Non-Central Chi-Squared Random Field

• Non-central chi-squared random field is a stochastic model of random variables defined by weighted sums of squared shifted Gaussian components, connecting Gaussian fields with quadratic transformations.
• The analytic framework employs characteristic functions, Laplace inversion, and density maximum bounds to derive dimension-free estimates and anti-concentration measures.
• Its Hilbert space formulation and geometric analysis support applications in long-memory processes, excursion theory, and enhanced Gaussian comparison techniques.

A non-central chi-squared random field is a collection of random variables $\{Y(t): t\in T\}$ whose fixed-index marginals are non-central chi-square-type laws, typically realized as weighted sums of squared shifted Gaussian components, or more generally as linear combinations of central and non-central chi-square variables. In the most direct Gaussian representation, one writes

$$Y(t)=\sum_{k=1}^n \lambda_k(t)\bigl(Z_k(t)-a_k(t)\bigr)^2,$$

so that, for each fixed $t$, every term is a scaled non-central chi-square variable with $1$ degree of freedom and non-centrality $a_k(t)^2$. In infinite-dimensional form, the same structure appears as the squared norm of a non-centered Gaussian element in a Hilbert space, which is why the topic sits at the intersection of Gaussian random fields, quadratic forms, anti-concentration theory, and statistical modeling (Bobkov et al., 2020).

1. Definition and canonical representations

For a standard non-central chi-square distribution with $\nu$ degrees of freedom and non-centrality parameter $\Lambda$, the density is

$$f_{\chi^2_{\nu}(\Lambda)}(x) = \frac12 e^{-(x+\Lambda)/2} \left(\frac{x}{\Lambda}\right)^{\nu/4-1/2} I_{\nu/2-1}\big(\sqrt{\Lambda x}\big),\qquad x>0,$$

with $I_\alpha$ the modified Bessel function of the first kind. In the weighted-sum viewpoint used by Bobkov, Naumov, and Ulyanov, the basic non-central object is

$$W_a=\lambda_1(Z_1-a_1)^2+\cdots+\lambda_n(Z_n-a_n)^2,\qquad \lambda_1\ge\cdots\ge \lambda_n>0,$$

where $$Y(t)=\sum_{k=1}^n \lambda_k(t)\bigl(Z_k(t)-a_k(t)\bigr)^2,$$0 are independent. Each term $$Y(t)=\sum_{k=1}^n \lambda_k(t)\bigl(Z_k(t)-a_k(t)\bigr)^2,$$1 is a scaled non-central $$Y(t)=\sum_{k=1}^n \lambda_k(t)\bigl(Z_k(t)-a_k(t)\bigr)^2,$$2 variable with $$Y(t)=\sum_{k=1}^n \lambda_k(t)\bigl(Z_k(t)-a_k(t)\bigr)^2,$$3 degree of freedom and non-centrality $$Y(t)=\sum_{k=1}^n \lambda_k(t)\bigl(Z_k(t)-a_k(t)\bigr)^2,$$4, and $$Y(t)=\sum_{k=1}^n \lambda_k(t)\bigl(Z_k(t)-a_k(t)\bigr)^2,$$5 is therefore a weighted sum of independent non-central chi-square variables (Bobkov et al., 2020).

This pointwise representation extends directly to random fields. If $$Y(t)=\sum_{k=1}^n \lambda_k(t)\bigl(Z_k(t)-a_k(t)\bigr)^2,$$6 and

$$Y(t)=\sum_{k=1}^n \lambda_k(t)\bigl(Z_k(t)-a_k(t)\bigr)^2,$$7

then, after diagonalizing $$Y(t)=\sum_{k=1}^n \lambda_k(t)\bigl(Z_k(t)-a_k(t)\bigr)^2,$$8 and applying an orthogonal transformation, one obtains a representation of the form

$$Y(t)=\sum_{k=1}^n \lambda_k(t)\bigl(Z_k(t)-a_k(t)\bigr)^2,$$9

so each marginal $t$0 is exactly of the weighted non-central chi-square type studied in the finite-dimensional theory. The same mechanism holds for non-centered Gaussian elements $t$1 in a Hilbert space, for which

$t$2

with $t$3; if $t$4 is indexed by $t$5, then $t$6 is a non-central chi-square-type random field (Bobkov et al., 2020).

