---
title: Divergence-Free Gaussian Reference Measure
url: https://www.emergentmind.com/topics/divergence-free-gaussian-reference-measure
type: topic
---

# Divergence-Free Gaussian Reference Measure

A divergence-free Gaussian reference measure is a Gaussian probability measure whose support lies in a divergence-free subspace of vector fields, so that almost every sample path satisfies the incompressibility constraint $\nabla \!\cdot u = 0$. In the recent literature, this object appears in two closely related settings: as a physically admissible prior for generative modeling of incompressible flows on the torus, constructed by pushing a scalar Gaussian through a curl operator [2603.24500], and as the law of a vector-valued Gaussian process on a bounded domain whose covariance kernel is divergence-free by construction [2511.12535]. Across both settings, the central purpose is structural rather than penalized enforcement of incompressibility: the Gaussian reference measure is defined directly on the divergence-free subspace, so admissibility is built into sampling, interpolation, conditioning, and generative transport.

## 1. Divergence-free structure as a measure-theoretic constraint

In the torus-based formulation, the ambient velocity-field space is
$$
H:=L^2_{\mathrm{per}}(\mathbb{T}^2;\mathbb{R}^2),
$$
the real Hilbert space of square-integrable, zero-mean, $2$-vector fields with periodic boundary conditions, and the scalar stream-function space is
$$
X:=L^2_{\mathrm{per}}(\mathbb{T}^2;\mathbb{R}).
$$
The divergence-free subspace is
$$
V=\{u\in H:\nabla\cdot u=0\},
$$
and the orthogonal complement is
$$
V^\perp=\{\nabla\phi:\phi\in H^1_{\mathrm{per}}(\mathbb{T}^2),\ \int \phi=0\}.
$$
This yields the Helmholtz-Hodge splitting
$$
H=V\oplus V^\perp,
$$
which provides the basic functional-analytic setting for incompressible Gaussian measures [2603.24500].

In the bounded-domain Gaussian-process formulation, one considers a compact domain $D\subset\mathbb{R}^d$ with $d=2,3$ and vector fields in $C(D)^d$ or Sobolev spaces $H^\tau(D)^d$. A continuously differentiable process $f\sim \mathrm{GP}(m,K)$ is called divergence-free if
$$
\nabla\!\cdot f(\omega,x)=\sum_{i=1}^d\frac{\partial f_i}{\partial x_i}(\omega,x)=0
\quad\text{for all }x\in D,\ \text{a.e. }\omega\in\Omega.
$$
Lemma 3.2 in the summarized report states that this is equivalent to the mean $m$ being divergence-free and, for each fixed $x'$, every column of the matrix-valued map $x\mapsto K(x,x')$ being divergence-free [2511.12535].

These two settings differ in representation—spectral on $\mathbb{T}^2$, kernel-based on $D\subset\mathbb{R}^d$—but they agree on the defining property of the measure: it is not merely centered on divergence-free data, but supported on a divergence-free function class.

## 2. Gaussian measures on vector-field spaces

A centered Gaussian measure $\mu$ on $H$ is determined by its covariance operator $K:H\to H$, assumed bounded, self-adjoint, nonnegative, and trace-class. Writing
$$
\mu=N(0,K),
$$
means that for any $h\in H$ the linear functional $\langle h,\cdot\rangle$ has law $N(0,\langle h,Kh\rangle)$. Under periodic boundary conditions, $K$ is diagonalized by the Fourier basis $\{e_k(x)=e^{ik\cdot x}\}_{k\in\mathbb{Z}^2\setminus\{0\}}$, so that
$$
K e_k=\lambda_k e_k,\qquad \lambda_k\ge 0,
$$
and a draw $u\sim\mu$ admits the spectral expansion
$$
u(x)=\sum_{k\neq 0}\sqrt{\lambda_k}\,\xi_k\,e_k(x),\qquad \xi_k\sim \text{i.i.d. }N(0,1)
$$
[2603.24500].

