Intrinsic Gaussian Vector Fields On Manifolds
Intrinsic Gaussian Vector Fields On Manifolds
Intrinsic Gaussian Vector Fields On Manifolds
Abstract ing spatial data (Chilès and Delfiner, 2012) and au-
tomated decision-making, e.g., optimization (Noskova
Various applications ranging from robotics to and Borovitskiy, 2023; Shields et al., 2021; Snoek et
climate science require modeling signals on al., 2012), or sensor placement (Krause et al., 2008).
non-Euclidean domains, such as the sphere. Gaussian processes can be scalar- or vector-valued (Al-
Gaussian process models on manifolds have varez et al., 2012). The important special case of the
recently been proposed for such tasks, in latter is Gaussian vector fields. These can for exam-
particular when uncertainty quantification is ple be used to model velocities or accelerations, either
needed. In the manifold setting, vector- as a target in itself or as means for exploring an un-
valued signals can behave very differently known dynamical system. When the input domain
from scalar-valued ones, with much of the is Euclidean, vector fields are just vector-valued func-
progress so far focused on modeling the lat- tions. However, when the domain is a submanifold of
ter. The former, however, are crucial for a Euclidean space, such as when modeling wind speeds
many applications, such as modeling wind or ocean currents on the surface of Earth, the situation
speeds or force fields of unknown dynami- can be quite different: geometry places additional con-
cal systems. In this paper, we propose novel straints on vector fields that need to be accounted for.
Gaussian process models for vector-valued As illustrated in Figure 1, the values of a vector field
signals on manifolds that are intrinsically de- ought to be tangential to the manifold, while those of
fined and account for the geometry of the a vector function can be arbitrary vectors.
space in consideration. We provide compu-
tational primitives needed to deploy the re- In recent years, two different formalisms were pro-
sulting Hodge–Matérn Gaussian vector fields posed for defining Gaussian vector fields on manifolds.
on the two-dimensional sphere and the hyper- Lange-Hegermann (2018) approached the problem by
tori. Further, we highlight two generalization considering linear constraints on vector-valued Gaus-
directions: discrete two-dimensional meshes sian processes, constraining them to lie in the tangent
and “ideal” manifolds like hyperspheres, Lie space of a submanifold of Rd . Additionally, one can
groups, and homogeneous spaces. Finally, we
show that our Gaussian vector fields consti-
tute considerably more refined inductive bi-
ases than the extrinsic fields proposed before.
1 INTRODUCTION
impose further linear constraints to the resulting fields, 1.1 Gaussian Processes
such as making them divergence-free. Hutchinson et
al. (2021)—which is closer in spirit to this work— Let X be a set. A random function f on X is called
instead considered projecting vector-valued Gaussian a Gaussian process (GP) with mean µ : X → R and
processes to a submanifold of Rd . While both these covariance (or kernel) k : X × X → R—denoted by
procedures can in principle produce any valid Gaus- f ∼ GP(µ, k)—if for any finite set of points x in X
sian vector field, they are fundamentally extrinsic and we have f (x) ∼ N(µ(x), Kxx ) where K• •′ = k(•, •′ ).
we will show that the fields one gets in practice intro- Without loss of generality, we usually assume µ(·) = 0.
duce undesirable inductive biases. Assuming a GP prior f ∼ GP(0, k) and a Gaussian
To remedy this, we propose a new approach: fully likelihood y | f = N (y | f (x), σε2 ) with a fixed noise
intrinsic Gaussian vector fields based on the Hodge variance σε2 , the posterior f | y is a GP (Rasmussen
Laplacian 1 that we name Hodge–Matérn Gaussian vec- and Williams, 2006) with mean and covariance
tor fields. For some simple manifolds, namely for the −1
two-dimensional sphere S2 and for the hypertori Td , we µf |y (·) = K·x Kxx + σε2 I y, (1)
−1
develop computational techniques that allow effortless kf |y (·, ·′ ) = k(·, ·′ ) − K·x Kxx + σε2 I
Kx·′ . (2)
use of these intrinsic fields in downstream applications.
Here, the function µf |y can be used to draw predictions
The aforementioned computational techniques hinge
and the function kf |y is used to quantify uncertainty.
on knowing the eigenvalues and eigenfields of the
Hodge Laplacian. To this end, we describe how to: When X = Rd , Matérn Gaussian processes (Ras-
mussen and Williams, 2006; Stein, 1999) are most
a) derive these from the eigenpairs of the Laplace– often used. Their respective kernels kν,κ,σ2 are the
Beltrami operator when the manifold is two- three-parameter family of Matérn kernels, whose
dimensional, using automatic differentiation only; limiting case k∞,κ,σ2 for ν → ∞ is known as the
heat (a.k.a. squared exponential, RBF, Gaussian,
diffusion) kernel, which is arguably the most popular.
b) compute these on product manifolds in terms of
the eigenvalues and eigenfields on the factors; and
1.2 Gaussian Processes on Manifolds
c) get these explicitly in the cases of the circle S1 , Now consider X = M, where M is a compact Rieman-
the hypertori Td , and the sphere S2 . nian manifold. Throughout this paper, manifolds are
always assumed to be connected. Using the SPDE-
We conjecture that (a) can also be used to define Gaus- based characterization of Matérn Gaussian processes
sian vector fields on meshes, by changing the analytic of Lindgren et al. (2011) and Whittle (1963), Borovit-
notions into their appropriate discretizations. Further- skiy et al. (2020) showed how to compute Matérn ker-
more, we conjecture that (c) can be done for more gen- nels on M in terms of the spectrum of the Laplace–
eral parallelizable manifolds, e.g. on Lie groups, which Beltrami operator ∆:
may then facilitate the generalization to the very gen- ∞
σ2 X
eral class of homogeneous spaces. The latter includes kν,κ,σ2 (x, x′ ) = Φν,κ (λn )fn (x)fn (x′ ), (3)
many manifolds of interest which are poorly amenable Cν,κ n=0
to discretization because of their higher dimension. ∞
where {fn }n=0 is an orthonormal basis of eigenfunc-
By showing how our intrinsic Hodge–Matérn Gaussian tions of ∆ such that ∆fn = −λn fn ,
vector fields improve over their naı̈ve extrinsic coun- ( −ν−d/2
2ν
terparts on the two-dimensional sphere, we hope to κ2 + λ ν < ∞,
motivate further research. First, into the development Φν,κ (λ) = 2 (4)
− κ2 λ
e ν = ∞,
of practical intrinsic Gaussian vector fields on other
domains. Second, into applying the proposed models d = dim(M), and
R Cν,κ is a normalization constant that
in areas like climate/weather modeling and robotics. ensures vol1M M kν,κ,σ2 (x, x) dx = σ 2 . There exist an-
alytical (Azangulov et al., 2022; 2023) and numerical
1
The concurrent work by Peach et al. (2024) studies (Borovitskiy et al., 2020; Coveney et al., 2020) tech-
Gaussian vector fields induced by the connection Lapla- niques for computing the eigenpairs λn , fn , or bypass-
cian rather than the Hodge Laplacian (see Appendix A.7 ing the computation thereof. In the end, a truncated
on the difference between these two notions of Laplacian).
