UND INFORMATIK. Maximal φ-inequalities for Nonnegative Submartingales. G. Alsmeyer und U. Rösler

Similar documents
Notes 18 : Optional Sampling Theorem

On the maximal density of sum-free sets

An essay on the general theory of stochastic processes

Lecture 4 Lebesgue spaces and inequalities

Random Bernstein-Markov factors

THE L 2 -HODGE THEORY AND REPRESENTATION ON R n

The Uniform Integrability of Martingales. On a Question by Alexander Cherny

Sobolev Spaces. Chapter 10

SOME REMARKS ON THE SPACE OF DIFFERENCES OF SUBLINEAR FUNCTIONS

TWO MAPPINGS RELATED TO SEMI-INNER PRODUCTS AND THEIR APPLICATIONS IN GEOMETRY OF NORMED LINEAR SPACES. S.S. Dragomir and J.J.

Notes 15 : UI Martingales

PREDICTABLE REPRESENTATION PROPERTY OF SOME HILBERTIAN MARTINGALES. 1. Introduction.

New Versions of Some Classical Stochastic Inequalities

On the minimum of several random variables

UND INFORMATIK. Asexual Versus Promiscuous Bisexual Galton-Watson Processes: The Extinction Probability Ratio. G. Alsmeyer und U.

CLOSED RANGE POSITIVE OPERATORS ON BANACH SPACES

Solution for Problem 7.1. We argue by contradiction. If the limit were not infinite, then since τ M (ω) is nondecreasing we would have

Pointwise multipliers on martingale Campanato spaces

Eötvös Loránd University, Budapest. 13 May 2005

On a Class of Multidimensional Optimal Transportation Problems

1 Brownian Local Time

Improvements of the Giaccardi and the Petrović inequality and related Stolarsky type means

5 Measure theory II. (or. lim. Prove the proposition. 5. For fixed F A and φ M define the restriction of φ on F by writing.

UTILITY OPTIMIZATION IN A FINITE SCENARIO SETTING

On duality theory of conic linear problems

Herz (cf. [H], and also [BS]) proved that the reverse inequality is also true, that is,

ON THE FOLIATION OF SPACE-TIME BY CONSTANT MEAN CURVATURE HYPERSURFACES

NORMS ON SPACE OF MATRICES

2 Sequences, Continuity, and Limits

If g is also continuous and strictly increasing on J, we may apply the strictly increasing inverse function g 1 to this inequality to get

Regularity of the density for the stochastic heat equation

Estimates for probabilities of independent events and infinite series

Wavelets and modular inequalities in variable L p spaces

Problem set 4, Real Analysis I, Spring, 2015.

APPROXIMATE IDENTITIES AND YOUNG TYPE INEQUALITIES IN VARIABLE LEBESGUE ORLICZ SPACES L p( ) (log L) q( )

Self-normalized laws of the iterated logarithm

A note on the growth rate in the Fazekas Klesov general law of large numbers and on the weak law of large numbers for tail series

Nonlinear stabilization via a linear observability

OPTIMAL SCALING FOR P -NORMS AND COMPONENTWISE DISTANCE TO SINGULARITY

L p Spaces and Convexity

Exponential stability of operators and operator semigroups

PARTIAL REGULARITY OF BRENIER SOLUTIONS OF THE MONGE-AMPÈRE EQUATION

On the Set of Limit Points of Normed Sums of Geometrically Weighted I.I.D. Bounded Random Variables

A CRITERION FOR A DEGREE-ONE HOLOMORPHIC MAP TO BE A BIHOLOMORPHISM. 1. Introduction

TAUBERIAN THEOREM FOR HARMONIC MEAN OF STIELTJES TRANSFORMS AND ITS APPLICATIONS TO LINEAR DIFFUSIONS

Convex Functions and Optimization

Fundamental Inequalities, Convergence and the Optional Stopping Theorem for Continuous-Time Martingales