In the Gaussian-process setting, a vector-valued process
$$
f:\Omega\times D\to \mathbb{R}^d,\qquad f\sim\mathrm{GP}(m,K),
$$
is characterized by the mean
$$
m(x)=\mathbb{E}[f(x)]\in\mathbb{R}^d
$$
and covariance kernel
$$
K(x,x')=\mathbb{E}\Bigl[\bigl(f(x)-m(x)\bigr)\bigl(f(x')-m(x')\bigr)^T\Bigr]\in\mathbb{R}^{d\times d}.
$$
For any finite set $X=\{x_1,\dots,x_N\}\subset D$, the joint random vector $(f(x_1),\dots,f(x_N))\in\mathbb{R}^{dN}$ is Gaussian with mean $(m(x_i))_{i=1}^N$ and block-matrix covariance $\bigl[K(x_i,x_j)\bigr]_{i,j=1}^N$ [2511.12535].

The measure-theoretic language is explicit in the GP construction: by Kolmogorov’s extension theorem, the law of $f$ is a Borel probability measure
$$
\mu\equiv\mathcal{L}(f)\quad\text{on }C(D)^d,
$$
completely determined by its mean function and covariance operator
$$
(C\varphi)(x)=\int_D K_{\rm div}(x,y)\,\varphi(y)\,dy,\qquad \varphi\in L^2(D)^d
$$
[2511.12535].

## 3. Curl-based pushforward on the torus

The torus construction in "Project and Generate: Divergence-Free Neural Operators for Incompressible Flows" [2603.24500] defines a divergence-free Gaussian reference measure by pushing forward a scalar Gaussian through the stream-function operator
$$
T:X\to H,\qquad T\phi:=\nabla^\perp\phi\equiv (\partial_y\phi,-\partial_x\phi)^T.
$$
By elementary calculus,
$$
\nabla\cdot(\nabla^\perp\phi)=0,
$$
so $\mathrm{Ran}\,T\subset V$.

Let $\mu_\phi=N(0,C_\phi)$ be any Gaussian measure on $X$; the divergence-free measure is then
$$
\mu_{\rm df}:=T_\#\mu_\phi,
$$
meaning that for any measurable $B\subset H$,
$$
\mu_{\rm df}(B)=\mu_\phi(T^{-1}(B)).
$$
Because $T$ is linear and bounded, $\mu_{\rm df}$ is again Gaussian, with mean zero and covariance
$$
K_{\rm df}=T\,C_\phi\,T^*,
$$
where $T^*:H\to X$ is the $L^2$-adjoint of $T$ [2603.24500].

When $C_\phi e_k=\sigma_k e_k$ in the scalar Fourier basis, the covariance on velocity modes is expressed as
$$
K_{\rm df}u=\sum_{k\neq 0}\sigma_k\,\bigl(i\,k^\perp e_k\bigr)\bigl(i\,k^\perp e_k\bigr)^*u,
$$
with $k^\perp=(-k_y,k_x)$. The resulting samples are exactly divergence-free, since for any $u=T\phi$,
$$
\nabla\cdot u=\nabla\cdot(\nabla^\perp\phi)=\partial_x\partial_y\phi-\partial_y\partial_x\phi=0
$$
pointwise, and hence up to discretization error on a grid [2603.24500].

This construction is closed-form and intrinsic: the incompressibility constraint is enforced by the map defining the measure itself, not by a subsequent correction step.

## 4. Spectral Leray projection and subspace geometry

The same torus-based framework pairs the divergence-free Gaussian measure with the spectral Leray projector. The $L^2$-orthogonal projection $P:H\to V$ is given in Fourier space by
$$
(Pw)^\wedge(k)=\biggl(I-\frac{k\,k^T}{|k|^2}\biggr)\,w^\wedge(k),\qquad k\neq 0,\qquad (Pw)^\wedge(0)=0,
$$
or equivalently in physical space by
$$
P=I-\nabla\Delta^{-1}\nabla\cdot.
$$
For every mode, one checks
$$
k\cdot (Pw)^\wedge(k)=0,
$$
which makes $P$ the canonical linear mechanism for restricting a field to the divergence-free subspace [2603.24500].

