On a class of martingale inequalities

Let Z = {Z₀, Z₁, Z₂,…} be a martingale, with difference sequence X₀ = Z₀, X_i = Z_i ? Z_{i ? 1}, i ≥ 1. The principal purpose of this paper is to prove that the best constant in the inequality λP(sup_i |X_i| ≥ λ) ≤ C sup_iE |Z_i|, for λ > 0, is C = (log 2)^?1. If Z is finite of length n, it is proved that the best constant is $\text{C}{}_{\mathrm{n}}\text{= [n(2}{}^{\mathrm{<mtext>1</mtext><mtext>n</mtext>}}\text{? 1)]}{}^{\mathrm{?1}}$. The analogous best constant C_n(z) when Z₀ ≡ z is also determined. For these finite cases, examples of martingales attaining equality are constructed. The results follow from an explicit determination of the quantity G_n(z, E) = sup_zP(max_i=1,…,n |X_i| ≥ 1), the supremum being taken over all martingales Z with Z₀ ≡ z and E|Z_n| = E. The expression for G_n(z,E) is derived by induction, using methods from the theory of moments.     

A refinement of the Riesz decomposition for amarts and semiamarts

A real-valued adapted sequence of random variables is an amart if and only if it can be written as a sum of a martingale and a sequence dominated in absolute value by a Doob potential, i.e., a positive supermartingale that converges to 0 in L¹. The same holds for vector-valued uniform amarts with the norm replacing the absolute value.     

Common strict character of some sharp infinite-sequence martingale inequalities

A procedure is given for proving strictness of some sharp, infinite-sequence martingale inequalities, which arise from sharp, finite-sequence martingale inequalities attained by degenerating extremal distributions. The procedure is applied to obtain strictness of the sharp inequalities of Cox and Kemperman

$\text{P(|X}{}_{\mathrm{i}}\text{|?1 for some i=1, 2,…)?(ln 2)}{}^{\mathrm{?1}}\text{sup}\text{n}\text{E}\text{∑}\text{i=0}\text{n}\text{X}{}_{\mathrm{i}}$

and of Cox (sharp form of Burkholder's inequality)

$\text{P}\text{∑}\text{i=0}\text{∞}\text{X}{}^2{}_{\mathrm{i}}\text{?1}\text{? e}{}^{\mathrm{<mtext>1</mtext><mtext>2</mtext>}}\text{sup}\text{n}\text{E}\text{∑}\text{i=0}\text{n}\text{X}{}_{\mathrm{i}}$

for all nontrivial martingale difference sequences X₀,X₁,….     

Limit distributions for the maxima of stationary Gaussian processes

Let {X_n} be a stationary Gaussian sequence with E{X₀} = 0, {X²₀} = 1 and E{X₀X_n} = rⁿ_n Let c_n = (2ln n)^{$\text{built1}\text{2}$}, b_n = c_n? $\text{1}\text{2}$c^-1_n ln(4π ln n), and set M_n = max_{0 ?k?n}X_k. A classical result for independent normal random variables is that

$\text{P[c}{}_{\mathrm{n}}\text{(M}{}_{\mathrm{n}}\text{?b}{}_{\mathrm{n}}\text{)?x]→exp[-e}{}^{\mathrm{-x}}\text{] as n → ∞ for all x.}$

Berman has shown that (1) applies as well to dependent sequences provided r_nlnn = o(1). Suppose now that {r_n} is a convex correlation sequence satisfying r_n = o(1), (r_nlnn)^-1 is monotone for large n and o(1). Then

$\text{P[r}{}_{\mathrm{n}}{}^{\mathrm{<mtext>-1</mtext><mtext>2</mtext>}}\text{(M}{}_{\mathrm{n}}\text{? (1?r}{}_{\mathrm{n}}\text{)}{}^{\mathrm{<mtext>1</mtext><mtext>2</mtext>}}\text{b}{}_{\mathrm{n}}\text{)?x] → Ф(x)}$

for all x, where Ф is the normal distribution function. While the normal can thus be viewed as a second natural limit distribution for {M_n}, there are others. In particular, the limit distribution is given below when r_n is (sufficiently close to) γ/ln n. We further exhibit a collection of limit distributions which can arise when r_n decays to zero in a nonsmooth manner. Continuous parameter Gaussian processes are also considered. A modified version of (1) has been given by Pickands for some continuous processes which possess sufficient asymptotic independence properties. Under a weaker form of asymptotic independence, we obtain a version of (2).     