A broader analytic framework replaces the one-degree-of-freedom building blocks by general terms $t$7, allowing arbitrary degrees of freedom and mixed central/non-central components. In that sense, a non-central chi-squared random field is often understood pointwise as a field whose marginals are linear combinations of central and non-central chi-square variables arising from quadratic forms in Gaussian fields (Kume et al., 2023).

2. Marginal densities and analytic representations

The marginal density of a non-central chi-square-type field is rarely available in a simple closed form once several weighted components are summed. For a single term $t$8, the density can be written explicitly, but the finite-sum theory in (Bobkov et al., 2020) proceeds instead through characteristic functions and inversion formulas. For

$t$9

the modulus of the characteristic function satisfies

$1$0

and the density maximum is controlled through

$1$1

This separates the effect of the weights $1$2 from the shifts $1$3, and it is the basis of the dimension-free density bounds discussed below (Bobkov et al., 2020).

A complementary analytic route is the branch cut approach to Laplace inversion. For a general linear combination

$1$4

the density is represented by inverting a Laplace transform whose integrand contains factors $1$5. When odd multiplicities occur, these factors create branch cuts on the real axis, and contour deformation converts the Bromwich integral into a finite number of feasible univariate integrals. In the central case with $1$6, the density becomes a finite alternating sum of one-dimensional integrals over intervals determined by the ordered $1$7; general degrees of freedom and non-centralities are then incorporated by recursive differentiation with respect to $1$8 and $1$9 (Kume et al., 2023).

Tail functionals of non-central chi-square marginals are naturally expressed through Marcum and Nuttall $a_k(t)^2$0-functions. The partial moments

$a_k(t)^2$1

encode truncated moments and exceedance functionals of non-central chi-square-type variables. For a random field with pointwise non-central chi-square marginals, exceedance probabilities $a_k(t)^2$2 and tail moments $a_k(t)^2$3 are therefore computable pointwise through the same special-function machinery (Gil et al., 2013).

3. Density maxima, anti-concentration, and dimension-free bounds

A central quantitative question is how peaked the marginal density of a non-central chi-square-type field can be. For

$a_k(t)^2$4

define

$a_k(t)^2$5

Bobkov, Naumov, and Ulyanov prove a two-sided bound for the density maximum $a_k(t)^2$6: $a_k(t)^2$7 and, under the typical condition $a_k(t)^2$8,

$a_k(t)^2$9

Thus the marginal density peak is of order $\nu$0, with matching dependence on the parameters and only universal constants separating upper and lower bounds (Bobkov et al., 2020).

In the central case

$\nu$1

the corresponding parameters are

$\nu$2

and there exist absolute constants $\nu$3 such that

$\nu$4

The dependence is dimension-free: no explicit $\nu$5 appears, only quadratic sums of the weights (Bobkov et al., 2020).

For a non-central chi-squared random field $\nu$6 with pointwise representation

$\nu$7

the same theorem yields, for each fixed $\nu$8,

$\nu$9

where

$\Lambda$0

This controls the peak of the marginal density at each index. If $\Lambda$1 and $\Lambda$2 vary smoothly or stay bounded away from $\Lambda$3 and $\Lambda$4, this suggests uniform anti-concentration over $\Lambda$5, for example

$\Lambda$6

which is precisely the type of bound needed in Gaussian comparison and small-ball analysis (Bobkov et al., 2020).