Within the paper’s broader framework, the projector is used for deterministic regression and for constrained generative dynamics. The abstract states that learning-based models for fluid dynamics often operate in unconstrained function spaces, leading to physically inadmissible, unstable simulations, while penalty-based methods offer soft regularization but provide no structural guarantees, resulting in spurious divergence and long-term collapse. The Leray projection is introduced as a differentiable spectral projection grounded in the Helmholtz-Hodge decomposition, restricting the regression hypothesis space to physically admissible velocity fields [2603.24500].

A central clarification made in the same work is that simply projecting model outputs is insufficient when the prior is incompatible. This identifies a common misconception in constrained generative modeling: output projection alone does not ensure that the reference measure, intermediate measures, and transported law all remain subspace-consistent. The divergence-free Gaussian reference measure is introduced precisely to resolve this mismatch.

## 5. Kernel-based divergence-free Gaussian processes on bounded domains

A second construction uses matrix-valued covariance kernels. Let $\Phi:\mathbb{R}^d\to\mathbb{R}$ be a scalar positive-definite radial kernel of sufficient smoothness, such as a Matérn or Wendland kernel, and define
$$
K_{\rm div}(x,y)=\bigl(-\Delta_x I_d+\nabla_x\nabla_x^T\bigr)\Phi(x-y),\qquad x,y\in D.
$$
Componentwise,
$$
\bigl(K_{\rm div}(x,y)\bigr)_{ij}
=-\delta_{ij}\,\Delta\Phi(x-y)+\frac{\partial^2}{\partial x_i\,\partial x_j}\Phi(x-y).
$$
By construction, $K_{\rm div}(x,y)=K_{\rm div}(y,x)^T$, is positive-definite, and for each fixed $y$ its columns lie in $C^1(D)^d$ and satisfy $\nabla_x\!\cdot K_{\rm div}(x,y)=0$. Hence
$$
f\sim\mathrm{GP}(0,K_{\rm div})
$$
is a centered divergence-free Gaussian process on $D$ [2511.12535].

The induced Gaussian reference measure is supported on
$$
\mathbb{H}_{\rm div}
=
\{v\in C(D)^d:\nabla\!\cdot v=0\text{ on }D\},
$$
and almost every sample path of $f$ lies in this closed subspace [2511.12535]. This kernel formulation is the bounded-domain analogue of the torus pushforward construction in the sense that divergence-freeness is encoded at the covariance level rather than imposed after sampling.

The associated reproducing-kernel Hilbert space is
$$
\mathcal{H}_{\rm div}
=
\overline{\mathrm{span}\{K_{\rm div}(\cdot,x)w:x\in D,\ w\in\mathbb{R}^d\}}
\subset C(D)^d,
$$
with reproducing property
$$
\langle v(\cdot),K_{\rm div}(\cdot,x)w\rangle_{\mathcal H}=v(x)^T w.
$$
By Mercer’s theorem and the Karhunen-Loève expansion, there exist eigenpairs $(\lambda_k,\phi_k)$ of the integral operator $C$ in $L^2(D)^d$, with $\phi_k$ divergence-free and
$$
K_{\rm div}(x,y)=\sum_{k=1}^\infty \lambda_k\,\phi_k(x)\phi_k(y)^T,
$$
$$
f(\omega,x)=\sum_{k=1}^\infty \sqrt{\lambda_k}\,\xi_k(\omega)\,\phi_k(x),
\qquad \xi_k\sim N(0,1)\ \text{i.i.d.}
$$
The RKHS is the Cameron-Martin space of $\mu$, with norm
$$
\|h\|_{\mathcal H}^2=\sum_{k=1}^\infty
\frac{\langle h,\phi_k\rangle_{L^2(D)^d}^2}{\lambda_k},
$$
or equivalently
$$
\|h\|_{\mathcal H}^2=\langle h,C^{-1}h\rangle_{L^2(D)^d},
$$
whenever $h$ lies in the range of $C^{1/2}$ [2511.12535].

