
« La Belgique, victime de cruelles violences, a rencontré, dans ses épreuves, de grands et de nombreux témoignages de sympathie. L’Université de Louvain en a recueilli sa part. Ses professeurs, dispersés après le désastre de la ville, ont été invités dans des Universités étrangères. Beaucoup y ont trouvé un asile, quelques-uns même la possibilité de poursuivre leur enseignement. Ainsi, j’ai été appelé a faire des conférences à l’Université Harvard en Amérique l’année dernière, puis cette année au Collège de France. Le présent volume contient la matière des Leçons que j’ai faites au Collège de France entre décembre 1915 et mars 1916. Il est, après bien d’autres, un modeste souvenir de ces événements. ... Une partie des résultats établis dans cet article se trouvaient dans la troisième édition du Tome II de mon Cours d’Analyse, qui était sous presse lors de l’incendie de Louvain. Les Allemands ont brulé cet ouvrage : toutes les installations de mon éditeur ont, en effet, partagé le sort de la bibliothèque de l’Université. ... » Charles de la Vallée Poussin in Intégrales de Lebesgue, fonctions d'ensemble, classes de Baire, Leçons professées au Collège de France (1916). Excerpt from the introduction.
This post is devoted to some probabilistic aspects of uniform integrability, a basic concept that I like very much. Let Φ be the class of non-decreasing functions φ:R+→R+ such that
limx→+∞φ(x)x=+∞.
This class contains for instance the convex functions x↦xp with p>1 and x↦xlog(x). Let us fix a probability space (Ω,A,P), and denote by Lφ the set of random variables X such that φ(|X|)∈L1=L1((Ω,F,P),R). We have Lφ⊊L1. Clearly, if (Xi)i∈I⊂L1 is bounded in Lφ with φ∈Φ then (Xi)i∈I is bounded in L1.
Uniform integrability. For any family (Xi)i∈I⊊L1, the following three properties are equivalent. When one (and thus all) of these properties holds true, we say that the family (Xi)i∈I is uniformly integrable (UI). The first property can be seen as a natural definition of uniform integrability.
- (definition of UI) limx→+∞supi∈IE(|Xi|1|Xi|≥x)=0;
- (epsilon-delta) the family is bounded in L1: supi∈IE(|Xi|)<∞, and moreover ∀ε>0, ∃δ>0, ∀A∈F, P(A)≤δ⇒supi∈IE(|Xi|1A)≤ε;
- (de la Vallée Poussin) there exists φ∈Φ such that supi∈IE(φ(|Xi|))<∞.
The second property is often referred to as the epsilon-delta'' criterion. The third and last property is a boundedness in Lφ⊊L1 and is due to Charles-Jean Étienne Gustave Nicolas de la Vallée Poussin (1866 - 1962), a famous Belgian mathematician, also well known for his proof of the prime number theorem.
Proof of 1⇒2. For the boundedness in L1, we write, for any i∈I, and some x≥0 large enough,
E(|Xi|)≤E(|Xi|1|Xi|<x)+E(|Xi|1|Xi|≥x)≤x+supi∈IE(|Xi|1|Xi|≥x)<∞.
Next, by assumption, for any ε>0, there exists xε such that supi∈IE(|Xi|1|Xi|≥x)≤ε for any x≥xε. If P(A)≤δε:=ε/xε then for any i∈I,
E(|Xi|1A)=E(|Xi|1|Xi|<xε1A)+E(|Xi|1|Xi|≥xε1A)≤xεP(A)+ε≤2ε.
Proof of 2⇒1. Since (Xi)i∈I is bounded in L1, for every j∈I, if A:=Aj:={|Xj|≥x}, then, by Markov inequality, P(A)≤x−1supi∈IE(|Xi|)≤δ for x large enough, uniformly in j∈I, and the assumption gives limx→+∞supi,j∈IE(|Xi|1|Xj|≥x)=0.