General theorems on rates of convergence in distribution of random variables I. General limit theorems

Let (X_n)_n?$\text{N}$ be a sequence of real, independent, not necessarily identically distributed random variables (r.v.) with distribution functions F_Xn, and S_n = Σ_i=1ⁿX_i. The authors present limit theorems together with convergence rates for the normalized sums ?(n)S_n, where ?: $\text{N}$ → $\text{R}$⁺, ?(n) → 0, n → ∞, towards appropriate limiting r.v. X, the convergence being taken in the weak (star) sense. Thus higher order estimates are given for the expression ∝_$\text{R}$f(x) d[F_?(n)Sn(x) ? F_X(x)] which depend upon the normalizing function ?, decomposability properties of X and smoothness properties of the function f under consideration. The general theorems of this unified approach subsume O- and o-higher order error estimates based upon assumptions on associated moments. These results are also extended to multi-dimensional random vectors.     

On the value of a stopped set function process

For certain types of stochastic processes {X_n | n ∈ $\text{N}$}, which are integrable and adapted to a nondecreasing sequence of σ-algebras $\text{F}$_n on a probability space (Ω, $\text{F}$, P), several authors have studied the following problems: IfSdenotes the class of all stopping times for the stochastic basis {$\text{F}$_n | n ∈ $\text{N}$}, when is$\text{sup}{}_{\mathrm{s}}\text{∫}{}_{\mathrm{\Omega }}\text{| X}{}_{\mathrm{\sigma }}\text{| dP}$finite, and when is there a stopping time $\text{σ}$ for which this supremum is attained? In the present paper we set the problem in a measure theoretic framework. This approach turns out to be fruitful since it reveals the root of the problem: It avoids the use of such notions as probability, null set, integral, and even σ-additivity. It thus allows a considerable generalization of known results, simplifies proofs, and opens the door to further research.     

Hyperamarts: Conditions for regularity of continuous parameter processes

The regularity of trajectories of continuous parameter process (X_t)_t∈R+ in terms of the convergence of sequence E(X_Tn) for monotone sequences (T_n) of stopping times is investigated. The following result for the discrete parameter case generalizes the convergence theorems for closed martingales: For an adapted sequence (X_n)_1≤n≤∞ of integrable random variables, lim X_n exists and is equal to X_∞ and (X_T) is uniformly integrable over the set of all extended stopping times T, if and only if lim E(X_Tn) = E(X_∞) for every increasing sequence (T_n) of extended simple stopping times converging to ∞. By applying these discrete parameter theorems, convergence theorems about continuous parameter processes are obtained. For example, it is shown that a progressive, optionally separable process (X_t)_t∈R+ with E{X_T} < ∞ for every bounded stopping time T is right continuous if lim E(X_Tn) = E(X_T) for every bounded stopping time T and every descending sequence (T_n) of bounded stopping times converging to T. Also, Riesz decomposition of a hyperamart is obtained.     

