Kac–Rice Formula Overview

• The Kac–Rice formula is a family of identities that computes expected counts of zeros, level crossings, and measures of level sets by integrating local densities of random fields and their derivatives.
• It is widely applied in areas such as random matrix theory, algebraic statistics, and adaptive regression, where it links geometric methods with probabilistic analysis.
• Modern variants extend the technique to high-dimensional settings, critical value analysis, and piecewise smooth processes, offering concrete tools for complex system evaluations.

The Kac–Rice formula is a family of identities that expresses expected counts of zeros, level crossings, critical points, or more generally Hausdorff measures of level sets, as integrals of local densities built from a random field and its derivatives. In the scalar one-dimensional case it computes the expected number of crossings of a level; in the vector-valued case $X:T\subset \mathbb{R}^D\to \mathbb{R}^d$ it computes $$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$; and modern variants extend the same principle to critical values, Euler characteristic, transverse intersections, random sections of vector bundles, and weighted counting measures (Berzin et al., 2022, Armentano et al., 2023, Stecconi, 2021).

1. Classical formulations

In its standard one-dimensional form, the formula concerns a real process $X(t)$ and a level $u$. If $Y(t)=\dot X(t)$ exists and $p(x,y,t)$ denotes the joint density of $(X(t),Y(t))$, then the crossing intensity is

$$\mu_u(t)=\int_{-\infty}^{\infty} |y|\,p(u,y,t)\,dy,$$

and the expected number of crossings on $[t_0,t]$ is

$$\mathbb{E}\big[N_u(t_0,t)\big]=\int_{t_0}^t \mu_u(t')\,dt' =\int_{t_0}^t\int_{-\infty}^{\infty}|y|\,p(u,y,t')\,dy\,dt'.$$

Equivalently,

$$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$0

which is the one-dimensional Kac–Rice form for level crossings (Masoliver et al., 2023). In the stationary Gaussian case with independent $$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$1 and $$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$2, this reduces to Rice’s classical formula

$$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$3

and at the mean level $$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$4 one obtains $$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$5 (Masoliver et al., 2023).

The higher-dimensional Euclidean version replaces the absolute derivative by the normal Jacobian. For a random field $$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$6, a typical Kac–Rice formula gives

$$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$7

and in the scalar case $$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$8 this becomes an integral of $$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$9 conditioned on $X(t)$0 (Berzin et al., 2022).

Complex-analytic variants are structurally similar but use the real dimension $X(t)$1. For a complex-valued process $X(t)$2, the expected number of zeros satisfies

$X(t)$3

where $X(t)$4 is the joint density of $X(t)$5 (Feng et al., 2012). This same dimensional principle reappears in formulas for critical values of complex Gaussian fields.

2. Geometric and manifold formulations

A central geometric underpinning is the coarea formula. For a $X(t)$6 map $X(t)$7, with

$X(t)$8

the coarea identity reads

$X(t)$9

and a weighted version inserts an extra measurable factor $u$0 inside both integrals (Berzin et al., 2022). A direct application gives Kac–Rice formulas for almost all levels, while continuity in the level variable is used to upgrade them to formulas valid for all levels, which is emphasized as necessary because many applications require the formula at level zero (Berzin et al., 2022).

A general modern statement for level-set measure is the formula

$u$1

for a random field $u$2 with $u$3 and almost surely $u$4 sample paths (Armentano et al., 2023). In that setting, a weak Bulinskaya condition,

$u$5

is sufficient for $u$6-rectifiability of the level set, and the proof for $u$7 uses Crofton’s formula to reduce Hausdorff measure to averaged intersection counts on affine slices (Armentano et al., 2023).