Importantly, they consider a very different setting: a pri- series from Equation (3) yields tractable Gaussian pro-
ori unknown manifolds that are estimated from finite data, cesses that respect the intrinsic geometry of the man-
showcasing an impressive neuroscience application. ifold (Rosa et al., 2023), as illustrated in Figure 2.
D. Robert-Nicoud, A. Krause, V. Borovitskiy
(b) Extrinsic kernel In order to define Gaussian vector fields that can be
used in practice, Hutchinson et al. (2021) introduced
Figure 2: Comparing an intrinsic Matérn kernel (ν = the notion of projected Gaussian processes. These
∞) of Equation (3) to an extrinsic one, the restriction are constructed by picking an isometric embedding
of a Euclidean Matérn kernel to the manifold. Note ϕ : M → RD into some Euclidean space.2 Then, a
the latter induces high correlation between the points tangent space Tx M can be identified with a subspace
across the minor axis of the ellipse, despite them being of Tϕ(x) RD ∼
= RD . Thus, there exists a projection Px
D
far from each other in terms of the intrinsic distance. from R to this subspace and f (x) = Px g(x) defines a
valid Gaussian vector field for any vector-valued GP g.
(a) Projected Matérn (b) Projected Matérn, rotated (c) Hodge–Matérn (d) Hodge–Matérn, rotated
Figure 3: GP regression from a single observation (the red vector) for a very large length scale κ. Black vectors
represent the prediction µf |y (·), the background color shows the uncertainty kf |y (·, ·) : yellow for high, blue for
low. On (a) and (b) we use a projected Matérn GP of Hutchinson et al. (2021); on (c) and (d) we use our Hodge–
Matérn Gaussian vector field, ν = ∞ in both cases. Unnaturally, uncertainty in (a) and (b) is non-monotonous
with respect to the distance on the sphere: it is considerably lower at the antipode than halfway to it.
certain types of vector field data (Berlinghieri et al., Having found a suitable adaptation of the heat kernel
2023), as we will clearly observe in Section 5. to the vector field case, we turn to making it explicit.
Similarly to the scalar case, a Hilbert space L2 (M; T M)
To overcome these challenges, we introduce a fully in-
of square integrable vector fields can be defined. There
trinsic class of Gaussian vector fields on manifolds. ∞
is always an orthonormal basis {sn }n=0 of L2 (M; T M)
such that ∆sn = −λn sn , i.e. sn are eigenfields of the
3 INTRINSIC GAUSSIAN VECTOR FIELDS
Hodge Laplacian. Moreover, we have λn ≥ 0 for n ≥ 0.
The kernel P t can be computed in terms of the eigen-
In this section, we present the main ideas behind the
fields just like its scalar counterpart (cf. (3)). Namely,
construction of the intrinsic Gaussian vector fields we
as discussed in Appendix A.4, we have
propose. The mathematical formalism for this section
is detailed in Appendix A. ∞
X
P t (x, x′ ) = e−tλn sn (x) ⊗ sn (x′ ), (11)
3.1 The Hodge Heat Kernel n=0
In practice, the series in Equation (11) should be trun- rameters, gives a more flexible family of kernels:
cated, with only a few terms corresponding to the
smallest eigenvalues λn used to approximately com- σ12 kdiv 2 curl 2 harm
ν,κ1 ,1 + σ2 kν,κ2 ,1 + σ3 kν,κ3 ,1 . (17)
pute the kernel, just like in the scalar case. Notice that
the functions Φν,κ (λ) are all decreasing in λ, so that By analogy with the concurrent paper by Yang et
the most significant terms are the ones corresponding al. (2024), we call this family Hodge-compositional
to the smallest eigenvalues. Matérn kernels. Unless prior knowledge suggests a
more specialized choice, this is the family we recom-
3.3 Divergence-Free and Curl-Free Kernels mend for use in practical applications, inferring all the
hyperparameters κi , σi from data. Our experimental
The celebrated Helmholtz decomposition (also known results in Section 5 support this recommendation.
as the fundamental theorem of vector calculus) states
that any vector field in Rd decomposes into the sum 3.5 Gaussian Process Regression
of its divergence-free and curl-free parts. The former,
intuitively, has no sinks and sources; the latter has All kernels defined above fall under the umbrella
no vortexes. Many vector fields in physics are known framework of Gaussian vector fields described in
to have only one of these parts. This suggests that Hutchinson et al. (2021). In practice, there are two
divergence-free and curl-free Gaussian vector fields can ways to perform Gaussian process regression in this
be useful inductive biases (Berlinghieri et al., 2023). setting. The first one—if the manifold is embedded
For manifolds, the analog of Helmholtz decomposition in RD —is to treat Gaussian vector fields as special
is the Hodge decomposition, see e.g. Rosenberg (1997, cases of Gaussian vector functions in RD . The second
Theorem 1.37). It states that any vector field u on M is to introduce a frame, i.e. a coordinate choice (not
can be represented as a sum of three fields: necessarily smooth) in all of the tangent spaces, and
describe all quantities in these coordinates. Depend-
u = u1 + u2 + u3 , (15) ing on whether an embedding or a frame is available,
either can be used. Both are merely modes of compu-
where u1 = ∇f1 for some function f1 , and thus is pure tation, not affecting the inductive biases of Gaussian
divergence—meaning that div u1 ̸= 0 and curl u1 = vector fields and not introducing any error per se.
