Central Limit and Brownian Bridge

• Central Limit Theorem is a foundational result in probability, showing that normalized sums of random variables converge to a normal distribution.
• Brownian bridge is a Gaussian process defined by fixed endpoints, emerging in empirical process limits and nonparametric deconvolution.
• Recent research extends these concepts to dependent data and non-i.i.d. cases, linking limit theorems with covariance structure and functional geometry.

The central limit theorem (CLT) and Brownian bridge are foundational concepts in probability theory and statistical inference, with significant roles in both classical and modern high-dimensional statistics. The CLT typically describes the convergence in distribution of suitably normalized sums of random variables to the normal distribution. The Brownian bridge, a fundamental example of a Gaussian process, characteristically emerges as the limit in various empirical and nonparametric function estimation problems, especially under constraints such as fixed initial and terminal conditions. Contemporary research extends these concepts to dependent data, stochastic processes, deconvolution problems, and non-i.i.d. settings, revealing intricate connections among limit theorems, covariance structures, and the geometry of function classes.

1. Central Limit Theorems for Stochastic Processes

In the classical setting, the CLT asserts that normalized sums of i.i.d. mean-zero random variables with finite variances converge in distribution to a standard Gaussian random variable. For stochastic processes, analogous results can be formulated, often yielding convergence in function space to a Gaussian process with a prescribed covariance structure.

For processes defined via stochastic differential equations, such as the α-Brownian bridge,

$$dX_t = -\alpha\frac{X_t}{T-t}dt + dW_t, \quad X_0 = 0,~ 0 \leq t < T,$$

the problem of statistical inference centers on the parameter $\alpha$. The continuous-time maximum likelihood estimator (MLE) for $\alpha$ admits a closed-form involving integrals of the observed process, exploiting the explicit solution representation: $X_t = (T-t)^\alpha \int_0^t (T-s)^{-\alpha} dW_s,$ where $W_t$ is standard Brownian motion (Es-Sebaiy et al., 2020).

2. Brownian Bridge and Its Generalizations

The Brownian bridge $(B_t^0)_{t\in[0,1]}$ is a Gaussian process with mean zero and covariance $\mathrm{Cov}(B_s^0, B_t^0) = s \wedge t - st$, representing Brownian motion conditioned to return to zero at time $1$. In statistical limit theory, the Brownian bridge naturally arises as the weak limit of the normalized empirical process: $$\left\{ \sqrt{n} \left( F_n(t) - F(t) \right) : t \in \mathbb{R} \right\}$$ under the null hypothesis, where $F_n$ is the empirical cumulative distribution function (CDF), and $\alpha$0 is the true CDF.

In deconvolution problems involving the model $\alpha$1, with $\alpha$2 an additive measurement error of known distribution, uniform central limit theorems (UCLTs) for estimators of linear functionals of the unknown $\alpha$3 yield Gaussian process limits whose covariance is governed by both the "test" function (functional of $\alpha$4) and the deconvolution operator. These limits are "generalized Brownian bridges," constructed by filtering indicator kernels through the inverse of the noise characteristic function in Fourier space (Söhl et al., 2012).

3. Asymptotic Normality and Berry–Esseen Bounds for the α-Brownian Bridge

For the α-Brownian bridge, the estimator $\alpha$5 admits the representation

$\alpha$6

The Fisher information accumulated until time $\alpha$7 is

$\alpha$8

which diverges only logarithmically. The normalized estimator satisfies, as $\alpha$9,

$\alpha$0

This is a manifestation of local asymptotic normality (LAN) in the context of diffusion processes, with the slow $\alpha$1-rate reflecting the singularity at the terminal time (Es-Sebaiy et al., 2020).

Berry–Esseen-type results provide quantitative rates of convergence in the Kolmogorov distance. For the MLE in the α-Brownian bridge model, the optimal rate is

$\alpha$2

where $\alpha$3 is the standard normal CDF. Matching upper and lower bounds demonstrate the minimax optimality of this logarithmic rate, in contrast to the $\alpha$4 rate in classical i.i.d. settings (Es-Sebaiy et al., 2020).