MAXIMALITY OF SUMS OF TWO MAXIMAL MONOTONE OPERATORS

Deviation Measures and Normals of Convex Bodies

1/12/05: sec 3.1 and my article: How good is the Lebesgue measure?, Math. Intelligencer 11(2) (1989),

Almost sure limit theorems for U-statistics

JUHA KINNUNEN. Harmonic Analysis

Convex Analysis and Economic Theory AY Elementary properties of convex functions

ASYMPTOTIC STRUCTURE FOR SOLUTIONS OF THE NAVIER STOKES EQUATIONS. Tian Ma. Shouhong Wang

Generalized Orlicz spaces and Wasserstein distances for convex concave scale functions

A NOTE ON STOCHASTIC INTEGRALS AS L 2 -CURVES

Universität des Saarlandes. Fachrichtung 6.1 Mathematik

arxiv: v1 [math.oc] 21 Mar 2015

LACK OF HÖLDER REGULARITY OF THE FLOW FOR 2D EULER EQUATIONS WITH UNBOUNDED VORTICITY. 1. Introduction

ON THE STRONG LIMIT THEOREMS FOR DOUBLE ARRAYS OF BLOCKWISE M-DEPENDENT RANDOM VARIABLES

LARGE DEVIATIONS OF TYPICAL LINEAR FUNCTIONALS ON A CONVEX BODY WITH UNCONDITIONAL BASIS. S. G. Bobkov and F. L. Nazarov. September 25, 2011

Asymptotic of Enumerative Invariants in CP 2

arxiv: v1 [math.ap] 18 May 2017

Bellman function approach to the sharp constants in uniform convexity

Boundedly complete weak-cauchy basic sequences in Banach spaces with the PCP

The Pedestrian s Guide to Local Time

arxiv: v1 [math.ca] 7 Aug 2015

The small ball property in Banach spaces (quantitative results)

A note on the convex infimum convolution inequality

A NEW PROOF OF THE ATOMIC DECOMPOSITION OF HARDY SPACES

ON THE UNIQUENESS PROPERTY FOR PRODUCTS OF SYMMETRIC INVARIANT PROBABILITY MEASURES

On Some Extensions of Bernstein s Inequality for Self-Adjoint Operators

ONLINE APPENDIX TO: NONPARAMETRIC IDENTIFICATION OF THE MIXED HAZARD MODEL USING MARTINGALE-BASED MOMENTS

BLOWUP THEORY FOR THE CRITICAL NONLINEAR SCHRÖDINGER EQUATIONS REVISITED

Coupling Binomial with normal

The primitive root theorem

EVALUATIONS OF EXPECTED GENERALIZED ORDER STATISTICS IN VARIOUS SCALE UNITS

ON ψ-direct SUMS OF BANACH SPACES AND CONVEXITY

Introduction to Real Analysis Alternative Chapter 1

(A n + B n + 1) A n + B n

EXISTENCE OF SET-INTERPOLATING AND ENERGY- MINIMIZING CURVES

The Dirichlet s P rinciple. In this lecture we discuss an alternative formulation of the Dirichlet problem for the Laplace equation:

A CLASS OF SCHUR MULTIPLIERS ON SOME QUASI-BANACH SPACES OF INFINITE MATRICES

Weighted Sums of Orthogonal Polynomials Related to Birth-Death Processes with Killing

SPACES ENDOWED WITH A GRAPH AND APPLICATIONS. Mina Dinarvand. 1. Introduction

Geometry and topology of continuous best and near best approximations

On groups of Hölder diffeomorphisms and their regularity

On the distributional divergence of vector fields vanishing at infinity

Self-equilibrated Functions in Dual Vector Spaces: a Boundedness Criterion

ON MATRIX VALUED SQUARE INTEGRABLE POSITIVE DEFINITE FUNCTIONS

Concentration behavior of the penalized least squares estimator

BIHARMONIC WAVE MAPS INTO SPHERES

THE SEMI ORLICZ SPACE cs d 1

A Concise Course on Stochastic Partial Differential Equations

4 Expectation & the Lebesgue Theorems

Yunhi Cho and Young-One Kim

Convergent Iterative Algorithms in the 2-inner Product Space R n

AN EFFECTIVE METRIC ON C(H, K) WITH NORMAL STRUCTURE. Mona Nabiei (Received 23 June, 2015)

A note on linear differential equations with periodic coefficients.