0—and in particular curl-free, u2 = ⋆∇f2 , and thus
is pure curl and divergence-free, and u3 is a harmonic 3.6 Kernel Evaluation and Sampling
form, ∆u3 = 0, both curl- and divergence-free. The
symbol ⋆ denotes the Hodge star operator, which we As mentioned in the end of Section 3.2, given that
recall in Appendix A.1. For intuition on divergence eigenfields and eigenvalues are known, we can approx-
and curl, see Appendix A.2. imately evaluate the kernels by truncating the series
Importantly, the orthonormal basis of eigenfields in Equation (14). Such a truncation is a well-defined
∞
{sn }n=0 may be chosen in such a way that each sn kernel, i.e. it corresponds to a Gaussian vector field.
is in exactly one of the three classes above. Let Idiv , Proposition 6. The Gaussian vector field
Icurl , and Iharm denote the index sets of the respective
classes of eigenfields. Using a single class, we can de- L
σ X
q
fine versions of Matérn Gaussian vector fields on the f (x) = p wn Φν,κ (λn )sn (x), (18)
Cν,κ n=0
manifold M with the associated inductive bias.
Theorem 5. There exists a Gaussian vector field f • — iid
where wn ∼ N (0, 1), corresponds to the kernel P
given
where • ∈ {div, curl, harm} —with kernel ∞
by the truncation of Equation (14) with the sum n=0
PL
σ2 X therein substituted by the sum n=0 .
kν,κ,σ2 (x, x′ ) = Φν,κ (λn )sn (x)⊗sn (x′ ). (16)
•
Cν,κ
n∈I• Importantly, Equation (18) allows to approximately
curl div sample Hodge–Matérn Gaussian vector fields in an ex-
What is more, div f = 0, curl f = 0 and
tremely computationally efficient way by simply draw-
∆f harm = 0 almost surely as long as f • is smooth
ing random wn ∼ N (0, 1). Efficiently sampling their
enough that div and curl are well-defined.
respective posteriors can be performed using pathwise
conditioning for Gaussian vector fields, as described in
3.4 Hodge-compositional Matérn Kernels Hutchinson et al. (2021).
Combining the pure divergence, pure curl, and har- Remark 7. Of course, direct analogs of Equation (18)
monic kernels, each with a separate set of hyperpa- also hold for the kernels of Theorem 5.
Intrinsic Gaussian Vector Fields on Manifolds
(a) Eigenfield √1 ∇Y3,2 (b) Eigenfield √1 ⋆ ∇Y3,2 (c) Eigenfield √1 ∇Y7,3 (d) Eigenfield √1 ⋆ ∇Y7,3
λ3 λ3 λ7 λ7
Figure 4: Eigenfunctions on the sphere S2 (represented by color) and the respective eigenfields.
In summary, (approximate) sampling, kernel evalua- Theorem 8. For each n ≥ 1, both ∇fn and ⋆∇fn are
tion and differentiation reduce to knowing eigenvalues eigenfields of the Hodge Laplacian, and the set
and eigenfields of the Hodge Laplacian on M . Thus,
we proceed to discuss how to obtain those in practice.
∇f ⋆∇f
√ n , √ n , gj n ≥ 1 and 0 ≤ j ≤ J (19)
λn λn
4 EXPLICIT EIGEN-VALUES AND -FIELDS
forms an orthonormal basis of L2 (M; T M).
The above allows defining intrinsic kernels on gen-
eral compact oriented Riemannian manifolds. How- All of these operators can easily be computed numer-
ever, actually computing these kernel requires solving ically, e.g. via automatic differentiation, which makes
for eigenfields and eigenvalues of the Hodge Laplacian. pointwise evaluation and differentiation of the kernels
Luckily, in some important cases this turns out to be an easy endeavor with modern computing systems.
tractable. We present them in this section.
The Sphere It is well known that the eigenfunctions
4.1 Surfaces and the Sphere of the Laplace–Beltrami operator on the sphere S2 are
given by the spherical harmonics Yℓ,m , for ℓ ≥ 0, −ℓ ≤
The main case of interest here is the sphere S2 . How- m ≤ ℓ, with eigenvalues λℓ = λℓ,m = ℓ(ℓ + 1). Addi-
ever, we start by considering the more general case tionally, the 0-eigenspace of the Hodge Laplacian on
of manifolds of dimension 2 (surfaces). We explain the sphere is trivial, see Appendix A.3. This, together
how to obtain the eigenfields and eigenvalues granted with Theorem 8, allows us to compute the eigenfields.
their scalar counterparts and a basis of harmonic vec- We visualize some of them in Figure 4.
tor fields are known.
The approximation of the full kernel can be made even
more efficient via the addition theorem, e.g. De Vito et
Surfaces Suppose M is a compact, oriented Rieman-
al. (2021, Section 7.3), that states
nian surface. We consider two intrinsic operators for
this case. The first is the gradient of a scalar func- X 2ℓ + 1
tion, giving us a vector field. The second operator Yℓ,m (x)Yℓ,m (x′ ) = Pℓ (x · x′ ), (20)
is the Hodge star operator ⋆ acting on vector fields, 4π
−ℓ≤m≤ℓ
which in the case of surfaces is just a 90◦ rotation of a
vector field in the positive direction, as shown in Ap- where Pℓ is the ℓ-th Legendre polynomial and the
pendix A.2. scalar product is taken in R3 after embedding the S2
Suppose we know all eigenfunctions {fn }n≥0 and their as the standard unit sphere. This reduces the compu-
respective eigenvalues 0 = λ0 < λ1 ≤ λ2 ≤ · · · of tations to a single simple function for each eigenvalue.
the Laplace–Beltrami operator on M. Further, assume As a result, we have the following.
