A functional Hungarian construction for sums of independent random variables
Ion Grama, Michael Nussbaum · 2024
Read A functional Hungarian construction for sums of independent random variables by Ion Grama, Michael Nussbaum (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-15043.txt.
Introduction
Ion Grama and Michael Nussbaum build a functional Hungarian construction for sums of independent random variables. The variables do not share one distribution. Their 2024 paper extends a classic strong approximation device. The partial sum process is indexed by functions from a class. The supremum over that class is taken outside the probability. That setup is just what a Komlós-Major-Tusnády inequality in bounded-function space requires.
Their build keeps the classic rate over wide classes of functions. Hölder balls are included, where the standard result strains. Keeping the rate at this breadth is the technical prize. Such strong fits power proofs that statistical experiments match in the limit. The result is core support for limit theory.
Catalog record
| CREATOR | Ion Grama, Michael Nussbaum |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 87 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-15043
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
A functional Hungarian construction for sums of independent random variables Thanks: Research of the first author supported by Deutsche Forschungsgemeinschaft, by the German Academic Exchange Service (DAAD) and partly by the Alexander von Humboldt Schtiftung. Thanks: Research of the second author supported by NSF Grant DMS-0072162 Ion Grama, Michael Nussbaum Abstract . We develop a Hungarian construction for the partial sum process of independent non-identically distributed random variables. The process is indexed by functions f f from a class ℋ \mathcal{H} , but the supremum over f ∈ f\in ℋ \mathcal{H} is taken outside the probability. This form is a prerequisite for the Komlós-Major-Tusnády inequality in the space of bounded functionals l ∞ ( ℋ ) l^{\infty}(\mathcal{H}) , but contrary to the latter it essentially preserves the classical n − 1 / 2 log n n^{-1/2}\log n approximation rate over large functional classes ℋ \mathcal{H} such as the Hölder ball of smoothness 1 / 2 1/2 . This specific form of a strong approximation is useful for proving asymptotic equivalence of statistical experiments. 1991 Mathematics Subject Classification Primary 60F17, 60F99; Secondary 62G07 1. Introduction Let X_{i}, i=1,...,n, be a sequence of independent random variables with zero means and finite variances. Let \mathcal{H} be a class of real valued functions on the unit interval [0,1] and t_{i}=i/n, i=1,...,n. The partial sum process indexed by functions is the process X^{n}(f)=n^{-1/2}\sum_{i=1}^{n}f(t_{i})X_{i},\;f\in\mathcal{H}. Suppose f\in\mathcal{H} are uniformly bounded; then X^{n}=\{X^{n}(f), f\in\mathcal{H}\} may be regarded as a random element with values in l^{\infty}(\mathcal{H}) - the space of real valued functionals on \mathcal{H}. The class \mathcal{H} is Donsker if X^{n} converges weakly in l^{\infty}(\mathcal{H}) to a Gaussian process. We are interested in associated coupling results, i.e. in finding versions of X^{n} and of this Gaussian process on a common probability space which are close as random variables. The standard coupling results of the type ”nearby variables with nearby laws” (cf. Dudley [ 3 ] , Section 11.6) naturally refer to the \sup -metric in l^{\infty}(\mathcal{H}): for an appropriate version of X^{n} ( \widetilde{X}^{n}=\{\widetilde{X}^{n}(f), f\in\mathcal{H}\}, say) and of a Gaussian process \widetilde{N}^{n}=\{\widetilde{N}^{n}(f), f\in\mathcal{H}\}, we have (1.1) P^{\ast}\left(\sup_{f\in\mathcal{H}}\left|\widetilde{X}^{n}\left(f\right)-\widetilde{N}^{n}\left(f\right)\right|>x\right)\rightarrow 0,\quad x>0, where P^{\ast} is the outer probability on the common probability space (cf. van der Vaart and Wellner [ 20 ] , 1.9.3, 1.10.4). Here we shall consider a different type of coupling. We are looking for versions \widetilde{X}^{n}, \widetilde{N}^{n} such that (1.2) \sup_{f\in\mathcal{H}}P\left(\left|\widetilde{X}^{n}(f)-\widetilde{N}^{n}(f)\right|>x\right)\rightarrow 0,\quad x>0, and such that additional exponential bounds of the Komlós-Major-Tusnády type are valid. Note that ( 1.2 ) is weaker than ( 1.1 ) since the supremum is taken outside the probability. More specifically we are interested in a construction involving also a rate sequence r_{n}\rightarrow 0 such that (1.3) \sup_{f\in\mathcal{H}}P\left(r_{n}^{-1}\left|\widetilde{X}^{n}(f)-\widetilde{N}^{n}(f)\right|>x\right)\leq c_{0}\exp\{-c_{1}x\},\quad x>0. Here c_{0}, c_{1} are constants depending on the class \mathcal{H}. The classical results of Komlós, Major and Tusnády [ 9 ] and [ 10 ] refer to a \sup inside the probability for \mathcal{H}=\mathcal{H}_{0}, where \mathcal{H}_{0} is the class of indicators f(t)=\mathbf{1}(t\leq s), s\in[0,1]. The following bound was established: for