RESEARCH INSTITUTE FOR MATHEMATICAL SCIENCES RIMS A Note on an Anabelian Open Basis for a Smooth Variety. Yuichiro HOSHI.

Transcription:

ANGEWANDTE MATHEMATIK UND INFORMATIK Maximal φ-inequalities for Nonnegative Submartingales G. Alsmeyer und U. Rösler Universität Münster und Universität Kiel e-mail: gerolda@math.uni-muenster.de, roesler@math.uni-kiel.de Bericht Nr.??/2-S UNIVERSITÄT MÜNSTER

Maximal φ-inequalities for Nonnegative Submartingales Gerold Alsmeyer Institut für Mathematische Statistik Fachbereich Mathematik Westfälische Wilhelms-Universität Münster Einsteinstraße 62 D-4849 Münster, Germany Uwe Rösler Mathematisches Seminar Christian-Albrechts-Universität Kiel Ludewig-Meyn-Straße 4 D-2498 Kiel, Germany Let M n n be a nonnegative submartingale and def = max kn M k, n the associated maximal sequence. For nondecreasing convex functions φ :[, [, with φ = Orlicz functions we provide various inequalities for Eφ in terms of EΦ am n where, for a, Φ a x def = x s a a φ r r dr ds, x >. Of particular interest is the case φx = xfor which a variational argument leads us to EM n + E log xdx /2 2. A further discussion shows that the given bound is better than Doob s classical bound e e +EM n log + M n whenever EM n + e 2.78. AMS 99 subject classifications. 6G42, 6E5. Keywords and phrases. Nonnegative submartingale, maximal sequence, Orlicz and Young function, Choquet representation, convex function inequality.

2. Introduction Let M n n be a nonnegative submartingale and def = max kn M k, n the associated maxima. Moment inequalities for the sequence n in terms of M n n are usually based on Doob s maximal inequality which states that P >t M n dp. t { >t} for n and t>. If p>, a combination of. with Hölder s inequality shows EM np p p E p.2 p for n [4, p. 255f], the constant being sharp see [3, p. 4]. In case p = one finds with. that E e +EM n log + M n.3 e for n. Clearly, these results apply to M n n if M n n is a martingale. Orlicz and Young functions..2 and.3 are only special cases within a whole class of convex function inequalities based on Doob s inequality which is the main topic of this paper. Let C denote the class of Orlicz functions, that is unbounded, nondecreasing convex functions φ :[, [, with φ =. If the right derivative φ is also unbounded then φ is called a Young function and we denote by C the subclass of such functions. Given any probability space Ω, A,P, each φ C induces the semi-banach space L φ P, φ of φ-integrable random variables X on Ω, A, P, where X φ def = inf{ >: EΦ X / } defines the underlying semi-norm. L φ P, φ is called an Orlicz space and equals the space of α-times integrable functions L α P, α in case φx =x α for some α [,. Since φx = x φ sds xφ x by convexity, the numbers p = p φ def xφ x = inf x> φx and p = p φ def xφ x = sup x> φx.4 are both in [, ]. φ is called moderate [5, p. 62] if p < or, equivalently [7, Thm. 3..], if for some and then all > there exists a finite constant c such that φx c φx, x..5 This property is shared by all φ C which are also regularly varying at infinity at order α [2], thus including φx =x α for α [,. Examples of non-moderate Orlicz functions are φx = expx α for any α.

