org/10.1007/s10208-021-09525-9.
  8. [8]K. Dong, D. Eriksson, H. Nickisch, D. Bindel, and A. G. Wilson, Scalable log determinants for Gaussian process kernel learning, Adv. Neural Inf. Process. Syst., 30 (2017).arxiv.org/abs/1711.03481
  9. [9]K. Dong, D. Eriksson, H. Nickisch, D. Bindel, and A. G. Wilson, Scalable log determinants for Gaussian process kernel learning, in Proceedings of the 31st International Conference on Neural Information Processing Systems, 2017, p. 6330–6340, https://dl.acm.org/doi/pdf/10.5555/3295222.3295380.arxiv.org/abs/1711.03481
  10. [10]E. N. Epperly, J. A. Tropp, and R. J. Webber, Embrace rejection: Kernel matrix approximation by accelerated randomly pivoted Cholesky, SIAM J. Matrix Anal. Appl., 46 (2025), pp. 2527–2557.arxiv.org/abs/2410.03969
  11. [11]J. Gardner, G. Pleiss, K. Q. Weinberger, D. Bindel, and A. G. Wilson, GPyTorch: Blackbox matrix-matrix Gaussian process inference with GPU acceleration, in Proceedings of the 32nd International Conference on Neural Information Processing Systems, 2018, pp. 7587–7597, https://dl.acm.org/doi/pdf/10.5555/3327757.3327857.
  12. [12]M. B. Giles, Collected matrix derivative results for forward and reverse mode algorithmic differentiation, in Advances in Automatic Differentiation, Springer, 2008, pp. 35–44.DOI
  13. [13]R. B. Gramacy, Surrogates: Gaussian process modeling, design, and optimization for the applied sciences, Chapman and Hall/CRC, 2020.DOI
  14. [14]H. Huang, T. Xu, Y. Xi, and E. Chow, Hogp: A high-performance python package for gaussian process, Journal of Open Source Software, 11 (2026), p. 8621, https://doi.org/10.21105/joss.08621, https://joss.theoj.org/papers/10.21105/joss.08621.
  15. [15]W. J. Maddox, A. Pattonski, and A. G. Wilson, Low-precision arithmetic for fast Gaussian processes, in Uncertainty in Artificial Intelligence, PMLR, 2022, pp. 1306–1316.arxiv.org/abs/2207.06856
  16. [16]R. A. Meyer, C. Musco, C. Musco, and D. P. Woodruff, Hutch++: Optimal stochastic trace estimation, in Symposium on Simplicity in Algorithms, 2021, pp. 142–155, https://doi.org/10.1137/1.9781611976496.16.
  17. [17]D. Persson, A. Cortinovis, and D. Kressner, Improved variants of the Hutch++ algorithm for trace estimation, SIAM J. Matrix Anal. Appl., 43 (2022), pp. 1162–1185, https://doi.org/10.1137/21M1447623.
  18. [18]J. Quinonero-Candela and C. E. Rasmussen, A unifying view of sparse approximate Gaussian process regression, J. Mach. Learn. Res., 6 (2005), pp. 1939–1959.
  19. [19]C. Rasmussen and C. Williams, Gaussian Processes for Machine Learning, Adaptive Computation and Machine Learning, The MIT Press, 2006, https://doi.org/10.7551/mitpress/3206.001.0001.
  20. [20]B. Schölkopf and A. J. Smola, Learning with kernels: support vector machines, regularization, optimization, and beyond, The MIT Press, 2001.
  21. [21]B. W. Silverman, Some aspects of the spline smoothing approach to non-parametric regression curve fitting, J. R. Stat. Soc. Ser. B Methodol., 47 (1985), pp. 1–21, https://doi.org/10.1111/j.2517-6161.1985.tb01327.x.
  22. [22]M. Skorski, A modern analysis of Hutchinson’s trace estimator, arXiv:2012.12895, (2020), https://doi.org/10.48550/arXiv.2012.12895.
  23. [23]E. Snelson and Z. Ghahramani, Sparse Gaussian processes using pseudo-inputs, in Proceedings of the 18th International Conference on Neural Information Processing Systems, 2005, p. 1257–1264.
  24. [24]S. Ubaru, J. Chen, and Y. Saad, Fast estimation of tr(f(A)) via stochastic Lanczos quadrature, SIAM J. Matrix Anal. Appl., 38 (2017), pp. 1075–1099, https://doi.org/10.1137/16M1104974.
  25. [25]T. Wagner, F. Nestler, T. Xu, Y. Xi, and M. Stoll, Preconditioned additive Gaussian processes with Fourier acceleration, Found. Data Sci., 12 (2027), pp. 131–157, https://doi.org/10.3934/fods.2026013.
  26. [26]J. Wenger, G. Pleiss, P. Hennig, J. Cunningham, and J. Gardner, Preconditioning for scalable Gaussian process hyperparameter optimization, in International Conference on Machine Learning, PMLR, 2022, pp. 23751–23780.arxiv.org/abs/2107.00243
  27. [27]C. K. Williams and C. E. Rasmussen, Gaussian processes for machine learning, vol. 2, The MIT Press, 2006.DOI
  28. [28]A. Wilson and H. Nickisch, Kernel interpolation for scalable structured Gaussian processes (KISS-GP), in Proceedings of the 32nd International Conference on International Conference on Machine Learning - Volume 37, 2015, pp. 1775–1784, https://dl.acm.org/doi/10.5555/3045118.3045307.arxiv.org/abs/1503.01057
  29. [29]K. Wu, J. Wenger, H. T. Jones, G. Pleiss, and J. Gardner, Large-scale Gaussian processes via alternating projection, in International Conference on Artificial Intelligence and Statistics, PMLR, 2024, pp. 2620–2628.arxiv.org/abs/2310.17137
  30. [30]T. Xu, D. Cai, H. Huang, E. Chow, and Y. Xi, Preconditioned truncated single-sample estimators for scalable stochastic optimization, 2026, https://arxiv.org/abs/2510.24587.
  31. [31]S. Zhao, T. Xu, H. Huang, E. Chow, and Y. Xi, An adaptive factorized Nyström preconditioner for regularized kernel matrices, SIAM Journal on Scientific Computing, 46 (2024), pp. A2351–A2376, https://doi.org/10.1137/23M1565139.

Paper details

vlnc:W_6thzt3f5
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Claude, ChatGPT

View statement
Looking for related papers…
Contents