that we have an orthonormal basis {gj }0≤j≤J of the Proposition 9. Writing
0-eigenspace of the Hodge Laplacian.3
2ℓ + 1
3
By the Hodge decomposition theorem, these form a Peℓ,ν,κ (z) = Φν,κ (λℓ )Pℓ (z), (21)
basis of the first de Rham cohomology group (a real vector 4πλℓ
space) of M, which is finite dimensional since M is compact
by assumption. the pure divergence and pure curl Hodge–Matérn ker-
D. Robert-Nicoud, A. Krause, V. Borovitskiy
Pd
4.2 Product Manifolds and Hypertori 1 ≤ j ≤ d, with eigenvalue i=1 λni , where xi ∈ S1 ,
and (fn , λn ) are eigenpairs of the Laplace–Beltrami op-
d
Another tractable setting is product manifolds: if we erator on S1 . In particular, with k T the scalar Matérn
know the eigenfields of the Hodge Laplacian for the kernel on Td , the vector Hodge–Matérn kernel is
factors, we can construct those of the product. We
give an overview of this in general before deriving the d 1 Td
kTν,κ,σ2 (x, x′ ) = k ′
2 (xi , xi )I d . (26)
spectrum of the circle from the scalar case and using d ν,κ,σ
it to resolve the case of the hypertori.
4.3 Possible Extensions
Product Manifolds Let M, N be two compact, ori-
ented Riemannian manifolds with scalar manifold heat We propose two prospective directions into which our
kernels P M , P N and Hodge heat kernels P M , P N . The results could be extended.
vector kernel on the product M × N is given by
P M×N
t (x, x′ ) = PtM (x1 , x′1 )P N ′
t (x2 , x2 ) + Meshes Neither Hodge–Matérn kernels nor the
(23) eigenfields can be analytically computed on a general
+ PtN (x2 , x′2 )P M ′
t (x1 , x1 )
two-dimensional manifold. This is true even for their
for x = (x1 , x2 ) and x′ = (x′1 , x′2 ) and x1 , x′1 ∈ M; scalar counterparts (Borovitskiy et al., 2020). How-
x2 , x′2 ∈ N. The details are explained in Appendix A.6. ever, we expect that Hodge–Matérn kernels can be
numerically approximated on surfaces discretized into
Knowing the Hodge heat kernels, the other Hodge–
meshes. To do this, one needs to apply suitable dis-
Matérn kernels can be derived using Equation (13).
crete counterparts of ⋆ and ∇ to numerically approx-
imate the scalar eigenfunctions the Laplace–Beltrami
Circle The circle S1 is the only4 compact Rieman-
operator and also take care of the harmonic forms.
nian manifold of dimension 1. The Hodge star oper-
4 5
To be precise: any 1-dimensional compact Riemannian Not to be confused with the torus defined as a “donut”
manifold is isometric to the circle of the same length. in R3 , which has a different intrinsic geometry.
Intrinsic Gaussian Vector Fields on Manifolds
(a) Projected Matérn (ν = 12 ) (b) Ground truth (January 2010) (c) Div.-free Hodge-Matérn (ν = 12 )
Figure 6: Interpolation of wind speed on the surface of Earth. The observations are the red vectors along a
meridian. Figures (a) and (c) report predictive mean (black vectors) and uncertainty (color: yellow is high, blue
is low). Note that in figure (a) sinks and sources are present, while the inductive bias of (c) prohibits that. We
advise the reader to examine the global Figures 8 and 9, located in the appendix because of space limitations.
Lie Groups and Related Manifolds It is possi- nels all with ν = 12 , ∞. The hyperparameters are the
ble to obtain the scalar manifold Matérn kernels on noise variance for the first kernel, and the length scale
homogeneous spaces via the representation theory of κ, the variance σ 2 , and the noise variance σε2 for the
their symmetry groups (Azangulov et al., 2022). We others. They were optimized by maximizing marginal
conjecture that the the vector case can be treated log-likelihood. After visual inspection of the ground
similarly—in particular, in view of the work of Ikeda truth, we selected the best performing kernel—the
and Taniguchi (1978). This could result in explicit for- divergence-free Hodge–Matérn kernel with ν = 21 —
mulas for eigenfields and eigenvalues for homogeneous and the projected Matérn- 21 kernel, and ran four fur-
spaces given in terms of algebraic quantities only. ther experiments by fixing their length scale to high
values κ = 0.5 and 1. Importantly, we also experi-
5 EXPERIMENTS mented with the Hodge-compositional Matérn (H.-C.
M.) kernels of Section 3.4 to understand if these could
We complement the theoretical motivations for the use detect the correct inductive bias.6
of intrinsic kernels on manifolds with a practical exper- We report the mean and standard deviation of mean
iment on weather data from the ERA5 dataset (Hers- squared error (MSE) of the predicted mean and pre-
bach et al., 2023). Further experiments on syntheti- dictive negative log-likelihood (PNLL) of the ground
cally generated data are available in Appendix C. truth against the predictive distribution, where each
test point is considered independently of the others.
5.1 Setup
An easy computation shows that the dual of the exte- Building on the case of R3 , we call
rior derivative d : Ωk (M) → Ωk+1 (M) is given by
∇ = d : Ω0 (M) −→ Ω1 (M), (41)
⋆ dM k+1 k+1 k
d = (−1) ⋆ d⋆ : Ω (M) −→ Ω (M). (36) ⋆ 1
div = d : Ω (M) −→ Ω (M), 0
(42)
1 0
curl = ⋆ d : Ω (M) −→ Ω (M), (43)
A.2 Gradient, Divergence, and Curl on
Surfaces under the identification of 1-forms with vector fields
and 2-forms with functions. This also corresponds—
In this section, we will link the abstract operators ⋆, d, potentially up to a sign, depending on conventions—
and d⋆ to quantities which are in some sense more con- to the definition of divergence in Riemannian geometry
crete. Before doing that, we start with a quick recap using the Levi–Civita connection (Lee, 2009, Equation
on vector fields, their curls and divergences in R3 . 13.11). Working in a local coordinate system on M, the
In R3 , a basis for 1-forms is given by dx1 , dx2 , dx3 , and explicit expression for these operators is
a basis for 2-forms is dx1 ∧ dx2 , dx1 ∧ dx3 , dx2 ∧ dx3 .