3 Given a Young function φ, the right continuous inverse ψ x def = inf{y : φ y >x}of φ is also unbounded and thus ψx def = x ψ s ds again an element of C, called the conjugate Young function to φ. Obviously, this conjugation is reflexive. A simple geometric argument shows [5, p. 63] that xφ x = φx + φ x ψ s ds = φx + ψφ x, x..6 and, by reflexivity, the same identity holds true with the roles of φ and ψ interchanged. With the help of.6, ψx p xψ x and ψ φ x x we infer as in [5, p. 69] that ψ xφ x = φx+ψφ x φx+ p φ xψ φ x φx+ ψ p xφ x ψ for x and thus p φ = inf x> xφ x φx p ψ Thm. 3..] so that we have the identity p ψ. The reverse inequality can also be shown [7, p φ = p ψ p ψ, or equivalently p φ + p ψ =..7 As further results stated in [7, Thm. 3..] we mention that for any φ C with p = p φ the assertions φx p φx for all > and x>;.8 φx x p.9 hold true, and that for moderate φ with p = p φ furthermore φx p φx for all > and x>;. φx. x p. The following two inequalities, the first of which may also be found in [5, p. 69], are easily deduced from another inequality stated as 2.2 in the next section. This latter inequality emerges as a special case of one of our results, see Corollary 2.2, but can also be derived by different arguments based on Doob s inequality and the Choquet representation of a function in C. For the interested reader this is briefly demonstrated in Appendix. Proposition.. Let φ be an Orlicz function with p = p φ > and M n n be a nonnegative submartingale. Then M n φ p p M n φ.2

for each n. Ifφis also moderate, i.e. p = p φ <, then furthermore 4 EφM n p p EφM n.3 p for each n. 2. Main Results Inequalities of type.3 are also the content of our main results to be presented in this section. They are based upon integration of a variational variant of Doob s inequality. to be proved in Proposition 3.2, namely P M n >t P M n />sds = t t t E + t 2. for all n, t>and,. Under additional contraints on φ we will see that by good choices of the constant in.3 can be improved considerably. We will also derive an inequality for E in terms of EM n log + M n which in many situations strictly beats.3. The following two subclasses of C will be of interest hereafter. We shall denote by C the set of all differentiable φ C whose derivative is concave or convex, and by C the set of φ C such that φ x x is integrable at and thus in particular φ =. Put C def = C C. Given φ C and a, define Φ a x def = x s a a φ r r dr ds, x > 2.2 and note that Φ a [a, C for a>. If φ C the same holds true for Φ def = Φ, whereas Φ otherwise. The function Φ will be of great importance in our subsequent analysis. If φ C then Φ is obviously again an element from this class. If in addition φ is concave or convex the same holds true for Φ x = x Φ x = φ x x φ r r dr, hence φ C implies Φ C. Use to see that φ and Φ are related through the differential equation xφ x Φx = φx, x 2.3 under the initial conditions φ = φ = Φ = Φ =. Ifφx=x p for some p>, then Φx = p xp, in particular Φ = φ in case φx =x 2. If φx=xthen Φ, but we have Φ x =xlog x x +. Further properties of Φ and its relation to φ are collected in Lemma 3. at the beginning of Section 3 where it will be seen particularly that Φ and φ grow at the same order of magnitude unless φ or its conjugate are non-moderate. Theorem 2.. Let M n n be a nonnegative submartingale and φ C. Then EφM n φa + EΦ am n / 2.4

for all a,, and n, in particular = 2 5 EφM n φa + EΦ a 2M n 2.5 for all a>and n. IfM n n is a positive martingale with M n+ cm n for some c> and all n, and if EΦ a M <, then Eφ c EΦ a M n EΦ a M 2.6 for all n. Of course, inequality 2.4 with a = is of interest only when Φ <, thus for φ C. The conditions on M n n implying 2.6 were given by Gundy [6] to demonstrate that the bound in.3 cannot be much improved, see 2.7 below and also [8, p. 7f]. If φx =x p for some p>, then Φx = p xp implies in 2.4 with a = EM np p p EMp n 2.7 for all n and,. Elementary calculus shows that p def p = argmin, = p p With this value of in 2.7 Doob s L p -inequality.2 comes out again. For an extension of it consider nonnegative increasing functions φ on [, such that φ /γ is also convex for some γ>. As before let M n n be a nonnegative submartingale. Then the same holds true for φ /γ M n n, the associated sequence of maxima being φ /γ n. Consequently,.2 implies γ γ Eφ EφM n 2.8 γ for each n. Interesting examples are φx =x p log r + x for p> and r choose γ = p, as well as φx =e rx for r>. For the latter example any γ>will do and since γ γ γ decreases to e as γ, we obtain Ee rm n eee rm n 2.9 for all n and r>. We will show in the next section that Φx p φx for each φ C with p = p φ > φ C. With this at hand the subsequent corollary follows from Theorem 2.. Corollary 2.2. Given the situation of Theorem 2., let φ C be such that p = p φ >. Then Eφ p EφM n/ 2.

