---
title: Functional Breuer–Major Theorem
url: https://www.emergentmind.com/topics/functional-breuer-major-theorem
type: topic
---

# Functional Breuer–Major Theorem

Searching arXiv for recent and foundational papers on the functional Breuer–Major theorem and related quantitative/Hilbert-valued extensions.
The functional Breuer–Major theorem is the path-space extension of the Breuer–Major central limit theorem for nonlinear functionals of stationary Gaussian inputs. In its discrete Gaussian form, one considers normalized partial sums
\[
Y_n(t)=\frac{1}{\sqrt n}\sum_{i=0}^{\lfloor nt\rfloor-1}\varphi(X_i), \qquad t\in[0,1],
\]
where \(X=\{X_n\}_{n\in\mathbb Z}\) is a zero-mean stationary Gaussian sequence with covariance \(\rho\), and \(\varphi\) has Hermite rank \(d\). Under the short-memory condition \(\sum_{r\in\mathbb Z}|\rho(r)|^d<\infty\), the finite-dimensional distributions converge to those of a scaled Brownian motion, and, under an additional moment assumption \(E[|\varphi(X_0)|^p]<\infty\) for some \(p>2\), one obtains weak convergence in path space. Modern formulations treat convergence in \(D([0,1])\), \(C([0,\infty))\), Hilbert-valued path spaces, Hölder topologies, and rough-path topologies, and also provide quantitative rates via Malliavin–Stein, Dirichlet-structure, and second-order Gaussian Poincaré methods [1808.02378, 1905.05127].

## 1. Discrete Gaussian formulation

Let \(X=\{X_n\}_{n\in\mathbb Z}\) be a zero-mean stationary Gaussian sequence with
\[
E[X_nX_{n+r}] = \rho(r), \qquad \rho(0)=1,
\]
and let \(\varphi:\mathbb R\to\mathbb R\) belong to \(L^2(\mathbb R,\gamma)\), where \(\gamma\) is the standard Gaussian law. Writing the Hermite expansion
\[
\varphi(x)=\sum_{q=d}^\infty c_qH_q(x),
\]
the integer \(d\ge 1\) is the Hermite rank, meaning \(c_d\neq 0\) and \(c_0=\cdots=c_{d-1}=0\). The normalized partial-sum process is
\[
Y_n(t)=\frac{1}{\sqrt n}\sum_{i=0}^{\lfloor nt\rfloor-1}\varphi(X_i), \qquad t\in[0,1].
\]

The classical Breuer–Major theorem yields convergence of the finite-dimensional distributions of \(Y_n\) to those of \(\sigma W\), where \(W\) is a standard Brownian motion, under the summability condition
\[
\sum_{r\in\mathbb Z}|\rho(r)|^d<\infty.
\]
The functional theorem strengthens this by establishing convergence in \(D([0,1])\) endowed with the Skorohod topology, provided one assumes in addition that \(E[|\varphi(X_0)|^p]<\infty\) for some \(p>2\) [1808.02378].

A Hilbert-valued formulation replaces path-space arguments by embedding the process into a separable Hilbert space \(K\) containing \(C([0,1])\) or \(D([0,1])\) isometrically, for example \(K=L^2([0,1])\). In that setting one views \(t\mapsto S_n(t)\) as a \(K\)-valued random variable and obtains
\[
S_n \Rightarrow \sigma W
\]
in distribution in \(C([0,1])\), hence in \(K\), again under \(\sum_{k\in\mathbb Z}|r(k)|^m<\infty\) for Hermite rank \(m\) [1905.05127].

## 2. Variance structure and the role of Hermite rank

The limiting variance is given by the Breuer–Major formula
\[
\sigma^2
=\Var\bigl(\varphi(X_0)\bigr)
+2\sum_{r=1}^\infty \Cov\bigl(\varphi(X_0),\varphi(X_r)\bigr)
=\sum_{q=d}^\infty q!\,c_q^2\sum_{r\in\mathbb Z}\rho(r)^q.
\]
Under \(\sum_r|\rho(r)|^d<\infty\), this series converges because \(|\rho(r)|\le 1\) [1808.02378].

