Asymptotic analysis of a quantitative genetics model with nonlinear integral

. — We study the asymptotic behavior of stationary solutions to a quantitative genetics model with trait-dependent mortality and a nonlinear integral reproduction operator with a parameter describing the deviation between the oﬀspring and the mean parental trait. Our asymptotic analysis encompasses the case when the parameter is typically small. Under suitable regularity and growth conditions on the mortality rate, we prove existence and local uniqueness of a stationary proﬁle that gets concentrated around a local optimum of mortality, with a Gaussian shape having small variance. Our approach is based on perturbative analysis techniques that require to describe accurately the correction to the Gaussian leading order pro-ﬁle. Our result extends previous results obtained with linear reproduction operator


Introduction
We investigate solutions (λ ε , F ε ) ∈ R × L 1 (R) of the following stationary problem: where B ε (f ) is the following non linear, homogeneous integral operator associated to the infinitesimal model [Fis18,BEV17]: In the context of quantitative genetics, the variable z denotes a multi-dimensional phenotypic trait, F ε (z) is the phenotypic distribution of the population and m(z) is the (trait-dependent) mortality rate which results in the selection of the fittest individuals.
The mixing operator B ε acts as a simple model for the inheritance of quantitative traits in a population with a sexual mode of reproduction.As formulated in (1.1), it is assumed that offspring traits are distributed normally around the mean of the parental traits (z 1 +z 2 )/2, with a variance which remains constant across generations, here ε 2 /2.
We are interested in the asymptotic behaviour of the trait distribution F ε as ε 2 vanishes.
This asymptotic regime was investigated thoroughly for various linear operators B ε associated with asexual reproduction such as for instance the diffusion operator F ε (z) + ε 2 ∆F ε (z), or the convolution operator 1 ε K( z ε ) * F ε (z) where K is a probability kernel with unit variance, see [DJMP05,Per07,BP07,BMP09,LMP11] for the earliest investigations, see further [MM15,Mir18,BGHP18] for the case of a fractional diffusion operator (or similarly a fat-tailed kernel K), and see further [Mir13, MP15, BM15, LL17, GM17, Mir17, MG18, CHM + 18] for the interplay between evolutionary dynamics and a spatial structure.In the linear case, the asymptotic analysis usually leads to a Hamilton-Jacobi equation for the Hopf-Cole transform U ε = −ε log F ε .This yields an original problem with non-negativity constraint that requires a careful well-posedness analysis [MR15,CL18].
From a mathematical viewpoint, the model (1.1) received recent attention in the field of probability theory [BEV17] and integro-differential equations [MR13,Rao17].In the latter couple of articles, a scaling different from (1.1) is studied: the variance is of order one, but there is a large reproduction rate that enforces the relaxation of the phenotypic distribution towards a Gaussian local equilibrium.Macroscopic equations are rigorously derived in [Rao17], in the case of an additional spatial structure, in the spirit of hydrodynamic limits for kinetic equations.
In a different context, a similar collisional operator as B ε (1.1) was introduced in the modelling of self-propelled particles with alignment interactions, see for instance [BDG06,DFR14].When two particles interact they tend to align with the mean velocity, with some possible noise.However, there are some discrepancies with our case study, since the operator is not conservative in our case, by definition of a reproduction operator.Moreover, it is normalized by the total mass of the phenotypic distribution: f (z 2 ) dz 2 .The rationale behind this choice is that during the mating process, the first parent chooses the trait of its partner depending on its frequency in the population.This is the neutral case without any assumption about assortative mating.Moreover, this dependency upon the frequency rather than the density discards any small population effects that could arise from a quadratic collisional operator.Such homogeneity of degree one is a key ingredient in our analysis.
The problem (P F ε ) is equivalent to the existence of special solutions of the form exp(λ ε t)F ε (z), for the following non-linear but one-homogeneous equation which will be the subject of future work: Alternatively speaking, the problem (P F ε ) expresses the balance between selection via trait-dependent mortality m(z), and the generation of diversity through reproduction B ε .The scalar λ ε is analogous to the principal eigenvalue of the operator B ε −m.However, it might not be unique, as in the Krein-Rutman theory, see Corollary 1.5.It measures the global fitness of the population: the population grows exponentially fast λ ε > 0 when the reproduction term B ε dominates, while it declines exponentially fast λ ε < 0 when the mortality m out-competes the reproduction.This preliminary analysis on the stationary profile paves the way for a systematic analysis of various quantitative genetics models, including time marching problems and the combination of multiple effects.