for all, and n. Ifφis also moderate p < then for each n. 6 Eφ p p p p p EφM n 2. A comparison of the constants going with EφM n in 2. and.3 shows that the one in 2. is strictly better unless p = p see Lemma A.2 in Appendix 2 for a rigorous proof. For large p, p and β def = p /p we also have that p p p p p βe, while p p p e β. Putting q = p p and choosing = q in 2., we obtain EφM n EφqM n 2.2 for each n and any φ C with p>. It is this inequality which will easily give the assertions of Proposition. see Section 3. By a similar argument as the one leading to 2.8, Doob s inequality.2 can be used to get refinements of.3 and 2. for functions φ C. However, the main tool for this is not 2. but rather a suitable Choquet representation of φ exploited in a similar manner as in [] for the special case of φ C with concave derivative. Theorem 2.3. Given the situation of Theorem 2., let φ C, k and φ k the k-th order derivative of φ. Then φ k C φ C implies EφM n k+ k + EφM n 2.3 k as well as for each n. M n φ k + k M n φ 2.4 In case φx =x p for integral p>wehaveφ p C and thus that 2.3 coincides with.2. We finally turn to the case φx =xfor which Φ but Φ x =xlog x x + as mentioned earlier. A more general version of inequality 2.4 proved in Section 3 Proposition 3.2 will be used to derive the following alternative to Doob s L -inequality.3. Theorem 2.4. If M n n is a nonnegative submartingale, then E b + b b E log xdx 2.5

7 for all b>and n. The value of b which minimizes the right hand side equals b def = + E M n log xdx /2 and gives EM n + /2 2 E log xdx 2.6 If M n n is a positive martingale with M n+ cm n for some c>and all n, and if EM log + M <, then E EM n log + M n EM log + M 2.7 c for all n. Since x log ydy=xlog+ x x for x, inequality 2.5 may be restated as E b + b EM n log + M n EM n + 2.8 b for all n and b>. For the special choices b = EM n + + and b = e, this yields for each n E +EM n + EM n + EM n log + M n 2.9 and E e + e EM n log + M n EM n +, 2.2 e respectively, where the right hand side of 2.9 is to be interpreted as if M n a.s. Note that even the choice b = e, though only suboptimal, leads to a better bound for E than in.3 whenever EM n + e 2.78. The previous upper bounds for EM n may be further improved if m def = EM >. To see this note that ˆM def n n =, M m, M m,... forms again a nonnegative submartingale whose maxima ˆM n satisfy = M n m for n. Consequently, 2.5 and 2.8 for ˆM n n give Corollary 2.5. EM n E ˆM n+ b + b ˆM n If M n n is a nonnegative submartingale with m = EM >, then b+ /m b E b b E m log+ log xdx m + 2.2 E m /2 2 for all b>and n, in particular /m E + E log xdx 2.22