Proof of 3⇒1. For any ε>0, since φ∈Φ, by definition of Φ there exists xε≥0 such that x≤εφ(x) for every x≥xε, and therefore
supi∈IE(|Xi|1|Xi|≥xε)≤εsupi∈IE(φ(|Xi|)1|Xi|≥xε)≤εsupi∈IE(φ(|Xi|)).
Proof of 1⇒3. Let us seek for a piecewise linear function φ of the form
φ(x)=∫x0φ′(t)dtwithφ′=∞∑n=1un1[n,n+1[andun=∑m≥11xm≤n
for a sequence xm+∞ to be constructed. We have φ∈Φ since un→+∞. Moreover, since φ≤∑∞n=1(u1+⋯+un)1[n,n+1[, we get, for any i∈I,
E(φ(|Xi|))≤∑∞n=1(u1+⋯+un)P(n≤|Xi|<n+1)=∑∞n=1unP(|Xi|≥n)=∑m≥1∑n≥xmP(|Xi|≥n)≤∑m≥1∑n≥xmnP(n≤|Xi|<n+1)≤∑m≥1E(|Xi|1|Xi|≥xm)∗≤∑m≥12−m<∞
where ∗≤ holds if for every m we select xm such that supi∈IE(|Xi|1|Xi|≥xm)≤2−m, which is allowed by assumption (we may replace 2−m by anything sumable in m).
This achieves the proof of the equivalence of the three properties: 1⇔2⇔3.
Alternative proof of 1⇒3. Following a suggestion made by Nicolas Fournier on an earlier version of this post, one can simply construct φ as follows:
φ(x)=∑m≥1(x−xm)+.
This function is non-decreasing and convex, and
φ(x)x=∑m≥1(1−xmx)+x→∞∑m≥11=∞.
It remains to note that
E(φ(|Xi|))≤∑m≥1E(|Xi|1|Xi|≥xm)≤1.
Convexity and moderate growth. In the de la Vallée Poussin criterion, one can always assume that the function φ is convex (in particular continuous), with moderate growth in the sense that for every x≥0,
φ(x)≤x2.
Indeed, the construction of φ that we gave above provides a function φ with piecewise constant and non-decreasing derivative, thus the function is convex (it is a Young function). Following Paul-André Meyer, to derive the moderate growth property, we may first observe that thanks to the way we constructed φ, for every x≥0,
φ′(2x)≤cφ′(x)
where one can take c=2. This follows from the fact that we have the freedom to take the xm's as large as we want, for instance in such a way that for every n≥1,
u2n≤un+∑m≥11n<xm≤2n≤2un.
Consequently, the function φ itself has also moderate growth since, denoting C:=2c,
φ(2x)=∫2x0g(u)du=∫x0g(2t)2dt≤2cφ(x)=Cφ(x).
Now φ(x)≤C1+kφ(2−kx) for any k≥1, and taking k=kx=⌈log2(x)⌉ we obtain
φ(x)≤C2Clog2(x)φ(1)=C2φ(1)2log2(C)log2(x)=C2φ(1)xlog2(C).
Since one can take c=2 we get C=4, which allows φ(x)≤x2 by scaling.
Examples of UI families.
- Any finite subset of L1 is UI;
- More generally, if supi∈I|Xi|∈L1 (domination: |Xi|≤X for every i∈I, with X∈L1) then (Xi)i∈I is UI. To see it, we may first observe that the singleton family {supi∈I|Xi|} is UI, and thus, by the de la Vallée Poussin criterion, there exists φ∈Φ such that φ(supi∈I|Xi|)∈L1, and therefore
supi∈IE(φ(|Xi|))≤E(supi∈Iφ(|Xi|))≤E(φ(supi∈I|Xi|))<∞,
which implies, by the de la Vallée Poussin criterion again, that (Xi)i∈I is UI;
- If U1⊂L1,…,Un⊂L1 is a finite collection of UI families then the union U1∪⋯∪Un is UI.