∂1 f (x)
The Hodge star operator sends the constant function ∇f (x) = , (44)
∂2 f (x)
1 to the volume form dx1 ∧ dx2 ∧ dx3 , and it maps
v (x)
div 1 = ∂1 v1 (x) + ∂2 v2 (x), (45)
⋆ dx1 = dx2 ∧ dx3 , (37) v2 (x)
⋆ dx2 = − dx1 ∧ dx3 , (38) v (x)
curl 1 = ∂1 v2 (x) − ∂2 v1 (x). (46)
v2 (x)
⋆ dx3 = dx1 ∧ dx2 , (39)
and vice versa. Vector fields are identified with 1-forms A.3 The Hodge Laplacian
by mapping dxi 7→ ei , and also with 2-forms via the
Hodge star operator. Let M be a d-dimensional manifold. The Hodge Lapla-
cian on differential forms is then defined as
With this in mind, we immediately see that the exte-
rior derivative of a function ∆ := − (d⋆ d + d d⋆ ) . (47)
df = ∂1 f dx1 + ∂2 f dx2 + ∂3 f dx3 (40) Remark 11. In some texts, including Rosenberg
(1997), the Hodge Laplacian has the opposite sign.
is identified with taking the gradient ∇f . A straight- Remark 12. For k = 0 (i.e. on functions) this recov-
forward calculation also shows that taking the exterior ers the Laplace–Beltrami operator
derivative of a 1-form corresponds to taking the curl of
the corresponding vector field, and the exterior deriva- ∆ = − d d⋆ = −∇∗ ∇, (48)
tive of a 2-form gives the divergence of the associated
where ∇ is the Levi–Civita connection. Cf. also ap-
vector field. This and the fact that d2 = 0 recover the
pendix A.7. In the dual context of vector fields, we
well known relations between ∇, curl, and div. Via the
obtain the classical divergence of the gradient
Hodge star operator, similar statements can be made
about the d⋆ operator. ∆ = div ∇. (49)
On surfaces, the situation is a bit different since we
only have two dimensions in which to move. The fact The following deep results—found in Rosenberg (1997,
that our manifold is oriented gives us an orientation Theorems 1.30, 1.37, and 1.45)—give us all we need to
on each cotangent space Tx∗ M. Picking an oriented know about the spectrum of the Hodge Laplacian.
orthonormal local basis dx1 , dx2 ∈ Tx∗ M, the volume Theorem 13 (Hodge). All the eigenvalues of the
form is given locally by dx1 ∧ dx2 and it follows that Hodge Laplacian ∆ on Ωk (M) are non-negative, they
⋆ dx1 = dx2 and ⋆ dx2 = − dx1 . This corresponds to have finite multiplicity, and they accumulate only at
a rotation by 90◦ in the cotangent space. infinity. The eigenforms span a dense subset of
In the case where our surface is embedded in R3 , the ΩL2 (M). In particular, there exists an orthonormal
choice of an orientation is equivalent to the choice of a basis of ΩL2 (M) consisting of smooth eigenforms of ∆.
global unit normal field for the manifold, i.e. a smooth Remark 14. The convention on eigenpairs is that ϕ
choice of a unit normal vector for each point. Then, is an eigenform of (minus) eigenvalue λ if ∆ϕ = −λϕ.
for tangent vectors the Hodge star is given extrinsically Theorem 15 (Hodge decomposition). The space of
by a rotation by 90◦ around the unit normal at each smooth k-forms decomposes as
point. This can also be written as the cross product
of a tangent vector with the unit normal. Ωk (M) = ker ∆ ⊕ im d ⊕ im d⋆ . (50)
Intrinsic Gaussian Vector Fields on Manifolds
The next result links the kernel of the Hodge Laplacian Proof. For (1), it is immediate to see that ∆⋆ = ⋆∆ so
with a purely topological property of the manifold: de that ⋆ sends eigenforms to eigenforms with the same
Rham cohomology. An accessible introduction is given eigenvalue. If α, β are any two k-forms, then
in Rosenberg (1997, Section 1.4). Z
Theorem 16 (Hodge). The kernel of the Hodge ⟨⋆α, ⋆β⟩L2 (M) = ⋆α ∧ ⋆ ⋆ β (55)
Laplacian on k-forms is naturally isomorphic to the ZM
k-th de Rham cohomology group, which is a real vector = ⋆α ∧ β (56)
space: ZM
ker ∆ ∼ k
= HdR (M). (51) = β ∧ ⋆α = ⟨β, α⟩L2 (M) (57)
M
Remark 17. The following facts about de Rham co- = ⟨α, β⟩L2 (M) , (58)
homology are often useful. Assume M is compact and
connected. showing that ⋆ is an isometry and concluding the
point.
1. HdR0
(M) ∼= R, spanned by the constant function For (2), one easily checks that ∆ d = d∆ so that d
f (x) = 1. sends eigenforms that are in im d⋆ , i.e. the complement
of ker d, to eigenforms with the same eigenvalue. Let
dM
2. HdR (M) ∼
= R, spanned by the volume form. ϕ, ψ ∈ im d⋆ be eigenforms with eigenvalues λϕ , λψ
k respectively, then
3. All of the HdR (M) are finite dimensional, see
e.g. Lee (2009, Theorem 10.17). ⟨dϕ, dψ⟩L2 (M) = ⟨ϕ, d⋆ dψ⟩L2 (M) (59)
= ⟨ϕ, −∆ψ⟩L2 (M) (60)
An example that is exploited in the main body of this
paper is the well known fact that = λψ ⟨ϕ, ψ⟩L2 (M) , (61)
1
HdR (S2 ) = 0. (52) where in the middle equality we used the fact that
ψ ∈ im d⋆ and that d⋆ d⋆ = 0. This concludes the
This can be computed using the Mayer–Vietoris se- proof of the point, and (3) is analogous.
quence, see e.g. Lee (2009, Section 10.1).