when choosing the minimizing b compare 2.6. 8 3. Proofs We begin with a collection of some useful properties of the function Φ = Φ in 2.2 associated with an element φ of C. Lemma 3.. For each φ C with p = p φ > the function Φ satisfies Φx φx, x. 3. p If φ is moderate, i.e. p = p φ <, then Φx p φx, x. 3.2 Finally, for each φ C the inequality holds true. lim inf x Φx x log x > 3.3 Φx The final inequality shows that lim x x = whenever φ C is a function close to the identity in the sense that φx = oxlog x asx. Another class of examples comprises φx =rx log r + x log r + x for r in which case Φx =+x log r + x use the differential equation 2.3. Proof. Using xφ x pφx, we obtain by partial integration s φ r r dr = φs s + s φr r 2 dr φs s + p s φ r r dr and thus s φ r r dr p φs p s p φ s for s. An integration of this inequality with respect to s obviously implies 3.. The second inequality 3.2 is obtained analogously when utilizing that p < implies xφ x p φx for all x. As to 3.3, choose a> such that φ a >. Then Φx Φ a x = x s a a φ r dr ds = φ a x logx/a x + a for all x a. The proofs of Theorem 2. and 2.4 are based on the following more general proposition.

9 Proposition 3.2. Let M n n be a nonnegative submartingale and φ C. Then inequality 2. holds true and furthermore EφM n φb+ Φ a Φ a b Φ {M n />b} ab b dp 3.4 for all n, a, b > and,. If φ x x valid for b =. is integrable at, i.e. φ C, then 3.4 remains Proof. Doob s maximal inequality gives P >t M n dp t { >t} = t = t t P M n >t,m n >sds P M n >tds + t P M n >t + t t t P M n >sds PM n />sds 3.5 and thus P >t P M n />sds t t for all n, t> and,, which is 2.. With this result we further infer EφM n φb+ φb+ = φb+ = φb+ = φb+ b φ tpm n >tdt b b b {M n />b} φ t PM n />sds dt t t s φ t dt P M n />sds b t Φ as Φ ab PM n />sds Φ a Φ a b Φ ab b dp for all n, b, t >,, and a>, where a = may also be chosen if φ x x at. is integrable Proof of Theorem 2.. Inequality 2.4 follows directly from Proposition 3.2 if we choose b = a and recall that Φ a a =Φ aa =. Hence we are left with the proof of 2.6. If M n n is a positive martingale satisfying M n+ cm n for all n and some c>, then P >t M n dp ct { >t} ct {M >t} M dp

for all n and t>, see [8, p. 72]. Consequently, P >t M n dp M dp ct {M n >t} ct {M >t} = P M n >s PM >s ds ct t + P M n >t PM >t c 3.6 for all n and t>. Assuming EΦ a M <, and also Eφ < there is nothing to prove otherwise, 2.6 now follows upon integration over, with respect to t of both sides of 3.6 multiplied with φ t. We must only note that φ t P M n >t PM >t dt = EφM n EφM because φ is convex and EφM k EφM n < for k n. We continue with a short proof of Proposition. stated in the Introduction. Proof of Proposition.. Let φ C with p = p φ >, put q def = p p and recall from 2.2 that Eφ def EφqM n for each n. Setting γ n = M n φ, this inequality implies as well as use also. EφM n/q n EφM n / n = EφM n EφqM n q p EφM n. The second inequality proves.3 while the first one yields φ q n and thus assertion.2. Proof of Theorem 2.3. The basic tool for the proof of Theorem 2.3 is to convert the assumption of smooth convexity φ k C into a suitable Choquet decomposition of φ. Each φ C can be written as φx = x t + Q φ dt [, where Q φ dt =φ δ + φ dt. Hence, if φ C, then φx = = = = x x φ y dy y t + Q φ dt dy [, [, [, x y t + dy Q φ dt x t + 2 Q φ dt. 2