- If U is UI then its convex envelope (or convex hull) is UI. Beware however that the vector span of U is not UI in general.
- If (Xn)n≥1,X∈L1 and XnL1→X then (Xn)n≥1∪{X} is UI and (Xn−X)n≥1 is UI. To see it, for any ε>0, we first select n large enough such that supk≥nE(|Xk−X|)≤ε, and then δ>0 with the epsilon-delta criterion for the finite family {X1,…,Xn}, which gives, for any A∈F such that P(A)≤δ,
supn≥1E(|Xn−X|1A)≤max(max1≤k≤nE(|Xk−X|1A);supk≥nE(|Xk−X|))≤ε.
- The de la Vallée Poussin criterion is often used with φ(x)=x2, and means in this case that every bounded subset of L2 is UI.
Integrability. The de la Vallée Poussin criterion, when used with a singleton UI family {X}, states that X∈L1 implies that φ(|X|)∈L1 for some φ∈Φ. In other words, for every random variable, integrability can always be improved in a sense. There is not any paradox here since φ depends actually on X. Topologically, integrability is in a sense an open statement rather than a closed statement. An elementary instance of this phenomenon is visible for Riemann series, in the sense that if ∑n≥1n−s<∞ for some s>0 then ∑n≥1n−s′<∞ for some s′<s, because the convergence condition of the series is s>1'', which is an open condition.
Integrability of the limit. If (Xn)n≥1 is UI and Xn→X in probability then X∈L1. Indeed, by the Borel-Cantelli lemma, we can extract a subsequence (Xnk)k≥1 such that Xnk→X almost surely, and then, by the Fatou Lemma,
E(|X|)=E(lim_k→∞|Xnk|)≤lim_k→∞E(|Xnk|)≤supn≥1E(|Xn|)<∞
because UI implies boundedness in L1.
Dominated convergence. For any sequence of random variables (Xn)n≥1, we have
XnL1→Xif and only ifXnP→X and (Xn)n≥1 is UI
This can be seen as an improved dominated convergence theorem (since (Xn)n≥1 is UI when supn|Xn|∈L1). The proof may go as follows. We already know that if Xn→X in L1 then X∈L1 and Xn→X in probability (Markov inequality) and (Xn)n≥1∪{X} is UI (see above). Conversely, if Xn→X in probability and (Xn)n≥1 is UI, we already known that X∈L1 (Fatou lemma for an a.s. converging subsequence). Moreover (Xn−X)n≥1 is UI since X∈L1, and thus, using the convergence in probability and the epsilon-delta criterion, we obtain that for any ε>0 and large enough n,
E(|Xn−X|)=E(|Xn−X|1|Xn−X|≥ε)+E(|Xn−X|1|Xn−X|<ε)≤2ε.
Martingales. The American mathematician Joseph Leo Doob (1910 - 2004) has shown that if a sub-martingale (Mn)n≥1 is bounded in L1 then there exists M∞∈L1 such that Mn→M∞ almost surely. Moreover, in the case of martingales, this convergence holds also in L1 if and only if (Mn)n≥1 is UI. In the same spirit, and a bit more precisely, if (Mn)n≥1 is a martingale for a filtration (Fn)n≥1, then the following two properties are equivalent:
- (Mn)n≥1 is UI;
- (Mn)n≥1 is closed, meaning that there exists M∞∈L1 such that Mn=E(M∞|Fn) for all n≥1, and moreover Mn→M∞ almost surely and in L1.