Amarts: A class of asymptotic martingales B. Continuous parameter

A continuous-parameter ascending amart is a stochastic process (X_t)_{t +} such that E[X_τn] converges for every ascending sequence (τ_n) of optional times taking finitely many values. A descending amart is a process (X_t)_{t +} such that E[X_τn] converges for every descending sequence (τ_n), and an amart is a process which is both an ascending amart and a descending amart. Amarts include martingales and quasimartingales. The theory of continuous-parameter amarts parallels the theory of continuous-parameter martingales. For example, an amart has a modification every trajectory of which has right and left limits (in the ascending case, if it satisfies a mild boundedness condition). If an amart is right continuous in probability, then it has a modification every trajectory of which is right continuous. The Riesz and Doob-Meyer decomposition theorems are proved by applying the corresponding discrete-parameter decompositions. The Doob-Meyer decomposition theorem applies to general processes and generalizes the known Doob decompositions for continuous-parameter quasimartingales, submartingales, and supermartingales. A hyperamart is a process (X_t) such that E[X_τn] converges for any monotone sequence (τ_n) of bounded optional times, possibly not having finitely many values. Stronger limit theorems are available for hyperamarts. For example: A hyperamart (which satisfies mild regularity and boundedness conditions) is indistinguishable from a process all of whose trajectories have right and left limits.     

Reduction of variance for Gaussian densities via restriction to convex sets

Let X be a random vector with values in $\text{R}$ⁿ and a Gaussian density f. Let Y be a random vector whose density can be factored as k · f, where k is a logarithmically concave function on $\text{R}$ⁿ. We prove that the covariance matrix of X dominates the covariance matrix of Y by a positive semidefinite matrix. When k is the indicator function of a compact convex set A of positive measure the difference is positive definite. If A and X are both symmetric Var(a · X) is bounded above by an expression which is always strictly less than Var(a · X) for every a ∈ $\text{R}$ⁿ. Finally some counterexamples are given to show that these results cannot be extended to the general case where f is any logarithmically concave density.     

Embedding EnG in Euclidean space

It is shown that if G is an arbitrary upper semicontinuous decomposition of Eⁿ for which π(N_G embeds in S^m for some m?3, then the decomposition space $\text{E}{}^{\mathrm{n}}\text{G}$ embeds as a closed subset of E^n+m+1. The proof consists of constructing a cell-like upper semicontinuous decomposition G? of E^n+m+1 which intersects Eⁿ to yield precisely G and using Edwards' Cell-Like Approximation Theorem to show that G? is shrinkable. As an immediate corollary, $\text{E}{}^{\mathrm{n}}\text{G}$ embeds in E^n+2k+2 whenever G is an arbitrary k-dimensional upper semicontinuous decomposition of Eⁿ. This is an improvement of (n?1)-dimensions over the corresponding dimension theoretic result and examples due to Daverman show that this result is sharp in case n is odd and off by no more than one dimension in case n is even.     

Hermite series estimates of a probability density and its derivatives

The following estimate of the pth derivative of a probability density function is examined: $\text{Σ}{}_{\mathrm{k\; =\; 0}}{}^{\mathrm{N}}\text{a}\text{?}{}_{\mathrm{k}}\text{h}{}_{\mathrm{k}}\text{(x)}$, where h_k is the kth Hermite function and $\text{a}\text{?}{}_{\mathrm{k}}\text{= (}\text{(?1)}{}^{\mathrm{p}}\text{n}\text{)}$Σ_{i = 1}ⁿh_k^(p)(X_i) is calculated from a sequence X₁,…, X_n of independent random variables having the common unknown density. If the density has r derivatives the integrated square error converges to zero in the mean and almost completely as rapidly as O(n^?α) and O(n^?α log n), respectively, where $\text{α =}\text{2(r ? p)}\text{(2r + 1)}$. Rates for the uniform convergence both in the mean square and almost complete are also given. For any finite interval they are O(n^?β) and $\text{O(n}{}^{\mathrm{<mtext>?\beta </mtext><mtext>2</mtext>}}\text{log}\text{n)}$, respectively, where $\text{β =}\text{(2(r ? p) ? 1)}\text{(2r + 1)}$.     