On compact Riemannian manifolds, deterministic “closed Kac–Rice type” formulas compute the volume of the nodal set of a nondegenerate $u$8 function $u$9. The basic divergence formula is

$Y(t)=\dot X(t)$0

for suitable $Y(t)=\dot X(t)$1, and later formulas express $Y(t)=\dot X(t)$2 directly in terms of $Y(t)=\dot X(t)$3, $Y(t)=\dot X(t)$4, $Y(t)=\dot X(t)$5, $Y(t)=\dot X(t)$6, and $Y(t)=\dot X(t)$7 (Jubin, 2019). This suggests a deterministic geometric counterpart of probabilistic Kac–Rice: the zero-set volume is represented as an ambient integral of local differential invariants rather than as an integral over the level set itself.

3. Critical points, critical values, and Euler characteristic

A major extension replaces zeros by critical points or critical values. For Gaussian $Y(t)=\dot X(t)$8 random polynomials

$Y(t)=\dot X(t)$9

the empirical measure of critical values is

$p(x,y,t)$0

and the expected density satisfies the complex Kac–Rice formula

$p(x,y,t)$1

where $p(x,y,t)$2 is the joint density of $p(x,y,t)$3 evaluated at $p(x,y,t)$4 (Feng et al., 2012). In this setting the covariance matrix of $p(x,y,t)$5 is explicitly invertible for $p(x,y,t)$6, yielding an exact integral representation of the density. The same analysis shows that the nonvanishing critical values of $p(x,y,t)$7 accumulate at infinity, that the natural scale is $p(x,y,t)$8, and that under logarithmic rescaling the normalized critical values converge in law to an exponential distribution on $p(x,y,t)$9 (Feng et al., 2012).

In two dimensions, the Kac–Rice philosophy extends from counting to topology. For a $(X(t),Y(t))$0 function $(X(t),Y(t))$1, the Euler primitive

$(X(t),Y(t))$2

admits the exact identity

$(X(t),Y(t))$3

for suitable quarter-planes $(X(t),Y(t))$4 in gradient space (Lachièze-Rey, 2016). For random fields, expectation can be passed under the integral under minimal requirements that do not involve density assumptions on the marginals of $(X(t),Y(t))$5 or of its derivatives (Lachièze-Rey, 2016). This produces a two-dimensional analogue of Kac–Rice for the Euler characteristic of excursion sets rather than for raw level-set measure.

For chi-fields on a manifold, with

$(X(t),Y(t))$6

Kac–Rice is applied to the critical set of $(X(t),Y(t))$7. When $(X(t),Y(t))$8, the expected number of critical points above threshold $(X(t),Y(t))$9 is

$$\mu_u(t)=\int_{-\infty}^{\infty} |y|\,p(u,y,t)\,dy,$$0

where $$\mu_u(t)=\int_{-\infty}^{\infty} |y|\,p(u,y,t)\,dy,$$1 is an explicit expectation involving an independent chi variable, a Wishart matrix, and a Hessian-like Gaussian matrix $$\mu_u(t)=\int_{-\infty}^{\infty} |y|\,p(u,y,t)\,dy,$$2 (Marinucci et al., 2024). For maxima, the auxiliary Gaussian field $$\mu_u(t)=\int_{-\infty}^{\infty} |y|\,p(u,y,t)\,dy,$$3, $$\mu_u(t)=\int_{-\infty}^{\infty} |y|\,p(u,y,t)\,dy,$$4, yields a second Kac–Rice formula with a density $$\mu_u(t)=\int_{-\infty}^{\infty} |y|\,p(u,y,t)\,dy,$$5. In the high-threshold regime, the density becomes, up to explicit constants, Hermite polynomials times a Gaussian density (Marinucci et al., 2024).

4. Beyond Gaussianity and beyond smooth Euclidean fields

The formula also extends far beyond smooth Gaussian fields on Euclidean domains. For piecewise smooth processes $$\mu_u(t)=\int_{-\infty}^{\infty} |y|\,p(u,y,t)\,dy,$$6 driven by deterministic flows between random jumps, the one-dimensional Kac–Rice formula takes the form

$$\mu_u(t)=\int_{-\infty}^{\infty} |y|\,p(u,y,t)\,dy,$$7

where $$\mu_u(t)=\int_{-\infty}^{\infty} |y|\,p(u,y,t)\,dy,$$8 is the deterministic velocity field and $$\mu_u(t)=\int_{-\infty}^{\infty} |y|\,p(u,y,t)\,dy,$$9 is the time-dependent density (Azaïs et al., 2017). In dimension $[t_0,t]$0, for a $[t_0,t]$1 compact hypersurface $[t_0,t]$2,