r_{n}=n^{-1/2} (1.4) P\left(r_{n}^{-1}\sup_{f\in\mathcal{H}_{0}}\left|\widetilde{X}^{n}\left(f\right)-\widetilde{N}^{n}\left(f\right)\right|>x\right)\leq c_{0}\exp\{-c_{1}x\},\quad x\geq c_{2}\log n, provided X_{1},...,X_{n} is a sequence of i.i.d. r.v.’s fulfilling Cramér’s condition (1.5) E\exp\{tX_{i}\}<\infty,\quad\left|t\right|\leq t_{0},\quad i=1,...,n, where c_{0}, c_{1}, c_{2} are constants depending on the common distribution of the X_{i} . Note that r_{n} in ( 1.4 ) can be interpreted as a rate of convergence in the CLT over l^{\infty}(\mathcal{H}_{0}) . The main reason for a construction with the supremum outside the probability is that an extension of ( 1.4 ) to larger functional classes \mathcal{H} in general implies a substantial loss of approximation rate r_{n} (cp. Koltchinskii ( [ 8 ] , theorem 11.1). Our goal is a construction where the almost n^{-1/2} -rate of the original KMT result is preserved despite the passage to large functional classes \mathcal{H} like Lipschitz classes. Couplings of the type ( 1.3 ) have first been obtained by Koltchinskii ( [ 8 ] , theorem 3.5) and Rio [ 18 ] for the empirical process of i.i.d. random variables, as intermediate results. They can be extended to a full functional KMT result, i.e. to a coupling in l^{\infty}(\mathcal{H}) with exponential bounds, but an additional control of the size of the functional class \mathcal{H} is required, usually in terms of entropy conditions. A reduced approximation rate r_{n} may occur as a result. We carry over the functional strong approximation result from the empirical process to the partial sum process under very general conditions: the distributions of X_{i} are allowed to be nonidentical and nonsmooth . That setting substantially complicates the task of a Hungarian construction. We can rely on the powerful methodology of Sakhanenko [ 19 ] , who established the classical coupling ( 1.4 ) for nonidentical and nonsmooth summands. We stress however that for the functional version ( 1.3 ) we need to perform the construction entirely anew. Our results relate to Sakhanenko’s [ 19 ] as Koltchinskii’s theorem 3.5 relates to Komlós, Major and Tusnády [ 9 ] and [ 10 ] . Further motivational discussion can be grouped under headings A)-C) below. A) Statistical applications. The Komlós-Major-Tusnády approximation has recently found an application in the asymptotic theory of statistical experiments. In [ 14 ] the classical KMT inequality for the empirical process was used to establish that a nonparametric experiment of i.i.d. observation on an interval can be approximated, in the sense of Le Cam’s deficiency distance, by a sequence of signal estimation problems in Gaussian white noise. The two sequences of experiments are then asymptotically equivalent for all purposes of statistical decision with bounded loss. This appears as a generalization of Le Cam’s theory of local asymptotic normality, applicable to ill-posed problems like density estimation. In particular it implies a nonparametric version of the Hàjek-Le Cam asymptotic minimax theorem. The control of the Le Cam distance is given by a relation to likelihood processes (see Le Cam and Yang [ 12 ] ). Assume that there is an element f_{0}\in\Sigma such that the measures in the experiments \mathcal{E}^{n} and \mathcal{G}^{n} are absolutely continuous w.r.t. P_{f_{0}}^{n} and Q_{f_{0}}^{n} respectively. If there are versions d\widetilde{P}_{f}^{n}/d\widetilde{P}_{f_{0}}^{n} and d\widetilde{Q}_{f}^{n}/d\widetilde{Q}_{f_{0}}^{n} of the likelihood ratios dP_{f}^{n}/dP_{f_{0}}^{n} and dQ_{f}^{n}/dQ_{f_{0}}^{n} on a common probability space (\Omega^{n},\mathcal{F}^{n},P^{n}), then \Delta\left(\mathcal{E}^{n},\mathcal{G}^{n}\right)\leq\sqrt{2}\sup_{f\in\Sigma}E_{P}^{n}\left(\sqrt{d\widetilde{P}_{f}^{n}/d\widetilde{P}_{f_{0}}^{n}}-\sqrt{d\widetilde{Q}_{f}^{n}/d\widetilde{Q}_{f_{0}}^{n}}\right)^{2} (here the expected value on the right side coincides with the Hellinger distance between \widetilde{P}_{f}^{n} and \widetilde{Q}_{f}^{n} ). Thus asymptotic equivalence of experiments \mathcal{E}^{n} and \mathcal{G}^{n} requires a ”good” coupling of the corresponding likelihood ratios dP_{f}^{n}/dP_{f_{0}}^{n} and dQ_{f}^{n}/dQ_{f_{0}}^{n} on a common probability space. This is achieved by constructing the linear terms (in f-f_{0} ) in the expansions of the log-likelihoods such that they are close as random variables; hence the demand for an inequality ( 1.3 ) with the supremum outside the probability. The Hungarian construction had been applied in statistics before, mostly for results on strong approximation of particular density and regression estimators (cf. Csörgő and Révész [ 2 ] ). It is typical for these results that the ” supremum inside the probability” is needed; for such an application of the functional KMT cf. Rio [ 18 ] . However for asymptotic equivalence of experiments, it turned out that it is sufficient, and indeed preferable, to have a coupling like ( 1.3 ) with the ”supremum outside the probability”. Applying theorem