A.4 The Heat Equation and its Kernel
Proposition 18. The various spaces of eigenforms
have the following relations. The heat equation for k-forms can now be defined as
1. The Hodge star ⋆ : Ωk (M) → Ωn−k (M) sends ∂t α(t, x) = ∆x α(t, x) (62)
eigenforms of the Hodge Laplacian to eigenforms with a given initial condition α(0, x) = β(x) ∈ Ωk (M).
with the same eigenvalue, and it preserves their
orthogonality and norm. A double (k-)form over M is a smooth section of the
bundle R ⊗ Λk T ∗ M ⊗ Λk T ∗ M over M × M, where the
2. The exterior derivative d : Ωk (M) → Ωk+1 (M) fibre above (x, x′ ) ∈ M × M is R ⊗ Λk Tx∗ M ⊗ Λk Tx∗′ M.
sends eigenforms in im d⋆ to eigenforms with the
A heat kernel for k-forms is a double form Pt (x, y)
same eigenvalue (and is zero on the other eigen-
such that
forms), it preserves orthogonality, and
√ 1. (∂t − ∆x )Pt (x, x′ ) = 0 and
∥ dϕ∥L2 (M) = λ∥ϕ∥L2 (M) (53)
2. limt→∞ M ⟨Pt (x, x′ ), α(x′ )⟩x′ dx′ = α(x) for any
R
for ϕ ∈ im d⋆ ⊆ Ωk (M) an eigenform of eigen- α ∈ Ωk (M), where the pointwise inner product
value −λ. and the integration are taken with respect to x′ ,
and integration is against the volume form of M.
3. The operator d⋆ : Ωk (M) → Ωk−1 (M) sends
eigenforms in im d to eigenforms with the same Theorem 19. Let ϕi ∈ Ωk (M) be an orthonormal ba-
eigenvalue (and is zero on the other eigenforms), sis of k-eigenforms of the Hodge Laplacian with (mi-
it preserves orthogonality, and nus) eigenvalues 0 ≤ λ1 ≤ λ2 ≤ · · · . The heat kernel
√ on k-forms exists, it is unique, and it can be expressed
∥ d⋆ ϕ∥L2 (M) = λ∥ϕ∥L2 (M) (54) by the following sum over eigenforms:
∞
for ϕ ∈ im d ⊆ Ωk (M) an eigenform of eigen-
X
Pt (x, x′ ) = e−λn t ϕn (x) ⊗ ϕn (x′ ). (63)
value −λ. n=0
D. Robert-Nicoud, A. Krause, V. Borovitskiy
Proof. See Rosenberg (1997, Proposition 3.1) for func- Fibrewise bilinearity is obvious, we are left to prove
tions, the discussion at the end of Rosenberg (1997, that our kernel is positive semi-definite in the sense of
Section 3.2) and Patodi (1971) for the general state- Hutchinson et al. (2021, Definition 3). In order to do
ment on forms. so, for xi ∈ M and αxi ∈ Tx∗i M, 1 ≤ i ≤ n, consider
the sequence
Given the heat kernel, the solution for the heat equa- D Ept
tion αim (x) = P m1 (x, xi ), αxi . (71)
xi
(∂t + ∆x )α(t, x) = 0 (64)
with initial condition α(0, x) = β(x) is given by Then we have
Λk Tx∗ M ⊗ Λk Tx∗′ M ∼
= Λk Tx∗′ M ⊗ Λk Tx∗ M. (68) n X
X n
k(αxi , αxj ) = (77)
−t∆ i=1 j=1
2. The propagator e satisfies the semigroup prop-
erty e−(t+s)∆ = e−t∆ e−s∆ . = lim ⟨αm (x), ⟨Pt (x, x′ ), αm (x′ )⟩x′ ⟩x (78)
m→∞
Theorem 5. There exists a Gaussian vector field f • — Theorem 21 (Stone–Weierstrass). Let A ⊆ C ∞ (M)
where • ∈ {div, curl, harm} —with kernel be a subalgebra of the algebra of smooth functions on M
which contains all constant functions and which sepa-
σ2 X rates points, i.e. such that for each x, y ∈ M there is
kν,κ,σ2 (x, x′ ) = Φν,κ (λn )sn (x)⊗sn (x′ ). (16)
•
A.6 Products Proof. First notice that we can choose a finite subset
e1 , . . . , en ∈ E that spans Λk T ∗ M everywhere. Indeed,
In this section, let M, N be two oriented, connected, consider the collection of open sets given by
compact Riemannian manifolds. We recall the Stone–
Weierstrass theorem, see e.g. Prolla (1994). {x ∈ M | e′1 (x), . . . , e′n′ (x) spans Λk Tx∗ M} (85)
D. Robert-Nicoud, A. Krause, V. Borovitskiy
for all possible choices of e′1 , . . . , e′n′ ∈ E. By our as- Proposition 10. The eigenfields on Td are
sumptions, this covers M. Thus, by compactness we !
d
can choose a finite sub-collection that also covers M.
fni (xi ) ej ∈ T S1 ⊗ · · · ⊗ T S1 ∼
Y
Taking all the elements of E appearing in this finite = T Td , (25)
i=1
sub-collection gives us the desired set. Every form
α ∈ Ωk (M) can then be written as Pd
1 ≤ j ≤ d, with eigenvalue i=1 λni , where xi ∈ S1 ,
n and (fn , λn ) are eigenpairs of the Laplace–Beltrami op-
d
X
α(x) = fi (x)ei (x) (86) erator on S1 . In particular, with k T the scalar Matérn
i=1 kernel on Td , the vector Hodge–Matérn kernel is
(1997, Section 2.2), while an extensive treatise on it, Take arbitrary x, x′ ∈ S2 . Then
including the existence of a heat kernel, in Berline et
al. (1995). Cov(f (x), f (x′ )) = Cov(Px g(x), Px′ g(x′ )) (94)
We will now explain what this normalization means Proposition 29. The appropriately normalized pro-
in practice and make these constants explicit for the jected Matérn kernel (with trivial coregionalization
kernels of Hodge–Matérn Gaussian vector fields and matrix) on a manifold M of dimension d embedded in
projected Matérn Gaussian vector fields. RN is given by
Proposition 27. Assume (105) holds, then
1
kπν,κ,σ2 (x, x′ ) = kν,κ,σ2 (x, x′ )PxT I N Px′ , (119)
1 h i
d
Ef ∼GP(0,k) ∥f ∥2L2 (M) = σ 2 . (106)
vol M
where kν,κ,σ2 is the scalar manifold Matérn kernel.