$[t_0,t]$3

with $[t_0,t]$4 the outward unit normal (Azaïs et al., 2017). These formulas hold for non-diffusive, non-stationary piecewise smooth processes under non-tangency, density continuity, and the condition that the process does not jump from or to the crossing set (Azaïs et al., 2017). A plausible implication is that Kac–Rice is fundamentally a “density times transverse speed” principle rather than a specifically Gaussian phenomenon.

A different generalization counts transverse intersections. In a regular setting, the expected cardinality of the preimage of a submanifold under a random map can be expressed as the integral of a density, and the same framework extends to other counting measures such as intersection degree (Stecconi, 2021). In the Gaussian specialization, this applies to smooth random sections of a vector bundle meeting a given submanifold of the total space, and it can be simplified by using appropriate connections (Stecconi, 2021). The same paper identifies a class of submanifolds called sub-Gaussian, for which the formula is locally finite and depends continuously on the covariance of the first jet, and it applies in particular to singularity conditions defined by jet prolongations meeting semialgebraic submanifolds of jet space (Stecconi, 2021).

5. Modern algebraic, statistical, and matrix-theoretic applications

In algebraic statistics and chemical reaction network theory, Kac–Rice can be adapted to parametrized polynomial systems that are linear in enough parameters. If

$[t_0,t]$5

with $[t_0,t]$6, then

$[t_0,t]$7

where $[t_0,t]$8 and $[t_0,t]$9 is the induced density factor (Feliu et al., 2020). This transforms expected counts of solutions into multivariate integrals that can be evaluated numerically and used to partition parameter space according to the number of solutions, with applications to multistationarity in chemical reaction networks (Feliu et al., 2020).

In adaptive regression, the Kac–Rice formula is used in the random-field sense to obtain an exact finite-sample $$\mathbb{E}\big[N_u(t_0,t)\big]=\int_{t_0}^t \mu_u(t')\,dt' =\int_{t_0}^t\int_{-\infty}^{\infty}|y|\,p(u,y,t')\,dy\,dt'.$$0-value for the global null in general regularized Gaussian regression problems. The basic Gaussian process is

$$\mathbb{E}\big[N_u(t_0,t)\big]=\int_{t_0}^t \mu_u(t')\,dt' =\int_{t_0}^t\int_{-\infty}^{\infty}|y|\,p(u,y,t')\,dy\,dt'.$$1

where $$\mathbb{E}\big[N_u(t_0,t)\big]=\int_{t_0}^t \mu_u(t')\,dt' =\int_{t_0}^t\int_{-\infty}^{\infty}|y|\,p(u,y,t')\,dy\,dt'.$$2 depends on the penalty, and the distribution of the global maximizer $$\mathbb{E}\big[N_u(t_0,t)\big]=\int_{t_0}^t \mu_u(t')\,dt' =\int_{t_0}^t\int_{-\infty}^{\infty}|y|\,p(u,y,t')\,dy\,dt'.$$3 is obtained by a Kac–Rice formula on a stratified set (Taylor et al., 2013). The resulting survival-function pivot is exactly $$\mathbb{E}\big[N_u(t_0,t)\big]=\int_{t_0}^t \mu_u(t')\,dt' =\int_{t_0}^t\int_{-\infty}^{\infty}|y|\,p(u,y,t')\,dy\,dt'.$$4 under the global null, and the framework covers the lasso, group lasso, principal components, and matrix completion (Taylor et al., 2013).

