Quadratic stochastic operators on Banach lattices

We study the convergence of iterates of quadratic stochastic operators that are mean monotonic. They are defined on the convex set of probability measures concentrated on a weakly compact order interval $$S = [0, f]$$S=[0,f] of a fixed Banach lattice F. We study their regularity and identify the limits of trajectories either as the “infimum” or “supremum” of the support of initial distributions.


Introduction
The theory of quadratic stochastic operators (q.s.o.) is rooted in works of Bernstein (cf. [8,9]). Their importance was rediscovered in 1950s by S. Ulam. His seminal report [28] prompted for many theoretical publications (cf. [4][5][6][7]11,12,16,20,21,29]), monographs (cf. [13,23]) and computer supported mathematical projects (cf. [3,14,15]) aiming to develop a unified theory illustrating (asymptotic) properties of q.s.o.. This subject has been intensively studied in mathematics and biology for almost three decades. In spite of that important questions have remained unsolved and the task is far from being completed. The asymptotic behavior of real or abstract mathematical models of biological systems, consisting with three or more biotypes, is not fully understood. We anticipate that q.s.o. will play in future an important role in genetics, population dynamics, social sciences as well as in other areas. Mathematically they perfectly suit to model the evolution of statistical configurations of all kinds of biotypes (genome, phenotypes e.t.c.) both for finite or continuous populations. The reader is referred to [13,17] for a comprehensive and updated review of the topic. The list of very recent articles and online presentations is long. Let us only mention [18,19,25,26], where the notion of quadratic stochastic operators is put into abstract vector (Banach) spaces with specific order or norm structures. In [24] the author studies nonhomogeneous Markov chains on ordered Banach spaces which are strongly linked to some picture of quadratic stochastic operators. Namely, as it has been proved in [7], generated by a quadratic stochastic operator V the so-called associated Markov chains may be efficiently used to determine the behavior of V n .
There has been much interest in recent years in self-organizing search methods in the q.s.o. field. Recently Ganikhodjaev, Saburov and Muhitdinov (see [15]) have generalized the notion of q.s.o. to bilinear forms on σ -additive measures on [0, 1]. In particular, points of the unit interval [0, 1] serve to code (continuum valued) traits attributed to each individual from a considered population. As usual children inherit their traits from (two) parents, who mate randomly. The paper [15] drops the Mendelian paradigm and introduces trends (however, only in a one-dimensional direction), which are steady in time. The authors proved in [15] regularity in the case when there is no mutation and an inherited trait comes (randomly) from one of its parents. In this paper we propose further extensions. In particular, we obtain regularity under essentially weaker constraints. Moreover, our methods applied to theorems and examples from [15] simplify existing proofs.
Instead of one particular trait we propose to characterize an individual by infinite dimensional set of parameters, encoded by vectors from an order interval S = [0, f ] in a fixed real Banach lattice (F, · , ≤). Let us very briefly recall basic notions necessary to formulate our results (regarding the theory of Banach lattices and other facts from functional analysis the reader is referred to [2] or [27]). A Banach space (F, · ) equipped with a partial order ≤ is a Banach lattice if the lattice operations (the infimum x ∧ y and the supremum x ∨ y) are well defined in F, satisfy axioms of Riesz spaces and are compatible with the norm topology (cf. [2], pp. 4, 5, 181 or [27], pp. 47-52). In particular, y ≤ x whenever |y| ≤ |x| in F. The modulus in F is defined as |x| = x ∨ (−x), and therefore |x| = x for all x ∈ F. The positive cone F + = {x ∈ F: 0 ≤ x} is a (weakly) closed subset of F (cf. [27], Proposition 5.2). Let us mention that for all x, y ∈ F we have x ∨ y = x+y+|x−y| 2 , and x ∧ y = x+y−|x−y| 2 . The order interval, with endpoints a ≤ b ∈ F, is defined as In the case F = R, with ordinary modulus | · |, order intervals are classical segments [a, b] = {x ∈ R: a ≤ x ≤ b}. It is well known that order intervals in Banach lattices are always bounded, convex and (weakly) closed. However, they are not weakly compact in general. Classical spaces (L p (μ), · p ) and (C(K ), · sup ) are important examples of Banach lattices. It follows from the Banach-Alaoglu theorem that order intervals are weakly compact if F is reflexive. In particular, if F = L p (μ), where 1 < p < ∞ (using different arguments also in L 1 (μ)), then order intervals are weakly compact. However, this does not hold for C(K ) or L ∞ (μ). We recall that the weak topology on a Banach space (F, · ) is the smallest topology T so that all norm continuous linear functionals ξ ∈ F are T continuous. The weak topology is denoted by T w and it is a locally convex topology generated by the basis consisting of neighborhoods U ξ 1 ,...,ξ n ,ε = {u ∈ F: |ξ j (u)| < ε, j = 1, . . . , n} of the zero vector 0 ∈ F, where n ∈ N, ξ 1 , . . . , ξ n ∈ F , and ε > 0. The net x α converges weakly to x in F if and only if for every fixed ξ ∈ F we have lim α ξ(x α ) = ξ(x). Weak compactness plays in the category of Banach lattices an important role (cf. [2], chapter 4.2).
In this paper we shall deal with order intervals (phase spaces) S ⊂ F, which are assumed to be compact and metrizable for the weak topology (in particular, by the Mazur theorem they are norm separable). Hence the weak and norm Borel structures on S coincide. We denote the Borel σ field in S by B. As usual δ x stands for the Dirac measure at x ∈ S. Compactness of S implies (cf. §6 in [10]) that the convex set P(S) of all probability measures μ, on the measurable space (S, B), is compact (metrizable) for the weak measure convergence. We recall that a sequence of probability measures μ n ∈ P(S) converges to μ ∈ P(S) in the weak convergence of measures if lim n→∞ S f dμ n → S f dμ holds for all T w continuous (bounded) functions f : S → R. Then we write μ n ⇒ μ. In order not to overexploit the term "weak" (especially in two different contexts) we shall use the notion of the Fortet-Mourier norm. We recall that given a finite Borel (signed) measure μ on S, the Fortet-Mourier norm is defined as μ FM = sup{| f dμ|: 0 ≤ f ≤ 1 and Lip f ≤ 1} (cf. [22], p. 48). As the reader may guess, given a continuous function f : where S is a fixed metric on S compatible with the relative weak topology T w | S on the set S (the choice of a specific metric S is not crucial). It is well known (see [22], p. 49, theorem 1.46) that on the space P(S) the convergence μ n ⇒ μ holds if and only if μ n − μ FM → 0. By bar(μ) we denote the barycenter of μ, as long as it exists. We understand it as the Pettis We denote i(μ) = inf supp(μ), s(μ) = sup supp(μ) and I (μ) = [i(μ), s(μ)] as long as all is well defined. For this, let us suppose that F has an order continuous norm (cf. [2], pp. 185-186) and supp(μ) is contained in an order interval I a,b = [a, b] ⊆ F + . Clearly, r n = x 1 ∧ · · · ∧ x n ∈ I a,b , if x j ∈ supp(μ), n = 1, 2, . . .. Hence r n ≥ a is bounded from below. Using induction method and the property of norm order continuity (cf. [2], Theorem 4.9, p. 186) we may construct a sequence (x j ) j≥1 such that lim n→∞ r n = r ∈ [a, b] exists and the norm r is as small as possible (we can take a dense sequence x j ∈ supp(μ) actually). Notice that r ≤ x for all x ∈ supp(μ).
If v ≤ x for all x ∈ supp(μ), then v ≤ x j for all j, hence v ≤ r n for all n and finally v ≤ r . In particular, r = i(μ) does exist (however not necessarily r ∈ supp(μ)). Similarly we obtain the existence of s(μ).