Proof. We have
1 Proof. Let x = x′ and pick a coordinate system such
h i
Ef ∼GP(0,k) ∥f ∥2L2 (M) = (107)
vol M that Tx M = span{e1 , . . . , ed }. Then we have
Z
1
= Ef ∼GP(0,k) ∥f (x)∥22 dx (108) PxT I N Px′ = diag(1, . . . , 1, 0, . . . , 0). (120)
vol M M
Z d
1
Ef ∼GP(0,k) ∥f (x)∥22 dx
= (109)
vol M M
Z It follows that
1
Ef (x)∼N (0,k(x,x)) ∥f (x)∥22 dx
= (110) Z
vol M M 1
tr kPν,κ,σ 2 (x, x) dx = (121)
vol M M
Z
1
= tr k(x, x) dx (111) Z
vol M M 1
= kν,κ,σ2 (x, x) dx (122)
= σ2 , (112) vol M M
= σ2 , (123)
where the last equality holds by (105).
Proposition 28. The constant Cν,κ for the Hodge– as desired.
Matérn kernel kν,κ,σ2 is given by
∞
B.3 Divergence of Gaussian Vector Fields
1 X
Cν,κ = Φν,κ (λn ), (113)
vol M n=0 We will now study the distribution of the (pointwise)
divergence of the Hodge–Matérn Gaussian vector fields
where the sum runs over all of the eigenfields of the and projected Matérn GPs. Similar techniques can be
Hodge Laplacian. Similar formulas are valid for the used to compute the full distribution for dα and d⋆ α
pure divergence, pure curl, and harmonic kernels by for α ∼ GP(0, kν,κ,σ2 ), which turns out to be another
restricting the sum to the appearing eigenfields. Gaussian process.
We fix a compact, oriented Riemannian manifold M of
Proof. We have
Z dimension dM ≥ 1 and we look at degree k = 1 dif-
ferential forms, although the computations straight-
tr(kν,κ,σ2 (x, x)) dx = (114)
M forwardly generalize to all other k. We write fn ∈
∞
! C ∞ (M) = Ω0 (M) for a basis of eigenfunctions with
σ2 X
Z
= tr Φν,κ (λn )sn (x) ⊗ sn (x) dx eigenvalues −λn , for n ≥ 0, where 0 = λ0 < λ1 ≤ · · · .
M Cν,κ n=0 We also set ϕn , n ∈ N a basis of 1-eigenforms with
(115) eigenvalues including the eigenforms
∞
σ2 X Z
= Φν,κ (λn ) tr (sn (x) ⊗ sn (x)) dx (116) 1
Cν,κ M
√ dfn (124)
n=0 λn
∞
σ2 X Z
= Φν,κ (λn ) ∥sn (x)∥22 dx (117) for n ≥ 1. We use Cν,κ k
to denote the normal-
Cν,κ n=0 M
ization constant for the Hodge–Matérn kernel on k-
∞
σ2 X forms (with k = 0 being the case of functions). For
= Φν,κ (λn ), (118)
Cν,κ α ∼ GP(0, k) a Gaussian differential form, we write
n=0
div α(x) for the random variable d⋆ α(x), where x ∈ M.
since ∥sn ∥L2 (M) = 1. Notice that here the trace is We assume div α(x) is well-defined, i.e. the Gaussian
taken with respect to the metric on M. The result process is smooth enough, which places restrictions on
follows immediately by requiring the left-hand side of the parameter ν. We leave the precise nature of these
the equation to equal σ 2 vol M. out of the scope of this paper.
Intrinsic Gaussian Vector Fields on Manifolds
Proposition 30. For the Hodge–Matérn Gaussian We now look at the projected kernel. Suppose ϕ : M →
form αν,κ,σ ∼ GP(0, kν,κ,σ2 ) on 1-forms we have RN is an isometric embedding and write kπν,κ,σ2 for the
∞ projected kernel obtained by projecting the vector ker-
σ2 X nel in RN where each component is a scalar manifold
Var(div αν,κ,σ (x)) = 1
λn Φν,κ (λn )fn (x)2 (125)
Cν,κ n=1 Matérn kernel with the same hyperparameters ν, κ, σ 2 .
In this case, we will talk about vectors and gradients
whenever αν,κ,σ is smooth enough for the divergence instead of differential forms (although a formulation in
to be well-defined. terms of 1-forms is also possible).
dM
X Corollary 34. On the sphere we have
= ∇ei w · e i + w · ∇ei e i − w · ∇ ei e i (145)
π
i=1 Var(div fν,κ,σ 2) =
∞
σ2
P
dM (155)
X n=1 λn Φν,κ (λn )
= + 1
= w · ∇ei ei − ∇ei ei (146) 2 0
4πCν,κ
i=1
dM
X
! for any x ∈ S2 .
=w· II(ei , ei ) (147)
i=1 Proof. By the same symmetry argument as in Corol-
= w · H(x), (148) lary 31 and using the fact that for the standard unit
sphere embedding we have H(x) = n(x) the unit nor-
where (142) comes from the fact that ∇ is the Levi– mal vector, we obtain
Civita connection, (143) from the fact that ϕ is an π
isometric embedding, and (145) comes from the fact 4π Var(div fν,κ,σ 2) =
π
The training points were selected by taking the obser-
whenever fν,κ,σ 2 is smooth enough for the divergence
vations at longitudes 90◦ E and 90◦ W, and then picking
to be well-defined. one every 180 of them, spaced regularly. This gives a
final training set of 34 points. The 1220 testing points
Proof. Once again, by Proposition 6 we have were generated randomly.8
π
fν,κ,σ 2 (x) = Using the results of Appendix B.3, it was possible to
∞ q quantify the variance of the divergence of the Gaussian
σ X (152)
=q Φν,κ (λn )fn (x)Px wn vector fields arising from the (prior) kernels that were
0
dM Cν,κ n=0 fitted in the experiments. The results are displayed
in Figure 10, confirming that the absolute divergence
where wn ∼ N (0, I N ) is a sequence of i.i.d. multivari- was higher for the Hodge heat and Hodge–Matérn ker-
ate normal vectors. Taking the divergence by applying nels, which explains the worse performance of these
Lemma 32 to each summand, we obtain intrinsic kernels against the projected kernels for this
π
experiment.
divfν,κ,σ 2 (x) =
In the last experiment on this data, we fitted the spective kernel performed best. On the rotation vec-
Hodge-compositional Matérn kernels, i.e. linear com- tor field, Hodge–Matérn and divergence-free Hodge–
binations of pure-divergence and pure-curl Hodge– Matérn vastly outperformed all other kernels.