4. Hilbert-space formulation and probabilistic uses

The Hilbert-space formulation is structurally fundamental. If $\Lambda$7 is a separable Hilbert space and $\Lambda$8 is a centered Gaussian element with covariance operator $\Lambda$9, there exists an orthonormal basis $$f_{\chi^2_{\nu}(\Lambda)}(x) = \frac12 e^{-(x+\Lambda)/2} \left(\frac{x}{\Lambda}\right)^{\nu/4-1/2} I_{\nu/2-1}\big(\sqrt{\Lambda x}\big),\qquad x>0,$$0 such that

$$f_{\chi^2_{\nu}(\Lambda)}(x) = \frac12 e^{-(x+\Lambda)/2} \left(\frac{x}{\Lambda}\right)^{\nu/4-1/2} I_{\nu/2-1}\big(\sqrt{\Lambda x}\big),\qquad x>0,$$1

and

$$f_{\chi^2_{\nu}(\Lambda)}(x) = \frac12 e^{-(x+\Lambda)/2} \left(\frac{x}{\Lambda}\right)^{\nu/4-1/2} I_{\nu/2-1}\big(\sqrt{\Lambda x}\big),\qquad x>0,$$2

For a shifted element $$f_{\chi^2_{\nu}(\Lambda)}(x) = \frac12 e^{-(x+\Lambda)/2} \left(\frac{x}{\Lambda}\right)^{\nu/4-1/2} I_{\nu/2-1}\big(\sqrt{\Lambda x}\big),\qquad x>0,$$3, one obtains

$$f_{\chi^2_{\nu}(\Lambda)}(x) = \frac12 e^{-(x+\Lambda)/2} \left(\frac{x}{\Lambda}\right)^{\nu/4-1/2} I_{\nu/2-1}\big(\sqrt{\Lambda x}\big),\qquad x>0,$$4

so the squared norm of a non-centered Gaussian element is an infinite weighted sum of non-central chi-square variables (Bobkov et al., 2020).

This representation is the bridge to Gaussian comparison on balls and to non-asymptotic central limit theory in finite and infinite dimensions. The density maximum bounds for $$f_{\chi^2_{\nu}(\Lambda)}(x) = \frac12 e^{-(x+\Lambda)/2} \left(\frac{x}{\Lambda}\right)^{\nu/4-1/2} I_{\nu/2-1}\big(\sqrt{\Lambda x}\big),\qquad x>0,$$5 feed into dimension-free estimates on the Kolmogorov distance between distributions of Gaussian norms, and hence into comparison of ball probabilities $$f_{\chi^2_{\nu}(\Lambda)}(x) = \frac12 e^{-(x+\Lambda)/2} \left(\frac{x}{\Lambda}\right)^{\nu/4-1/2} I_{\nu/2-1}\big(\sqrt{\Lambda x}\big),\qquad x>0,$$6 and $$f_{\chi^2_{\nu}(\Lambda)}(x) = \frac12 e^{-(x+\Lambda)/2} \left(\frac{x}{\Lambda}\right)^{\nu/4-1/2} I_{\nu/2-1}\big(\sqrt{\Lambda x}\big),\qquad x>0,$$7. The same source states that these refined density bounds also support dimension-free Gaussian comparison bounds and non-asymptotic CLTs for sums of random vectors in $$f_{\chi^2_{\nu}(\Lambda)}(x) = \frac12 e^{-(x+\Lambda)/2} \left(\frac{x}{\Lambda}\right)^{\nu/4-1/2} I_{\nu/2-1}\big(\sqrt{\Lambda x}\big),\qquad x>0,$$8 or in Banach and Hilbert spaces, and that they improve over Pinsker-type bounds via KL divergence by providing sharper anti-concentration behavior for norms (Bobkov et al., 2020).

A logarithmic transformation yields another derived field of interest. If $$f_{\chi^2_{\nu}(\Lambda)}(x) = \frac12 e^{-(x+\Lambda)/2} \left(\frac{x}{\Lambda}\right)^{\nu/4-1/2} I_{\nu/2-1}\big(\sqrt{\Lambda x}\big),\qquad x>0,$$9 pointwise and one defines $I_\alpha$0, then the moments of $I_\alpha$1 are computable via the Poisson mixture representation of the non-central chi-square law. In particular,

$I_\alpha$2

so log non-central chi-square fields provide an analytically tractable model for log-variance or log-energy processes derived from quadratic Gaussian structure (Pav, 2015).