An inductive argument now gives that φx = [, x t + k+ k +! Q φ kdt 3.7 for each φ C with φ k C. Put ϕ k,t x def = x t+ k+ k+! for k and t [,. Note that is convex for each t. Thus we infer with the help of 3.7 and the argument which ϕ /k+ k,t proved 2.8 that EφM n = = [, Eϕ k,t M n Q φ kdt k+ k + Eϕ k,t M n Q k φ kdt [, k+ k + EφM n k for each n. Replacing with /γ def n in the previous estimation, where γ n = k+ k M n φ, further gives 2.4 by a similar argument as in the proof of Proposition.. Proof of Theorem 2.4. If φx =x, then Φ x =xlog x x + and Φ x = log x for x>. Inequality 3.4 with these functions reads E b + {M n >b} log M n + b log b M n dp = b + M n log M n M n log + log b + +b dp {M n >b} for all b> and,. Choose b> and = b to obtain 2.5 in its equivalent form 2.8. 2.6 follows by elementary calculus. Finally, if M n n is a positive martingale with M n+ cm n for some c> and all n, and if EM log + M <, then a use of the first inequality in 3.6 gives after partial integration E M n dp M dp dt ct {M n >t} {M >t} M = c E M n t dt M = EM n log + M n EM log + M c t dt for all n and hence the asserted inequality 2.7. Appendix Inequality 2.2 from which Proposition. was deduced may be viewed as a specialization of inequality A.2 below to the pair X, Y =M n,. The purpose of this short

2 appendix is to provide a proof of A.2 based upon a Choquet representation of the involved convex function φ. Lemma A.. Let X, Y be nonnegative random variables satisfying Doob s inequality tpy t E {Y t} X for all t. Then EφY EφqX holds for each Orlicz function φ, where q def = q φ = p φ p φ. A. A.2 Proof. If p = p φ =,thusq=, there is nothing to prove. So let p> and put V = qx. Write φ in Choquet decomposition, that is φx = x t + φ dt, x. Then we obtain EφV EφY = E = E = E + E [, V t + Y t + φ dt [, = E {V Y } V t φ dt [,V ] V t φ dt [,V ] V t φ dt [,Y ] + E V Y φ dt [,Y ] = I + I 2. [,Y ] [,Y ] [,Y ] Y t φ dt V t φ dt Y t φ dt V t φ dt + {V<Y} Y,V ] V,Y ] t V φ dt It is easily seen that I. So it remains to prove that I 2 = EV φ V EY φ Y. To this end write EY φ Y peφy = pe Y t + φ dt [, = peyφ Y pe tφ dt {Y t} which, after rearranging terms and using A., leads to EY φ Y qe tφ dt qe {Y t} {Y t} Xφ dt = EV φ Y and thus the desired conclusion.

3 Appendix 2 We claimed in Section 2 that y y y x y y x A.3 x holds true for all y x> and that the inequality is strict unless x = y. A.3 may be rewritten as y x y Taking logarithms we arrive at y logx logy y x. y y log x log y. The desired conclusion is now an immediate consequence of the following lemma. Lemma A.2. For each y>, the function f y :,y] IR, f y x def = y logx logy y log x log y is strictly decreasing with f y y =. Proof. It suffices to note that for all x,y. f yx = y x y x < References [] ALSMEYER, G. and RÖSLER, U. 23. The best constant in the Topchii-Vatutin inequality for martingales. To appear in Statist. Probab. Letters. [2] BINGHAM, N.H., GOLDIE, C.M. and TEUGELS, J.L. 987. Regular Variation. Cambridge Univ. Press, Cambridge. [3] BURKHOLDER, D.L. 99. Explorations in martingale theory and applications. Ecole d Eté de Probabilité de Saint Flour XIX -989, Ed. P. Hennequin, Lect. Notes Math. 464, -66. Springer, Berlin. [4] CHOW, Y.S. and TEICHER, H. 997. Probability Theory: Independence, Interchangeability, Martingales 3rd Edition. Springer, New York. [5] DELLACHERIE, C. and MEYER, P.A. 982. Probabilities and Potential B. Theory of Martingales. North-Holland, Amsterdam. [6] GUNDY, R.F. 969. On the class L log L, martingales, and singular integrals. Studia Math. 33, 9-8. [7] LONG, R. 993. Martingale Spaces and Inequalities. Peking University Press and Vieweg, Wiesbaden. [8] NEVEU, J. 975. Discrete-Parameter Martingales. North-Holland, Amsterdam.