The Hermite rank is the decisive threshold in the dependence condition. It is the smallest chaos order that contributes nontrivially, and therefore the minimal power of the covariance that must be summable. In the continuous stationary version,
\[
Z_\varepsilon(t)=\sqrt{\varepsilon}\int_0^{t/\varepsilon}f(Y(s))\,ds,
\]
the corresponding condition is
\[
\int_{\mathbb R}|\rho(s)|^d\,ds<\infty,
\]
and the limiting variance becomes
\[
\sigma^2=\sum_{q=d}^\infty q!\,c_q^2\int_{\mathbb R}\rho(s)^q\,ds.
\]
In that formulation the condition \(\int_{\mathbb R}|\rho(s)|^d\,ds<\infty\) is both necessary and sufficient for nondegeneracy of \(\sigma^2\) [1807.09740].

The same structural principle persists in examples. For fractional Gaussian noise \(X_k=B_{k+1}-B_k\) of Hurst index \(H\in(0,\tfrac12)\), the functional theorem applies when \(H<1-\tfrac1{2d}\). For \(H>1-\tfrac1{2d}\), one enters the non-Gaussian Hermite regime of Taqqu, so the Brownian conclusion is no longer the relevant limit statement [1808.02378].

## 3. Tightness, Malliavin calculus, and proof architecture

A central difficulty is that finite-dimensional convergence does not imply convergence in \(D([0,1])\) or \(C([0,\infty))\). The functional Breuer–Major theorem is therefore a tightness theorem as much as a central limit theorem. In the discrete Gaussian case, the proof proceeds by combining the classical Breuer–Major finite-dimensional convergence with a Billingsley-type tightness criterion. It suffices to prove that for some \(p>2\),
\[
\big\|Y_n(t)-Y_n(s)\big\|_{L^p(\Omega)}
\le C\Bigl(\frac{\lfloor nt\rfloor-\lfloor ns\rfloor}{n}\Bigr)^{1/2},
\qquad 0\le s<t\le 1.
\]
This yields the required two-parameter moment condition and hence tightness in \(D([0,1])\) [1808.02378].

The key analytical input is a Malliavin representation of each summand as an iterated divergence. If \(d\) is the Hermite rank and \(e_i\) is the isonormal kernel of \(X_i\), then
\[
\varphi(X_i)=\delta^d\bigl(\varphi_d(X_i)\,e_i^{\otimes d}\bigr),
\]
where
\[
\varphi_d(x)=\sum_{q=d}^\infty c_qH_{q-d}(x).
\]
Meyer inequalities imply boundedness of \(\delta^d\) from a Sobolev-type norm into \(L^p(\Omega)\), and hypercontractivity yields uniform \(L^p\)-bounds on the relevant chaos components. Together with stationarity and \(\sum|\rho(r)|^d<\infty\), one recovers the increment estimate needed for tightness [1808.02378].

The same scheme appears in continuous and vector-valued settings. In the continuous stationary theorem one writes
\[
f\bigl(Y(u)\bigr)=\delta^d\Bigl(\bigl(-D\,L^{-1}\bigr)^d[f(Y(u))]\Bigr)
\]
and then applies Meyer inequalities and Kolmogorov–Chentsov to obtain tightness in \(C([0,\infty))\) [1807.09740]. For Gaussian vector fields \(\xi:\Omega\times\mathbb R^n\to\mathbb R^m\), one proves a \(k\)-fold divergence representation
\[
G(\xi(x))=\delta^k(u_k(x)),
\]
with \(u_k\) involving shift operators \(T_{i_1\ldots i_k}G(\xi(x))\), and again derives
\[
E|Z_s(t_2)-Z_s(t_1)|^p \le C_p\,|t_2-t_1|^{p/2},
\]
which yields tightness in \(C([0,\infty))\) [1901.02317].

## 4. Quantitative functional CLTs and Hilbert-valued formulations

A major development is the transition from qualitative weak convergence to explicit rates in functional metrics. Bourguin–Campese formulate the theorem in a separable Hilbert space \(K\) and work with the distance
\[
d_2(U,V)=\sup_{h\in C^2_1(K)} |E[h(U)]-E[h(V)]|,
\]
where \(C^2_1(K)\) consists of twice-Fréchet differentiable test functions with
\[
\|Dh\|_\infty+\|D^2h\|_\infty\le 1.
\]
This metric metrizes weak convergence in \(K\) [1905.05127].