5. Related constructions and terminological variants

Several adjacent constructions broaden the notion of a non-central chi-squared random field. One is the zero-degree-of-freedom non-central chi-square distribution $I_\alpha$3, defined through the Poisson mixture

$I_\alpha$4

with a point mass at zero

$I_\alpha$5

Its scaled form satisfies

$I_\alpha$6

In ensemble postprocessing, a station–time indexed family $I_\alpha$7 is already a non-central chi-square random field in the marginal sense, and the paper explicitly points to future research on multivariate postprocessing methods making use of this distribution (Groß et al., 2024).

Another extension normalizes several non-central chi-square components by their sum, producing a non-central Dirichlet law. If

$I_\alpha$8

then $I_\alpha$9 has a non-central Dirichlet distribution. Conditionally on the Multi-Poisson mixture counts, it is a finite mixture of ordinary Dirichlet laws, and the paper derives both a conditional independence statement $$W_a=\lambda_1(Z_1-a_1)^2+\cdots+\lambda_n(Z_n-a_n)^2,\qquad \lambda_1\ge\cdots\ge \lambda_n>0,$$0 and a stochastic representation of $$W_a=\lambda_1(Z_1-a_1)^2+\cdots+\lambda_n(Z_n-a_n)^2,\qquad \lambda_1\ge\cdots\ge \lambda_n>0,$$1 as a convex combination of a central Dirichlet component and a purely non-central component. This supplies a natural simplex-valued field construction whenever the underlying non-central chi-square variables are indexed by space or time (Orsi, 2021).

A related but distinct field-building mechanism comes from products of correlated Gaussian variables. If $$W_a=\lambda_1(Z_1-a_1)^2+\cdots+\lambda_n(Z_n-a_n)^2,\qquad \lambda_1\ge\cdots\ge \lambda_n>0,$$2 with jointly Gaussian $$W_a=\lambda_1(Z_1-a_1)^2+\cdots+\lambda_n(Z_n-a_n)^2,\qquad \lambda_1\ge\cdots\ge \lambda_n>0,$$3, then the marginal law of $$W_a=\lambda_1(Z_1-a_1)^2+\cdots+\lambda_n(Z_n-a_n)^2,\qquad \lambda_1\ge\cdots\ge \lambda_n>0,$$4 is a scaled difference of two independent non-central chi-square variables. The paper on the noncentral chi-square difference distribution shows that sums of independent copies of such products admit an exact representation as

$$W_a=\lambda_1(Z_1-a_1)^2+\cdots+\lambda_n(Z_n-a_n)^2,\qquad \lambda_1\ge\cdots\ge \lambda_n>0,$$5

with independent $$W_a=\lambda_1(Z_1-a_1)^2+\cdots+\lambda_n(Z_n-a_n)^2,\qquad \lambda_1\ge\cdots\ge \lambda_n>0,$$6 non-central chi-square. This does not produce a non-central chi-square field in the strict positive sense, but it yields a closely related quadratic-field class with explicit pdfs, moments, sign probabilities, and Stein operators (Gaunt, 2024).

A final terminological caution is essential. In the long-memory literature on chi-squared fields, “non-central” may refer not to non-central $$W_a=\lambda_1(Z_1-a_1)^2+\cdots+\lambda_n(Z_n-a_n)^2,\qquad \lambda_1\ge\cdots\ge \lambda_n>0,$$7 marginals but to non-central limit theorems. In particular, functionals of central chi-squared random fields subordinated to long-range dependent Gaussian fields can converge to Rosenblatt-type non-Gaussian limits, and that usage is distinct from the marginal non-centrality parameter of a $$W_a=\lambda_1(Z_1-a_1)^2+\cdots+\lambda_n(Z_n-a_n)^2,\qquad \lambda_1\ge\cdots\ge \lambda_n>0,$$8 law (Leonenko et al., 2015).