In random dynamical systems, the mean number of fixed points of a single- or multi-layer random Gaussian map is computed by Kac–Rice and reduces to a random-matrix expectation: $$\mathbb{E}\big[N_u(t_0,t)\big]=\int_{t_0}^t \mu_u(t')\,dt' =\int_{t_0}^t\int_{-\infty}^{\infty}|y|\,p(u,y,t')\,dy\,dt'.$$5 where the $$\mathbb{E}\big[N_u(t_0,t)\big]=\int_{t_0}^t \mu_u(t')\,dt' =\int_{t_0}^t\int_{-\infty}^{\infty}|y|\,p(u,y,t')\,dy\,dt'.$$6 are independent real Ginibre matrices with variances $$\mathbb{E}\big[N_u(t_0,t)\big]=\int_{t_0}^t \mu_u(t')\,dt' =\int_{t_0}^t\int_{-\infty}^{\infty}|y|\,p(u,y,t')\,dy\,dt'.$$7 (Ipsen et al., 2018). Random-matrix asymptotics then show a third-order phase transition between a phase with a single fixed point and a phase with exponentially many fixed points, and the asymptotic complexity depends only on the scaled variance parameter $$\mathbb{E}\big[N_u(t_0,t)\big]=\int_{t_0}^t \mu_u(t')\,dt' =\int_{t_0}^t\int_{-\infty}^{\infty}|y|\,p(u,y,t')\,dy\,dt'.$$8 (Ipsen et al., 2018).

A recent non-Hermitian random-matrix application starts from the characteristic polynomial

$$\mathbb{E}\big[N_u(t_0,t)\big]=\int_{t_0}^t \mu_u(t')\,dt' =\int_{t_0}^t\int_{-\infty}^{\infty}|y|\,p(u,y,t')\,dy\,dt'.$$9

and the Kac–Rice identity

$$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$00

Using the integral identity

$$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$01

this becomes a joint empirical density of an eigenvalue $$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$02 and a normalized right eigenvector $$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$03 (Fyodorov, 26 Jun 2025). The method yields explicit finite-$$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$04 joint densities for ensembles interpolating between complex and real Ginibre and for additive deformations of complex Ginibre matrices, including the emergence of weak non-reality and the structure of outliers (Fyodorov, 26 Jun 2025).

6. Assumptions, variants, and structural themes

Across these variants, Kac–Rice requires some combination of smoothness, nondegeneracy, and control of conditional laws. In the general level-set formula, the hypotheses are almost surely $$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$05 sample paths, continuous one-point densities bounded on compact sets, and conditional laws that depend continuously on the conditioning value in the $$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$06-topology (Armentano et al., 2023). In piecewise smooth processes, the hypotheses are instead non-tangency of the deterministic flow, continuity of time-dependent densities near the crossing surface, and the condition that the process does not jump from or to the crossing set (Azaïs et al., 2017). In adaptive regression, the exact pivot requires Gaussianity of the errors and a Morse-type condition ensuring a unique maximizer of the Gaussian field on the stratified index set (Taylor et al., 2013).

A recurring distinction is between formulas valid for almost every level and formulas valid for all levels. The coarea formula gives the almost-every-level identity directly, but many applications require a fixed level, especially level zero; continuity in the level variable or stronger conditional-law assumptions are then used to pass from almost every level to all levels (Berzin et al., 2022, Armentano et al., 2023). This suggests that one of the central technical issues in Kac–Rice theory is not the local Jacobian factor itself, but the global regularity needed to make level-wise conditioning stable.

Different geometric choices can materially alter the resulting statistics. In the $$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$07 polynomial setting, critical points are defined using the usual complex derivative $$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$08, not the Chern connection on $$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$09, and the paper explicitly contrasts the two: with $$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$10, the density of critical values decays like $$\mathbb{E}\big[\mathcal{H}^{D-d}(X^{-1}(u)\cap B)\big]$$11 and the critical values accumulate at infinity, whereas the Chern-connection setting leads to exponentially decaying densities (Feng et al., 2012). More broadly, the literature in these papers repeatedly uses the same blueprint—compute the relevant joint law, isolate the normal Jacobian or Hessian determinant, condition on the constraint, and integrate the resulting local density—while the ambient geometry, the notion of event being counted, and the regularity assumptions determine the precise form of the formula.