Are there non UI martingales? Yes, but they are necessarily unbounded: supn≥0|Mn|∉L1, otherwise we may apply dominated convergence. A nice counter example is given by critical Galton-Watson branching process, defined recursively by M0=1 and
Mn+1=Xn+1,1+⋯+Xn+1,Mn,
where Xn+1,j is the number of offspring of individual j in generation n, and (Xj,k)j,k≥1 are i.i.d. random variables on N with law μ of mean 1 and such that μ(0)>0. The sequence (Mn)n≥1 is a non-negative martingale, and thus it converges almost surely to some M∞∈L1. It is also a Markov chain with state space N. The state 0 is absorbing and all the remaining states can lead to 0 and are thus transient. It follows then that almost surely, either (Mn)n≥0 converges to 0 or to +∞, and since M∞∈L1, it follows that M∞=0. However, the convergence cannot hold in L1 since this leads to the contradiction 1=E(Mn)→E(M∞)=0 (note that (Mn)n≥1 is bounded in L1).
Topology. The Dunford-Pettis theorem, due to the American mathematicians Nelson James Dunford (1906 - 1986) and Billy James Pettis (1913 - 1979), states that for every family (Xi)i∈I∈L1, the following propositions are equivalent.
- (Xi)i∈I is UI;
- (Xi)i∈I is relatively compact for the weak σ(L1,L∞) topology;
- (Xi)i∈I is relatively sequentially compact for the weak σ(L1,L∞) topology.
The proof, which is not given in this post, can be found for instance in Delacherie and Meyer or in Diestel. We just recall that a sequence (Xn)n≥1 in L1 converges to X∈L1 for the weak σ(L1,L∞) topology when ℓ(Xn)→ℓ(X) for every ℓ∈(L1)′=L∞, in other words when E(YXn)→E(YX) for every Y∈L∞.
The Dunford-Pettis theorem opens the door for the fine analysis of closed (possibly linear) subsets of L1, a deep subject in functional analysis and Banach spaces.
Tightness. If (Xi)i∈I⊂L1 is bounded in Lφ for φ:R+→R+ non-decreasing and such that limx→∞φ(x)=+∞, then the Markov inequality gives
supi∈IP(|Xi|≥R)≤supi∈IE(φ(|Xi|))φ(R)⟶0R→∞
and thus the family of distributions (PXi)i∈I is tight, in the sense that for every ε>0, there exists a compact subset Kε of R such that supi∈IP(|Xi|∈Kε)≥1−ε. The Prokhorov theorem states that tightness is equivalent to being relatively compact for the narrow topology (which is by the way metrizable by the bounded-Lipschitz Fortet-Mourier distance). The Prokhorov and the Dunford-Pettis theorems correspond to different topologies on different objects (distributions or random variables). The tightness of (PXi)i∈I is strictly weaker than the UI of (Xi)i∈I.
UI functions with respect to a family of laws. The UI property for a family (Xi)i∈I⊂L1 depends actually only on the marginal distributions (PXi)i∈I and does not feel the dependence between the Xi's. In this spirit, if (ηi)i∈I is a family of probability measures on a Borel space (E,E) and if f:E→R is a Borel function then we say that f is UI for (ηi)i∈I on E when
limt→∞supi∈I∫{|f|>t}|f|dηi=0.
This means that (f(Xi))i∈I is UI, where Xi∼ηi for every i∈I. This property is often used in applications as follows: if ηn→η narrowly for some probability measures (ηn)n≥1 and η and if f is continuous and UI for (ηn)n≥1 then
∫|f|dη<∞and∫fdηn→∫fdη.
Logarithmic potential. What follows is extracted from the survey Around the circular law (random matrices). We have already devoted a previous post to the logarithmic potential. Let P(C) be the set of probability measures on C which integrate log|⋅| in a neighborhood of infinity. The logarithmic potential Uμ of μ∈P(C) is the function Uμ:C→(−∞,+∞] defined for all z∈C by
Uμ(z)=−∫Clog|z−w|dμ(w)=−(log|⋅|∗μ)(z).