Matérn kernels, as reported also in Table 1. We
see that these kernels recover almost exactly the re-
sults of fitting a divergence-free Hodge–Matérn ker-
nel. A detailed analysis of the fitted hyperparameters
shows that the resulting length scales and variances
put full weight on the divergence-free part of the ker-
nel with an almost exact match of length scales. The
only exceptions are when the length scale converges
to a local optimum, which fully explains the minimal
advantage in performance of the linear combination
kernel over the divergence-free Hodge–Matérn kernel
with ν = 12 . This supports the fact that Hodge-
compositional Matérn kernels are able to automati-
cally recover appropriate inductive biases in such sit-
uations, mirroring their discrete counterparts in Yang
et al. (2024).
curl-free
H.–M.– 12 sample H.–M.–∞ sample P. M.– 21 sample Rotation field
Kernel H.–M.– 12 sample
Mean Std Mean Std Mean Std Mean Std Mean Std
Pure noise 0.17 0.04 1.18 0.32 0.22 0.06 0.68 0.02 0.08 0.02
P. M.– 12 0.14 0.03 0.87 0.42 0.16 0.05 0.06 0.02 0.07 0.02
P. M.–∞ 0.19 0.04 0.71 0.26 0.20 0.07 0.34 0.36 0.07 0.03
H.–M.– 12 0.14 0.03 0.84 0.38 0.16 0.05 0.02 0.01 0.07 0.02
H.–M.–∞ 0.17 0.04 0.65 0.25 0.20 0.06 0.00 0.00 0.08 0.02
curl-free
0.20 0.06 1.15 0.47 0.22 0.07 0.68 0.02 0.05 0.01
H.–M.– 21
curl-free
0.16 0.04 1.11 0.37 0.23 0.06 0.69 0.03 0.08 0.02
H.–M.–∞
div-free
0.15 0.05 1.00 0.60 0.19 0.05 0.01 0.00 0.08 0.02
H.–M.– 12
div-free
0.16 0.04 0.81 0.46 0.18 0.04 0.00 0.00 0.08 0.02
H.–M.–∞
Table 3: MSE for synthetic experiments. The columns are datasets, the rows are models.
curl-free
H.–M.– 12 sample H.–M.–∞ sample P. M.– 21 sample Rotation field
Kernel H.–M.– 12 sample
Mean Std Mean Std Mean Std Mean Std Mean Std
Pure noise 0.41 0.25 2.39 0.34 0.67 0.29 1.76 0.04 -0.31 0.25
P. M.– 12 0.15 0.26 2.18 0.78 0.31 0.27 -0.65 0.20 -0.51 0.34
P. M.–∞ 0.61 0.39 1.47 0.63 0.58 0.40 -3.43 5.47 -0.38 0.73
H.–M.– 12 0.13 0.24 2.12 0.71 0.33 0.28 -1.41 0.15 -0.58 0.28
H.–M.–∞ 0.41 0.25 1.27 0.48 0.53 0.31 -9.42 0.04 -0.31 0.25
curl-free
0.66 0.59 2.66 0.71 0.67 0.37 1.77 0.03 -0.79 0.21
H.–M.– 21
curl-free
0.38 0.30 2.55 0.58 0.73 0.32 1.77 0.05 -0.31 0.25
H.–M.–∞
div-free
0.25 0.31 2.37 1.16 0.48 0.31 -2.20 0.17 -0.37 0.31
H.–M.– 12
div-free
0.33 0.25 1.52 0.66 0.46 0.18 -9.63 0.00 -0.31 0.25
H.–M.–∞
Table 4: Predictive NLL for synthetic experiments. The columns are datasets, the rows are models.
Intrinsic Gaussian Vector Fields on Manifolds
(b) Predictive mean and uncertainty (blue is low and yellow is high).
Figure 8: Robinson projection of Figures 6a, 6b and 7a displaying the ground truth, observations, predictive
mean, uncertainty, and a posterior sample of the GP with projected Matérn kernel with ν = 21 and length
scale κ = 0.5. The vectors in the sample are scaled independently from the ground truth and predictive mean.
Visually, the sample does not have structures reminiscent of that of the ground truth.
D. Robert-Nicoud, A. Krause, V. Borovitskiy
(b) Predictive mean and uncertainty (blue is low and yellow is high).
Figure 9: Robinson projection of Figures 6b, 6c and 7b displaying the ground truth, observations, predictive
mean, uncertainty, and a posterior sample of the GP with divergence-free Hodge–Matérn kernel with ν = 12 and
length scale κ = 0.5. The vectors in the sample are scaled independently from the ground truth and predictive
mean. Visually, and to the contrary of Figure 8, the sample appears to have structures reminiscent of that of
the ground truth, such as a strong west to east current in the southern hemisphere.
Intrinsic Gaussian Vector Fields on Manifolds
Kernel
3.5 P. M.−∞
P. M.−12
H.−M.−∞
3.0 H.−M.−12
2.5
Var(div f)
2.0
1.5
1.0
0.5
0.0
0 1 2 3 4 5 6 7 8 9 10 11
Experiment number
Figure 10: Variance of divergence for the prior kernels with fitted hyperparameters in the weather modeling
experiments. Note that the variance of the divergence does not depend on the input location in this case. The
divergence-free kernels and the kernels with fixed length scale are not represented here.