Our work is inspired by similar asymptotics in the case of linear operator B ε , see the seminal work by [DJMP05] and references cited above.Accordingly, our goal is to analyze problem (P F ε ) in the limit of vanishing variance ε 2 → 0. As there is few diversity generated in this asymptotic regime, we expect that the variance of the distribution solution F ε vanishes as well.Actually, there is strong evidence that the leading order profile of F ε is a Gaussian distribution with variance ε 2 .As a matter of fact, any Gaussian distribution with variance ε 2 is invariant by the infinitesimal operator B ε in the absence of selection (m ≡ 0, λ ε = 1) [TB94,MR13].This motivates the following decomposition of the solution: The latter (1.3) is similar to the Hopf-Cole transform used in the asymptotic analysis of adaptative evolutionary dynamics in asexual populations.In our case U ε is a corrector term that measures the deviation from the leading Gaussian distribution of variance ε 2 .Our analysis reveals that selection determines the center of the distribution z 0 , as expected, and also reshapes the distribution F ε via the corrector U ε .
The operator B ε is invariant by translation.Up to a translation of m, we may assume that the leading order Gaussian distribution is centered at the origin, i.e., z 0 = 0. Next, up to a change of λ ε ← λ ε + m(0), we may assume that m(0) = 0. Note that we may also assume U ε (0) = 0 without loss of generality, as the original problem is homogeneous.
Plugging the transformation of (1.3) into (P F ε ) yields the following equivalent problem for U ε : The residual term from the integral contribution is the following non-local term I ε (U ε ), see Section 2.1 for the details of the derivation.Set This decomposition appears to be relevant because a formal computation shows that I ε (U ε ) → 1 as ε → 0. Establishing uniform convergence is actually a cornerstone of our analysis.Thus for small ε, the problem (P U ε ) is presumably close to the following corrector equation, obtained formally at ε = 0: Interestingly, this finite difference equation admits explicit solutions by means of an infinite series: However, two difficulty remains: identify (i) the linear part γ 0 ∈ R d and (ii) the unknown λ 0 ∈ R. On the one hand, the linear part γ 0 cannot be recovered from (P U 0 ) because linear contributions cancel in the right-hand-side of (P U 0 ).Thus, identifying the coefficient γ 0 will be a milestone of our analysis.On the other hand, two important conditions must be fulfilled to guarantee that the series above converges, namely: The latter is a constraint on the possible translations that can be operated: the origin must be located at a critical point of m.The former prescribes the value of λ 0 J.É.P. -M., 2019, tome 6 accordingly.These two conditions are necessary conditions for the resolvability of problem (P U 0 ).Indeed, evaluating (P U 0 ) at z = 0, we get the first identity.Next, differentiating and evaluating again at z = 0, we get the second identity.
In the sequel we make this formal discussion rigourous, following a perturbative approach for ε small enough.Before stating our main result, we need to prescribe the appropriate functional space for the corrector U ε .
Definition 1.1 (Functional space for U ε ).-For any positive parameter α 2/5, we define the functional space For any bounded set K of E α , we use the notation K α = sup u∈K u α .Occasionally we use the notation ϕ α for the weight function ϕ α (z) = (1 + |z|) α .Although 2/5 is not the critical threshold, it happens that the exponent α cannot be taken too large in our approach.We set implicitly α = 2/5 in the following results, however we leave it as a parameter to emphasize its role in the analysis, and to pinpoint the apparition of the threshold.Note that α > 0 is required in our approach, as one constant collapses in the limit α → 0 (see estimate (5.6) below).
Then, we detail the assumptions on the selection function m.
Definition 1.2 (Assumptions on m).-The function m is a C 3 (R d ) function, bounded below, that admits a local non-degenerate minimum at 0 such that m(0) = 0, and there exists µ 0 > 0 such that D 2 m(0) µ 0 Id in the sense of symmetric matrices.Furthermore we suppose that (∀z) 1 + m(z) > 0 and Remark 1.3.-Our result is insensitive to the sign of the local extremum.Indeed, one can replace the hypothesis that m admits a "local non degenerate minimum" at 0 with a "local non degenerate maximum" at 0, and that there exists µ 0 < 0 such that D 2 m(0) µ 0 Id.However, we leave our main assumption as in Definition 1.2 as it is the most natural one from the point of view of stability analysis for the time-marching problem (1.2).