Let D′(C) be the set of Schwartz-Sobolev distributions on C. We have P(C)⊂D′(C). Since log|⋅| is Lebesgue locally integrable on C, the Fubini-Tonelli theorem implies that Uμ is a Lebesgue locally integrable function on C. In particular, we have Uμ<∞ almost everywhere and Uμ∈D′(C). By using Green's or Stockes' theorems, one may show, for instance via the Cauchy-Pompeiu formula, that for any smooth and compactly supported function φ:C→R,
−∫CΔφ(z)log|z|dxdy=2πφ(0)
where z=x+iy. Now can be written, in D′(C),
Δlog|⋅|=2πδ0
In other words, 12πlog|⋅| is the fundamental solution of the Laplace equation on R2. Note that log|⋅| is harmonic on C∖{0}. It follows that in D′(C),
ΔUμ=−2πμ,
i.e. for every smooth and compactly supported test function'' φ:C→R,
⟨ΔUμ,φ⟩D′=−∫CΔφ(z)Uμ(z)dxdy=−2π∫Cφ(z)dμ(z)=−⟨2πμ,φ⟩D′
where z=x+iy. Also −12πU⋅ is the Green operator on R2 (Laplacian inverse). For every μ,ν∈P(C), we have
Uμ=Uν almost everywhere ⇒μ=ν.
To see it, since Uμ=Uν in D′(C), we get ΔUμ=ΔUν in D′(C), and thus μ=ν in D′(C), and finally μ=ν as measures since μ and ν are Radon measures. (Note that this remains valid if Uμ=Uν+h for some harmonic h∈D′(C)). As for the Fourier transform, the pointwise convergence of logarithmic potentials along a sequence of probability measures implies the narrow convergence of the sequence to a probability measure. We need however some strong tightness. More precisely, if (μn)n≥1 is a sequence in P(C) and if U:C→(−∞,+∞] is such that
- (i) for a.a. z∈C, Uμn(z)→U(z);
- (ii) log(1+|⋅|) is UI for (μn)n≥1;
then there exists μ∈P(C) such that
- (j) Uμ=U almost everywhere;
- (jj) μ=−12πΔU in D′(C);
- (jjj) μn→μ narrowly.
Let us give a proof inspired from an article by Goldsheid and Khoruzhenko on random tridiagonal matrices. From the de la Vallée Poussin criterion, assumption (ii) implies that for every real number r≥1, there exists φ∈Φ, which may depend on r, which is moreover convex and has moderate growth φ(x)≤1+x2, and
supn∫φ(log(r+|w|))dμn(w)<∞.
Let K⊂C be an arbitrary compact set. Take r=r(K)≥1 large enough so that the ball of radius r−1 contains K. Therefore for every z∈K and w∈C,
φ(|log|z−w||)
The couple of inequalities above, together with the local Lebesgue integrability of (log|⋅|)2 on C, imply, by using Jensen and Fubini-Tonelli theorems,
supn∫Kφ(|Un(z)|)dxdy≤supn∬1K(z)φ(|log|z−w||)dμn(w)dxdy<∞,
where z=x+iy as usual. Since the de la Vallée Poussin criterion is necessary and sufficient for UI, this means that the sequence (Uμn)n≥1 is locally Lebesgue UI. Consequently, from (i) it follows that U is locally Lebesgue integrable and that Uμn→U in D′(C). Since the differential operator Δ is continuous in D′(C), we find that ΔUμn→ΔU in D′(C). Since ΔU≤0, it follows that μ:=−12πΔU is a measure (see e.g. Hormander). Since for a sequence of measures, convergence in D′(C) implies narrow convergence, we get μn=−12πΔUμn→μ=−12πΔU narrowly, which is (jj) and (jjj). Moreover, by assumptions (ii) we get additionally that μ∈P(C). It remains to show that Uμ=U almost everywhere Indeed, for any smooth and compactly supported φ:C→R, since the function log|⋅| is locally Lebesgue integrable, the Fubini-Tonelli theorem gives
∫φ(z)Uμn(z)dz
Now φ∗log|⋅|:w∈C↦∫φ(z)log|z−w|dz is continuous and is O(log|1+⋅|). Therefore, by (i-ii), Uμn→Uμ in D′(C), thus Uμ=U in D′(C) and then almost everywhere, giving (j).