Convergence in the pth-mean and some Weak Laws of Large Numbers for weighted sums of random elements in separable normed linear spaces

Theorem. Let X_n, n ≥ 1, be a sequence of tight random elements taking values in a separable Banach space B such that |X_n|, n ≥ 1, is uniformly integrable. Let a_nk, n ≥ 1, k ≥ 1, be a double array of real numbers satisfying Σ_{k ≥ 1} |a_nk| ≤ Γ for every n ≥ 1 for some positive constant Γ. Then Σ_{k ≥ 1}a_nkX_k, n ≥ 1, converges to 0 in probability if and only if Σ_{k ≥ 1}a_nkf(X_k), n ≥ 1, converges to 0 in probability for every f in the dual space B¹.     

Pointwise properties of convergence in probability

If a sequence of random variables X_n converges to X in probability we know little about the pointwise behavior X_n(ω). In this note we show that if X_n converges to X quickly enough (for example, like n^?α for α > 0) then, for almost all ω, X_n(ω) converges to X(ω) outside a set of density zero.     

A remark on the tail probability of a distribution

Let {X_n}_n≥1 be a sequence of independent and identically distributed random variables. For each integer n ≥ 1 and positive constants r, t, and ?, let S_n = Σ_j=1ⁿX_j and $\text{E{N}{}_{\mathrm{\infty }}\text{(r, t, ?)} = Σ}{}_{\mathrm{n=1}}{}^{\mathrm{\infty }}\text{n}{}^{\mathrm{r?2}}\text{P{|S}{}_{\mathrm{n}}\text{| > ?n}{}^{\mathrm{<mtext>r</mtext><mtext>t</mtext>}}\text{}}$. In this paper, we prove that (1) lim_?→0+?^α(r?1)E{N_∞(r, t, ?)} = K(r, t) if E(X₁) = 0, Var(X₁) = 1, and E(| X₁ |^t) < ∞, where 2 ≤ t < 2r ≤ 2t, $\text{K(r, t) =}\text{{2}{}^{\mathrm{<mtext>\alpha (r?1)</mtext><mtext>2</mtext>}}\text{Γ(}\text{(1 + α(r ? 1))}\text{2}\text{)}}\text{{(r ? 1) Γ(}\text{1}\text{2}\text{)}}$, and $\text{α =}\text{2t}\text{(2r ? t)}$; (2) $\text{lim}{}_{\mathrm{?\to 0+}}\text{G(t, ?)}\text{H(t, ?)}\text{= 0}$ if 2 < t < 4, E(X₁) = 0, Var(X₁) > 0, and E(|X₁|^t) < ∞, where G(t, ?) = E{N_∞(t, t, ?)} = Σ_n=1^∞n^t?2P{| S_n | > ?n} → ∞ as ? → 0+ and $\text{H(t, ?) = E{N}{}_{\mathrm{\infty }}\text{(t, t, ?)} = Σ}{}_{\mathrm{n=1}}{}^{\mathrm{\infty }}\text{n}{}^{\mathrm{t?2}}\text{P{| S}{}_{\mathrm{n}}\text{| > ?n}{}^{\mathrm{<mtext>2</mtext><mtext>t</mtext>}}\text{} → ∞}\text{as}\text{? → 0+}$, i.e., H(t, ?) goes to infinity much faster than G(t, ?) as ? → 0+ if 2 < t < 4, E(X₁) = 0, Var(X₁) > 0, and E(| X₁ |^t) < ∞. Our results provide us with a much better and deeper understanding of the tail probability of a distribution.     

Hecke operators and an identity for the dedekind sums

If h, k ∈ Z, k > 0, the Dedekind sum is given by

$\text{s(h,k) =}\text{∑}\text{μ=1}\text{k}\text{μ}\text{k}\text{hμ}\text{k}$

, with

$\text{((x)) = x ? [x] ?}\text{1}\text{2}\text{, x?Z}$

,

$\text{=0 , x∈Z}$

. The Hecke operators T_n for the full modular group SL(2, Z) are applied to log η(τ) to derive the identities (n ∈ Z⁺)

$\text{∑ ∑ s(ah+bk,dk) = σ(n)s(h,k)}$

,

$\text{ad=n b(}\text{mod}\text{d)}$

$\text{d>0}$

where (h, k) = 1, k > 0 and σ(n) is the sum of the positive divisors of n. Petersson had earlier proved (1) under the additional assumption k ≡ 0, h ≡ 1 (mod n). Dedekind himself proved (1) when n is prime.     