Their framework combines an infinite-dimensional version of Stein’s method, due to Shih, with a diffusive Dirichlet structure and \(\Gamma\)-calculus. For a centered Gaussian \(Z\) on \(K\) with covariance \(S\), and a \(K\)-valued random variable \(F\), the carré-du-champ bound takes the form
\[
d_2(F,Z)\le C\cdot
\big\|\Gamma(F,-L^{-1}F)-S\big\|_{L^2(\Omega;\ell_2(K))}.
\]
Expanding
\[
F=\sum_{p=0}^\infty I_p(f_p)
\]
into Wiener–Itô chaos then leads to fourth-moment and contraction estimates in terms of \(f_p\otimes_r f_q\) [1905.05127].

Applied to the partial-sum process
\[
S_n(t)=\frac1{\sqrt n}\sum_{k=1}^{\lfloor nt\rfloor} g(X_k),
\]
this yields a fully quantitative functional Breuer–Major theorem. If
\[
r(k)=|k|^{-\alpha}\ell(k), \qquad \alpha>\frac1m,
\]
with \(\ell\) slowly varying at \(\infty\), then there exists \(C<\infty\) such that
\[
d_2(S_n,\sigma W)\le C\,R(n)=O\bigl(n^{-\gamma}\bigr),
\qquad
\gamma=\min\Bigl\{\frac12,\frac{m\alpha-1}{2}\Bigr\},
\]
with more refined cases when \(\alpha\in(1/m,1)\) [1905.05127].

For \(g=H_p\) and \(r(k)=|k|^{-a}\ell(k)\) with \(a>1/p\), this recovers the rates of Nourdin-Peccati–Podolskij:
- if \(a<1\), then \(R(n)=n^{-1/2}\);
- if \(a\in(1,(2p-1)/p]\), then \(R(n)=n^{-(ap+1)/2}\);
- if \(a>(2p-1)/p\), then \(R(n)=n^{-(p-1/2)}\) [1905.05127].

Related quantitative work includes contraction-based bounds for the one-dimensional Breuer–Major CLT on Wiener space, where the rate depends on the Hermite rank and the chaotic gap, and convex-distance bounds for the finite-dimensional distributions of the functional theorem [1610.01456, 2001.02188].

## 5. Extensions across settings and topologies

The functional Breuer–Major paradigm now spans several ambient spaces and noise structures. In the continuous stationary setting,
\[
Z_\varepsilon(t)=\sqrt{\varepsilon}\int_0^{t/\varepsilon}f(Y(s))\,ds
\]
converges in \(C([0,\infty))\) under \(\int_{\mathbb R}|\rho(s)|^d\,ds<\infty\) and \(f\in L^p(\mathbb R,\gamma)\) for some \(p>2\) [1807.09740]. For Gaussian vector fields \(\xi:\Omega\times\mathbb R^n\to\mathbb R^m\), one obtains
\[
Z_s(\cdot)\Rightarrow \sigma W(\cdot)
\quad\text{in } C([0,\infty)),
\]
assuming \(G\in L^p(\mathbb R^m,\gamma_m)\) for some \(p>2\) and \(r_{i,j}\in L^d(\mathbb R^n)\) [1901.02317].

A genuinely infinite-dimensional version appears in the Hilbert space-valued theorem for stationary Gaussian processes \(\{X_k\}_{k\in\mathbb Z}\) with values in \(\mathcal H_1\) and measurable \(G:\mathcal H_1\to\mathcal H_2\). If \(q=\mathrm{rk}(G)\) and
\[
\sum_{v\in\mathbb Z}
\Bigl(\sup_{r\ge 1}\sum_{s\ge 1}|\rho_{rs}(v)|\Bigr)^q<\infty,
\]
then
\[
S_n=\frac1{\sqrt n}\sum_{k=1}^n \bigl(G[X_k]-E\,G[X_k]\bigr)
\Rightarrow Z
\quad\text{in }\mathcal H_2,
\]
and the continuous-time version
\[
V_n(t)=\frac1{\sqrt n}\sum_{k=1}^{\lfloor nt\rfloor}\bigl(G[X_k]-E\,G[X_k]\bigr)
\]
converges in \(\mathcal L^2([0,1])\otimes\mathcal H_2\) to \(B\otimes Z\) [2405.11452].