Bonjour,
In the proof of 2==>1 , you argued that" .... if \Omega = B_1 .... \cup B_n is a cover, then ..... " But in general does such a cover exists ?
Sincerely,
Luc.
you are right, something was missing: either one has to ask for the epsilon-delta for any probability space realizing (X_i), or one has to add the boundedness in L^1 to the epsilon-delta criterion. This is fixed now. Thanks.
Bonjour,
C'est moi encore.
In the proof of 1 => 3, you select a strictly increasng sequecne {xm}, then construct un and claimed that φ∈Φ, since un increases to infinity. I think here we need to clarify further about the choice of xm. For example, the following construction fails:
for m, choose xm, an integer greater than 2m such that supi∈IE[|Xi|⋅1(|Xi|≥xm)<2−m,
then pick another integer xm+1≥(xm+1)∨2m+1, then
u2n≤n so that
limn→+∞u1+u2+…+u2n2n=0
and thus
φ(x)xx→+∞→0.
sorry, the counterexample is wrong!@Luc
To be more explicit, since limn→∞un=+∞, we know that for every A>0 there exists N=NA such that un≥A for all n≥N. Now, by definition of φ we have φ′(x)≥A for all x≥N, which gives φ(x)/x≥(1/x)∫xNφ′(t)dt≥(1−N/x)A for any x≥N. In particular φ(x)/x≥A/2 for any x≥2N, and therefore limx→∞φ(x)/x=+∞.
1.
sorry again , the arithmetic mean 1n+1∑nk=1ukn→+∞→+∞. Here is an elementary proof:
For any M>0, there exists N1∈N∗ such that un>M as n≥N1
and there exists N2∈N∗ such that N2>N1 and un>2M, as n≥N2,
then for n>N1+N2+1, we have
u1+u2+…+unn+1=u1+u2+…+uN1n+1+uN1+1+…+uN2n+1+uN2+1+…+unn+1≥uN1+1+…+uN2n+1+uN2+1+…+unn+1≥(N2−N1)×Mn+1+(n−N2)×(2M)n+1=2M−2+N2+N1n+1⋅M>M.
This implies φ(n+1)n+1n→+∞→+∞ hence φ(x)x≥φ(⌊x⌋)⌊x⌋⋅x−1xx→+∞→+∞.
2.
In the proof of (1⟹3) , the very last equality should be changed to the inequality ≤ ? see the following ?
……=∑m≥1∑n≥xmP(|Xi|≥n)=∞∑m=1[∑n≥xm∞∑k=nP(k≤|Xi|<k+1)]=∞∑m=1[∑k≥xmk∑n=xmP(k≤|Xi|<k+1)]=∞∑m=1∑k≥xm(k−xm)P(k≤|Xi|<k+1)≤∞∑m=1∑k≥xmk⋅P(k≤|Xi|<k+1)
1. All right. 2. Youare right, and this glitch is now fixed, thanks.
For French readers, the concept of uniform integrability, under the name équi-intégrabilité, is considered in the book « Bases mathématiques du calcul des probabilités » by Jacques Neveu. In particular the de la Vallée Poussin criterion is considered, sligltly and without the name and the equivalence, in Exercice II-5-2 page 52. This precise reference was communicated by Laurent Miclo. For a more substantial treatment, the best remains the book by Claude Delacherie and Paul-André Meyer.
Bonjour,
It is not true that the span of a family of UI is UI, the span of 1 (i.e. R) is not UI. Is it correct?
Merci
Thank you very much, you are perfectly right. It does not work with the span. Actually it is a matter of tightness, because it works nicely with the convex envelope. I have updated the post accordingly.