## 6. Conditioning, approximation, and generative transport

The bounded-domain GP formalism provides explicit conditioning formulae. For noiseless observations $y_j=f(x_j)$, the posterior process is
$$
f_N\sim\mathrm{GP}(m_N,K_N),
$$
with
$$
m_N(x)=m(x)+K_{\rm div}(x,X)\,K_{\rm div}(X,X)^{-1}\bigl(Y-m(X)\bigr),
$$
$$
K_N(x,x')
=
K_{\rm div}(x,x')
-
K_{\rm div}(x,X)\,K_{\rm div}(X,X)^{-1}K_{\rm div}(X,x')^T.
$$
For noisy observations $y_j=f(x_j)+\varepsilon_j$ with $\varepsilon_j\sim N(0,\sigma^2 I_d)$ i.i.d., the predictive mean and covariance become
$$
m_{N,\sigma}(x)
=
m(x)+K_{\rm div}(x,X)\bigl(K_{\rm div}(X,X)+\sigma^2I_{dN}\bigr)^{-1}\bigl(Y-m(X)\bigr),
$$
$$
K_{N,\sigma}(x,x')
=
K_{\rm div}(x,x')
-
K_{\rm div}(x,X)\bigl(K_{\rm div}(X,X)+\sigma^2I_{dN}\bigr)^{-1}K_{\rm div}(X,x')^T
$$
[2511.12535].

The same report states error estimates in Sobolev norms under assumptions linking the prior RKHS to $H^\tau(D;\mathrm{div})$ and using the fill distance
$$
h_{X,D}=\sup_{x\in D}\min_{x_j\in X}\|x-x_j\|.
$$
For interpolation without noise, Theorem 5.1 gives rates depending on smoothness, target norm, and fill distance; for approximation with noise and Tikhonov regularization, Theorem 5.4 yields bounds containing terms of the form
$$
h_{X,D}^{\tau-s-d(\tfrac12-\tfrac1q)_+}
\quad\text{and}\quad
\sqrt{\lambda}\,h_{X,D}^{\,d/\gamma-s},
$$
plus terms involving the noise norm $\|\varepsilon\|_{\ell_2}$, and in the correctly specified Gaussian noise case $\lambda=\sigma^2$ the expectation over the noise yields rates of the same order plus an additional term $O(h_{X,D}^{-d/2}\sigma)$ [2511.12535].

In the torus generative-model setting, the divergence-free Gaussian reference measure is used as the initial law in a flow-matching or normalizing-flow-based model. If $\nu$ is the data distribution on $V$, $u_0\sim\mu_{\rm df}$, and $u_1\sim\nu$, then one sets
$$
u_t=(1-t)\,u_0+t\,u_1,\qquad 0\le t\le 1,
$$
so each $u_t\in V$. A time-dependent vector field $v_\theta:V\times[0,1]\to H$ is learned by minimizing
$$
L(\theta)=\int_0^1 \mathbb{E}_{u_t}\|P\,v_\theta(u_t,t)-\dot u_t\|_{L^2}^2\,dt.
$$
Because $P\,v_\theta(u_t,t)$ itself lies in $V$ and the support of $u_t$ is $V$, the generative ODE
$$
\partial_t x(t)=P\,v_\theta(x(t),t),\qquad x(0)\sim\mu_{\rm df}
$$
remains exactly in $V$ for all $t$ and produces incompressible velocity fields [2603.24500].