A note on the J1-convergence of a first-passage process

Let X₁, X₂, … be a sequence of independent and identically distributed random variables with mean zero such that the common distribution function belongs to the domain of attraction of a stable law G_α,β with 1<α<2 and β=1 or α=2. If S_n=X₁+…X_n and N(ξ)=min{k:S_k>ξ}, ξ>0, then it is shown that $\text{N(nt)}\text{B}{}^1\text{(n)}$, 0<t<1, converges weakly under the Skorohod J₁-topology to a stable subordinator of index $\text{1}\text{α}$, where B¹(n) depends on the norming constant for S_n.     

Local asymptotic normality for progressively censored likelihood ratio statistics and applications

Let X_n,1 ≤ X_n,2 ≤ … ≤ X_n,n be the ordered variables corresponding to a random sample of size n with respect to a family of probability measures {P_θ:θ ∈ Θ} where Θ is an open subset of the real line. In many practical situations the X_n,i are the observables and experimentation must be curtailed prior to X_n,n. If τ_n is a stopping variable adapted to the σ-fields {σ(X_n,1,…,X_n,k): 1 ≤ k ≤ n} and P_n,θ the projection of P_θ onto σ(X_n,1,…,X_n,τn), the local asymptotic normality of the stopped progressively censored likelihood ratio statistics is established with θ, $\text{θ}{}_{\mathrm{n}}\text{= θ + un}{}^{\mathrm{<mtext>?1</mtext><mtext>2</mtext>}}\text{∈ Θ}$ and θ, u held fixed, under certain conditions on the underlying distribution and on τ_n. Conditions are also given to ensure the local asymptotic normality of likelihood ratio statistics where the underlying observations are given in a series scheme.     

Sampling distribution for a class of estimators for nonregular linear processes

Let {X_t; t = 1, 2,…} be a linear process with a location parameter θ defined by X_t ? θ = Σ₀^∞g_rZ_t?r where {Z_t; t = 0, ±1,…} is a sequence of independent and identically distributed random variables, with E∥Z₁∥^δ < ∞ for some δ > 0. If δ ? 1 we assume further than E(Z₁) = 0. Let η = δ if 0 < δ < 2, and η = 2 if δ ? 2. Then assume that Σ₀^∞∥ g_r ∥^η < ∞. Consider the class of estimators $\text{θ}\text{∧}{}_{\mathrm{n}}$ given by $\text{θ}\text{∧}{}_{\mathrm{n}}\text{= Σ}{}_1{}^{\mathrm{n}}\text{c}{}_{\mathrm{nt}}\text{X}{}_{\mathrm{t}}\text{where}\text{c}{}_{\mathrm{nt}}$ is of the form c_nt = Σ_{p = 0}^sβ_npt^p for some s ? 0. An attempt has been made to investigate the distributional properties of $\text{θ}\text{∧}{}_{\mathrm{n}}$ in large samples for various choices of β_np (0 ? p ? s), s, and the distribution of Z₁ under the constraints Σ₀^∞r^kg_r = 0, 0 ? k ? q where q in an arbitrary integer, 0 ? q ? s.     

On the law of large numbers for stationary sequences

We show that if X_i is a stationary sequence for which $\text{S}{}_{\mathrm{n}}\text{B}{}_{\mathrm{n}}$ converges to a finite non zero random variable of constant sign, where S_n=X₁+X₂+?+X_n and B_n is a sequence of constants, then B_n is regularly varying with index 1. If in addition $\text{Σ}\text{P}\text{(|X}{}_1\text{|>B}{}_{\mathrm{n}}$ is finite, then $\text{E}\text{|X}{}_1\text{|}$ is finite, and if in addition to this X_i satisfies an asymptotic independence condition, $\text{E}\text{X}{}_1\text{≠ 0}$.