6. Geometry, peaks, excursions, and asymptotic field theory

Excursion and sojourn functionals provide one of the main interfaces between non-central chi-squared random fields and random-field geometry. On compact two-point homogeneous spaces, the central field

$$W_a=\lambda_1(Z_1-a_1)^2+\cdots+\lambda_n(Z_n-a_n)^2,\qquad \lambda_1\ge\cdots\ge \lambda_n>0,$$9

has $$Y(t)=\sum_{k=1}^n \lambda_k(t)\bigl(Z_k(t)-a_k(t)\bigr)^2,$$00 marginals, and the paper on time-dependent $$Y(t)=\sum_{k=1}^n \lambda_k(t)\bigl(Z_k(t)-a_k(t)\bigr)^2,$$01-random fields studies the large-$$Y(t)=\sum_{k=1}^n \lambda_k(t)\bigl(Z_k(t)-a_k(t)\bigr)^2,$$02 behavior of the time-averaged excursion volume of $$Y(t)=\sum_{k=1}^n \lambda_k(t)\bigl(Z_k(t)-a_k(t)\bigr)^2,$$03. The limiting law depends on the temporal memory parameters: short memory gives Gaussian limits, whereas long memory yields composite Rosenblatt random variables. The same paper states that a non-central extension would take

$$Y(t)=\sum_{k=1}^n \lambda_k(t)\bigl(Z_k(t)-a_k(t)\bigr)^2,$$04

so that, for each $$Y(t)=\sum_{k=1}^n \lambda_k(t)\bigl(Z_k(t)-a_k(t)\bigr)^2,$$05,

$$Y(t)=\sum_{k=1}^n \lambda_k(t)\bigl(Z_k(t)-a_k(t)\bigr)^2,$$06

with the same harmonic-analysis framework available in principle (Caponera et al., 2024).

Peak theory offers another route to effective non-centrality. For a central chi-squared field

$$Y(t)=\sum_{k=1}^n \lambda_k(t)\bigl(Z_k(t)-a_k(t)\bigr)^2,$$07

built from centered, stationary, isotropic Gaussian fields, conditioning on a stationary point of specified amplitude biases the underlying Gaussian components. One component acquires a non-zero mean and modified covariance, so the neighborhood of the conditioned point becomes a generalized, effectively non-central chi-square random field. This is the basis for the expected radial profile, spherical-harmonic mode statistics, and asphericity measures developed for peaks in $$Y(t)=\sum_{k=1}^n \lambda_k(t)\bigl(Z_k(t)-a_k(t)\bigr)^2,$$08 fields (Bloomfield et al., 2018). Complementarily, the number density of stationary points at fixed field amplitude in central $$Y(t)=\sum_{k=1}^n \lambda_k(t)\bigl(Z_k(t)-a_k(t)\bigr)^2,$$09 fields is handled through Kac–Rice integrals over the field, gradient, and Hessian variables, yielding exact asymptotics in several limits and a numerical framework for general spectra (Bloomfield et al., 2016).

Maximum theory is mostly developed for central fields, but the extension mechanism is explicit. On multi-dimensional lattice index sets, the maximum of a central chi-square random field with direct-product covariance admits high-threshold approximations from nonlinear renewal theory and the volume-of-tube method. The same source states that the constructions extend naturally to non-central chi-square fields by modifying the mean of the underlying Gaussian field. In practice, this separates two tasks: the null extreme-value problem is central, while alternatives correspond to locally non-central fields whose mean structure encodes signal (Kuriki et al., 2010).

Taken together, these strands show that the subject is not a single model but a family of Gaussian-quadratic field constructions. The core object is the pointwise non-central chi-square law generated by squared shifted Gaussian modes; the main analytic themes are exact or semi-exact density representations, sharp density-peak bounds, excursion and anti-concentration estimates, and geometric descriptions of peaks and extremes. The unifying principle is that non-centrality enters either explicitly through mean shifts in Gaussian components or implicitly through conditioning, and in both cases the resulting field remains governed by Gaussian covariance structure together with quadratic transformation rules (Bobkov et al., 2020).