A second-order Gaussian Poincaré approach gives another Hilbert-valued functional theorem. With
\[
F_T(r)=\frac1{\sqrt T}\int_{-rT}^{rT}\bigl(f(Y_t)-E[f(Y_t)]\bigr)\,dt,
\qquad r\in[0,1],
\]
and an envelope condition on the kernels \(h_t\), one has
\[
d_2(F_T,\mathcal B_T)\le \frac{C}{\sqrt T},
\]
where \(\mathcal B_T\) is a centered Gaussian random element in \(L^2([0,1])\) with the same covariance operator as \(F_T\). Moreover \(\mathcal B_T\) converges in \(d_2\) to \(\sigma B\), so \(d_2(F_T,\sigma B)\to 0\) [2506.13571].

The theorem also extends beyond Gaussian input. In the Poisson setting, if \(\eta\) is a Poisson point process on \(\mathbb R\), \(X_j=I_1(\psi_j)\), \(|\psi(x)|\le C_\psi(1+|x|)^{-\alpha}\) with \(\alpha>\tfrac12+\tfrac1{2d}\), and \(\phi\in C^{\gamma_0}_{M_0}\) with \(\gamma_0>d_\alpha+1\), then the \(d\)-th-and-higher-order chaos truncation \(\mathcal T^{\ge d}Y_n\) converges in law in \(D([0,1])\) to \(\mu B\). Tightness is proved by an \(L^p\)-spectral-gap inequality for Poisson functionals [2510.26216].

Recent work strengthens the topology. In the summable-covariance regime, interpolated partial sums
\[
S_n(t)= n^{-1/2}\sum_{k=1}^{\lfloor nt\rfloor}g(X_k)
+\bigl(nt-\lfloor nt\rfloor\bigr)n^{-1/2}g(X_{\lfloor nt\rfloor+1})
\]
converge to \(\sigma_gB\) in \(C^\alpha([0,1])\) for every \(\alpha<1/2\), assuming \(g\in L^p(\gamma)\) for some \(p>6\) and \(\sum_{h\in\mathbb Z}|r(h)|<\infty\) [2606.04179]. A rough-path version lifts the theorem to the càdlàg rough-path space \(\mathcal D^{r\text{-var}}([0,1];\mathbb R^m)\), with a Brownian rough-path limit carrying explicit symmetric and antisymmetric area corrections [2602.16615].

## 6. Optimality, misconceptions, and non-Gaussian boundaries

A persistent misconception is that the classical Breuer–Major theorem already implies path-space convergence. It does not. The classical theorem gives finite-dimensional convergence under \(L^2\)-integrability and covariance summability, but functional convergence requires tightness, and tightness cannot in general be obtained from mere \(L^2\)-bounds because one needs control of higher moments of increments [1808.02378, 1807.09740].

The additional assumption \(\varphi\in L^p(\gamma)\) or \(f\in L^p(\gamma)\) for some \(p>2\) is therefore not a cosmetic strengthening. In the discrete functional theorem it is described as a “sufficient (and almost necessary) natural condition,” and in the detailed exposition it is said to be minimal in the sense that it exactly produces the moment power \(p/2>1\) needed in Billingsley’s criterion [1808.02378]. Earlier criteria, such as those of Ben Hariz and Chambers–Slud, required rather unnatural summability conditions on all Hermite coefficients; the Malliavin–Meyer method reduces the problem to a single extra \(L^p\)-assumption [1808.02378].

Another conceptual boundary concerns the dependence regime. The Brownian limit is tied to the summable-covariance regime \(\sum|\rho|^d<\infty\) or \(\int|\rho|^d<\infty\). When this fails, the limit need not remain Gaussian. In the continuous non-stationary self-similar setting, the non-central case \(\alpha>2-\tfrac1d\) leads to a Hermite process of order \(d\); in the critical case \(\alpha=2-\tfrac1d\), the Gaussian limit survives only after logarithmic renormalization [1807.09740]. Similarly, for fractional Gaussian noise with \(H>1-\tfrac1{2d}\), one leaves the Brownian regime and enters the non-Gaussian Hermite regime [1808.02378].

Finally, the modern literature clarifies that the functional theorem is not tied to the Skorohod topology. Hölder convergence is strictly stronger than Skorohod \(J_1\) convergence, and rough Hölder or rough-path convergence is stronger still. This matters because many operations on paths, including Young integration and solution maps of differential equations, are continuous in Hölder or rough Hölder topologies but not in Skorohod topology [2606.04179]. The functional Breuer–Major theorem has thus evolved from a pathwise strengthening of a classical CLT into a family of invariance principles across Gaussian, Poisson, Hilbert-valued, and rough-path frameworks.

Source: https://www.emergentmind.com/topics/functional-breuer-major-theorem