Basics on quadratic stochastic operators
Given a weakly compact metrizable set S ⊆ F, let P = {P(x, y, ·)} x,y∈S be a family of set functions (defined on the measurable space (S, B), where B stands for the Borel σ -field) satisfying: Then the system P = {P(x, y, ·)} x,y∈S is called a quadratic transition probability function. If moreover it satisfies Let us mention briefly, that quadratic transition functions can be used to define perhaps the simplest class of nonlinear Markov processes {ξ k } k≥0 . We will not dwell on this subject but rather focus only on the evolution and weak limits of distributions P(ξ k ∈ ·). However, the problem of behavior of trajectories ξ k (ω), or the rate of convergence, remains untouched.
Any measurable quadratic transition probability function P may be uplifted to a bilinear mapping of P(S). Namely, define It is not hard to prove (cf. [4,7] as long as μ n ∈ P(S) converges in the norm · FM Indeed, let μ n − μ FM → 0 and g: S → R be weakly continuous. Then is weakly continuous. It is well known that convergence in the Fortet-Mourier norm is preserved if we bring it to direct products. In other words if μ n − μ FM → 0 and ν n − ν FM → 0 then The natural notion of regularity, in the context of q.s.o., was introduced very early (see [17,23]). We adopt it for the Fortet-Mourier topology.
Depending on the (topological) point of view different modes of convergence are distinguished. To connect our results with other contemporary studies we recall (cf. [4,7]) the following notions: Strong and norm mixing q.s.o. and their geometric structure have been recently described in [4,5,7]. The weak mixing is studied in [6]. However, Bartoszek et al. [6] is restricted to kernel quadratic stochastic operators. In particular, P(x, y, ·) are assumed to be absolutely continuous with respect to the Lebesgue measure and additionally z P(x, y, dz) = x+y 2 for all x, y (roughly speaking-the offspring is on average the mean of the parents). Such operators are called centred kernel q.s.o. On the other hand, the models considered in [6] generally apply to unbounded phase spaces. Their weak limit of V n ( f ) may be (depending on f ) both discrete (Dirac δ) or other probability measures, depending to what extent the CLT works. Mathematical methods used in their proofs come from the theory of characteristic functions, making it difficult to obtain closed form statements about the limits. In our paper we focus solely on weak regularity, assuming bounded domains and we look what happens if the evolution has a trend. In particular, our quadratic transition probabilities are not centred. We may consider our approach as generalizations of [15] and complementary to [6].
If for all x, y ∈ S the measure P(x, y, ·) is a convex combination of δ x , δ y , then the q.s.o. V is called Volterra. Clearly the Volterra q.s.o. can only model the mutation free evolution. The Mendelian situation occurs when P(x, y, ·) = 0.5δ x + 0.5δ y . We notice that in the last case V(μ) = μ for all μ ∈ P(S).
Our approach brings further extensions, with two folded generalizations. Firstly, in the case of Lebesgue operators V on S = [0, 1], we only assume that P(x, y, ·) = α(x, y)δ x∧y (·) + (1 − α(x, y))δ x∨y (·), where α(x, y) > 0.5 for all x = y ∈ [0, 1] (or α(x, y) < 0.5 for all x = y ∈ [0, 1] respectively). We will show that this condition may be relaxed further (see our Corollaries 5.1 and 5.2). In particular, we do not require the Volterra condition. Secondly, we shall extend Ganikhodjaev's, Saburov's and Muhitdinov's results from [15] to general phase spaces, assuming that S = I is a weakly compact order interval. Let us recall (cf. [2], Theorem 4.9) that order intervals in Banach lattices with order continuous norm are weakly compact. And finally, in general Banach lattices the notion of Lebesgue q.s.o. may split in different directions. For this let us only mention that α(x, y)δ x∧y (·) + (1 − α(x, y))δ x∨y (·) is not necessarily a convex combination of δ x , δ y .