4. Uniform Central Limit Theorems and the Generalized Brownian Bridge

In nonparametric deconvolution, one estimates linear functionals of $\alpha$5 using kernel deconvolution estimators constructed from noisy observations $\alpha$6. Under suitable smoothness conditions on $\alpha$7 and the functional $\alpha$8, and provided the order of ill-posedness $\alpha$9 of the noise is not too large relative to the smoothness of $X_t = (T-t)^\alpha \int_0^t (T-s)^{-\alpha} dW_s,$0 (specifically, if $X_t = (T-t)^\alpha \int_0^t (T-s)^{-\alpha} dW_s,$1), the process

$X_t = (T-t)^\alpha \int_0^t (T-s)^{-\alpha} dW_s,$2

converges in $X_t = (T-t)^\alpha \int_0^t (T-s)^{-\alpha} dW_s,$3 to a mean-zero Gaussian process $X_t = (T-t)^\alpha \int_0^t (T-s)^{-\alpha} dW_s,$4 with covariance

$X_t = (T-t)^\alpha \int_0^t (T-s)^{-\alpha} dW_s,$5

where $X_t = (T-t)^\alpha \int_0^t (T-s)^{-\alpha} dW_s,$6 is the "deconvolution score-function" $X_t = (T-t)^\alpha \int_0^t (T-s)^{-\alpha} dW_s,$7. In the absence of error ($X_t = (T-t)^\alpha \int_0^t (T-s)^{-\alpha} dW_s,$8), this reduces to the classical Brownian bridge indexed by $X_t = (T-t)^\alpha \int_0^t (T-s)^{-\alpha} dW_s,$9 (Söhl et al., 2012).

A key aspect is the analysis of the deconvolution operator as a Fourier multiplier. Its mapping properties on Besov spaces govern the uniformity and tightness properties of the empirical process, and therefore the class of "test functions" for which a uniform CLT is valid.

5. Comparative Rates: Logarithmic Versus Root-n

For α-Brownian bridge inference, the effective "sample size" grows only logarithmically: $W_t$0 which is substantially slower than the linear growth in the i.i.d. case. This intrinsic limitation manifests in the Berry–Esseen rate for the MLE of $W_t$1. As $W_t$2, the rate $W_t$3 contrasts with the standard $W_t$4, indicating that parameter learning is intrinsically slow in this context, and substantial time extension towards $W_t$5 is necessary for precise estimation (Es-Sebaiy et al., 2020).

In nonparametric deconvolution, however, under a sufficiently smooth functional and favorable noise decay, the uniform CLT recovers $W_t$6-rate asymptotics, with the limiting process a generalized Brownian bridge. When the noise is absent, root-n rates and the classical Brownian bridge covariance structure are reinstated (Söhl et al., 2012).

6. Statistical Implications and Limitations

The slow variance decay $W_t$7 in the α-Brownian bridge model necessitates observing the process very near the terminal time $W_t$8 for accurate inference. Confidence intervals based on normal approximation have error of order $W_t$9, requiring skewness correction when $(B_t^0)_{t\in[0,1]}$0 is not extremely close to $(B_t^0)_{t\in[0,1]}$1 (Es-Sebaiy et al., 2020). In deconvolution settings, the generalized Brownian bridge limit governs the efficient estimation of broad classes of linear functionals, provided a delicate balance between the smoothness of the functionals and the ill-posedness of the deconvolution problem is satisfied (Söhl et al., 2012).

The Fourier-multiplier mappings of the deconvolution operator are instrumental in establishing these uniform CLTs, dictating both the moment and tightness properties required for functional weak convergence.

---

References:

• "Optimal Berry-Esséen bound for Maximum likelihood estimation of the drift parameter in $(B_t^0)_{t\in[0,1]}$2-Brownian bridge" (Es-Sebaiy et al., 2020)
• "A uniform central limit theorem and efficiency for deconvolution estimators" (Söhl et al., 2012)