A plausible implication is that the two lines of work address complementary operations on divergence-free measures: one emphasizes posterior inference and approximation under pointwise observations, while the other emphasizes transport between a divergence-free reference law and a target law under learned dynamics.

## 7. Support, regularity, and computation

For the curl-pushforward construction, $\mu_{\rm df}=N(0,K_{\rm df})$ on $V$, and its support is the closure of $\mathrm{Ran}\,K_{\rm df}=\mathrm{Ran}\,T$ in $V$. If one chooses $C_\phi$ with eigenvalues $\sigma_k\sim |k|^{-2\alpha}$, then draws $\phi$ lie almost surely in the Sobolev space $H^s_{\mathrm{per}}$ with $s<\alpha-1$, and hence
$$
u=\nabla^\perp\phi\in H^{s-1}
$$
[2603.24500].

The same source gives an FFT-based spectral discretization. On a uniform $N\times N$ grid, one implements $T$ and $P$ by FFTs in $O(N^2\log N)$ operations. Sampling $\mu_{\rm df}$ amounts to three steps: draw scalar Fourier modes $\hat\phi(k)\sim \sqrt{\sigma_k}N(0,1)$, set $\hat u(k)=i\,k^\perp \hat\phi(k)$, and apply the inverse FFT to obtain $u(x)$. Each projection or curl-pushforward requires two forward and two inverse FFTs, and the reported practical remark is that this overhead is minor compared to the cost of evaluating a large neural-operator network [2603.24500].

For the kernel-based GP construction, a Karhunen-Loève truncation is
$$
f(\omega,x)\approx \sum_{k=1}^M \sqrt{\lambda_k}\,\xi_k(\omega)\,\phi_k(x),
\qquad \xi_k\sim N(0,1),
$$
where $\{(\lambda_k,\phi_k)\}$ are the leading eigenpairs of the integral operator $C$. Numerically, one approximates the eigendecomposition of the matrix $K_{\rm div}(X,X)$ and extends to new points by
$$
f(\omega,x)\approx K_{\rm div}(x,X)\,U_M\,\Lambda_M^{-1/2}\,\xi_M,
$$
where $U_M\Lambda_M U_M^T$ is the rank-$M$ eigendecomposition of $K_{\rm div}(X,X)$ [2511.12535].

Kernel evaluation requires computation of
$$
\Delta\Phi(r),\qquad \frac{\partial^2\Phi}{\partial r_i\,\partial r_j}(r),\qquad r=x-y.
$$
For a radial kernel $\Phi(r)=\varphi(\|r\|)$, the report gives
$$
\Delta\Phi(r)=\varphi''(\|r\|)+\frac{d-1}{\|r\|}\varphi'(\|r\|),
$$
and
$$
\frac{\partial^2\Phi}{\partial r_i\,\partial r_j}
=
\varphi''(\|r\|)\frac{r_i r_j}{\|r\|^2}
+
\varphi'(\|r\|)\frac{\delta_{ij}-r_i r_j/\|r\|^2}{\|r\|}.
$$
The algorithmic summary given there is: form the Gram matrix $K_{ij}=K_{\rm div}(x_i,x_j)$; solve $K\alpha=Y$ in the noiseless case or $(K+\sigma^2I)\alpha=Y$ otherwise; predict by
$$
f_N(x)=m(x)+K_{\rm div}(x,X)\alpha;
$$
and, for uncertainty quantification, compute the predictive covariance using the standard Schur-complement formula already stated [2511.12535].

Across both implementations, the defining computational theme is subspace consistency. In the torus setting, both the reference $\mu_{\rm df}$ and the learned dynamics lie in $V$, so the entire generative flow is well-defined on the divergence-free subspace without further corrections or penalties [2603.24500]. In the bounded-domain GP setting, the covariance kernel itself has divergence-free columns, so interpolation, regression, and posterior sampling remain inside the divergence-free class [2511.12535].

Source: https://www.emergentmind.com/topics/divergence-free-gaussian-reference-measure