Monotonicity of means and regularity Theorem 3.1 Let I = [0, f ] ⊂ F be a weakly metrizable order interval in a Banach lattice (F, · , ≤) with order continuous norm and let {P(x, y, ·)} x,y∈I be a quadratic transition probability function satisfying:
(3.1) I z P(x, y, dz) ≤ x+y 2 , for all x, y ∈ I .
Let n j ∞ be any increasing sequence such that V n j (μ) converges weakly to some ν ∈ P(S). By the Feller assumption It follows that We obtain bar(V(ν)) = bar(ν). Hence for any positive ξ ∈ F we get By positivity of ξ and condition (3.1) we get ξ( x+y 2 − I z P(x, y, dz)) = 0 for ν × ν almost all (x, y). Now, applying the additional assumption (3.2), we get x = y for ν × ν almost all (x, y). Hence ν = δ c(μ) . Since n j was arbitrary (with only restriction that V n j (μ) converges), thus lim n→∞ V n (μ) = δ c(μ) .

Finite dimensional case
In the finite dimensional case, when F = R d and the q.s.o. is Lebesgue, the last result may be strengthen. We assume that in R d we have the standard order, so the positive cone is R d The lattice norm · may be taken arbitrary as all norms on finite dimensional vector spaces are equivalent (or apply Corollary 4.4 from [2]).
Clearly i(V(μ)) = i(μ). Therefore, we can iterate the above estimation to obtain We have already noticed that i(μ) = lim n→∞ r n , where r n = x 1 ∧ · · · ∧ x n , for some sequence x j ∈ supp(μ). It follows from the definition of Lebesgue quadratic transition function that r n ∈ supp(V n (μ)). Now for arbitrarily fixed ε > 0 we find Then for every j ∈ J we have u k, j = s(μ) j as well. Define * = min{s(μ) j − r k, j : j / ∈ J }. Clearly the set U (r k ) = {x ∈ I : x − r k ∞ < } is an open neighborhood of r k (both for · and · ∞ norms) in the relative topology on I . Notice that, whenever < ε * , then U (r k ) ⊆ [i(μ), u k ] = W u k . In fact, for all j ∈ J if x ∈ U (r k ), then x j ≤ s(μ) j = u k, j . On the other hand, if j / ∈ J , then x j < r k, j + = u k, j + εr k, j − εs(μ) j + = u k, j − ε(s(μ) j − r k, j ) + ≤ u k, j − ε * + < u k, j .
This implies c(μ) ∈ [i(μ), u k ] and by the triangle inequality We get c(μ) = i(μ) as ε > 0 may be as small as we wish.
Problem Describe weakly regular Lebesgue quadratic operators in the onedimensional case, for general symmetric (continuous) α(x, y), and identify limit measures.
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.