The condition (1.6) is clearly verified if m is a polynomial function.It would be tempting to write, in short, that log(1 + m) ∈ E α , which is indeed a consequence of (1.6).However, the latter condition also contains the decay of the first order derivative D log(1 + m) with rate |z| −α , which is not contained in the definition of E α (1.5) for good reasons.
We also introduce the subset E α 0 : (1.7) The main result of this article is the following theorem.
Theorem 1.4 (Existence and convergence) (i) There exist K 0 a ball of E α , and ε 0 a positive constant, such that for any ε ε 0 , the problem (1.11) Moreover, the convergence U ε → U 0 is locally uniform up to the second derivative.
An immediate remark is that the regularity required by (1.6), and particularly the C 3 regularity of m, is consistent with formula (1.10) which involves the pointwise value of third derivatives of m.Alternatively speaking we think that our result is close to optimal in terms of regularity.
It is important to notice that our result holds true for any local minimum z 0 such that One should define the functional spaces E α and E α 0 accordingly (and particularly replace the conditions u(0) = 0 and Du(0) = 0 by the conditions u(z 0 ) = 0 and Du(z 0 ) = 0), and then adapt (1.9)-(1.10)as follows, for the one-dimensional case: where γ 0 is defined by the same formula as in (1.11) but evaluated at z 0 .Immediately, one sees that the compatibility condition (1.12) is necessary to have the positivity of the term inside the log in (1.13).As a consequence, we have: Numerical simulations of the stationary problem (P F ε ) with ε = 0.1 in an asymmetric double-well mortality rate (grey line).The numerical equilibrium is in yellow plain line.The only difference between the two simulations is the initial data (red dashed line).The simulations illustrate the lack of uniqueness for problem (P F ε ).
Corollary 1.5 (Lack of uniqueness).-If the selection function m has at least two different local non-degenerate minima that verify the compatibility condition (1.12), there exists at least two pairs (λ ε , F ε ) solutions of problem (P F ε ) for ε small enough.
We performed numerical simulations to illustrate this phenomenon (see Figure 1).The function m is an asymmetric double well function.We solved the time marching problem (1.2) but on the renormalized density F ε / F ε in order to catch a stationary profile.We clearly observed the co-existence of two equilibria for the same set of parameters, that were obtained for two different initializations of the scheme.However, let us mention that the question of uniqueness in the case of a convex selection function m is an open question, to the extent of your knowledge.
This result is in contrast with analogous eigenvalue problems where B ε is replaced with a linear operator, say F ε + ε 2 ∆F ε as in various quantitative genetics models with asexual mode of reproduction, see e.g.[BMP09] and references mentioned above, or in the semi-classical analysis of the Schrödinger equation, see e.g.[DS99].In the linear case, λ ε ∈ R and F ε 0, F ε ≡ 0 are uniquely determined (up to a multiplicative constant for F ε ) under mild assumptions on the potential m.This is the signature that B ε (1.1) is genuinely non-linear and non-monotone, so that possible extensions of the Krein-Rutman theorem for one-homogeneous operators, as in [Mah07], are not applicable.
The existence part (i) has already been investigated in [BCGL17] using the Schauder fixed point theorem and very loose variance estimates.But the approach was not designed to catch the asymptotic regime ε → 0. The current methodology gives much more precise information on the behavior of the solutions of the problem (P F ε ) in the regime of vanishing variance.Theorem 1 provides a rigorous background for the connection between problem (P U ε ) and problem (P U 0 ) in a perturbative setting.It justifies that the problem (P F ε ) is well approximated by the solution (λ 0 , U 0 ) of the problem (P U 0 ).Quite surprisingly, the value γ 0 of the linear part of the corrector function U 0 is resolved during the asymptotic analysis although it cannot be obtained readily from problem (P U 0 ) as mentioned above.It coincides with the heuristics of [BBC + 18] where the same coefficient was obtained by studying the formal expansion up to the next order in ε 2 : U ε = U 0 + ε 2 U 1 + o(ε 2 ), and by identifying the equation on U 1 in which the value of γ 0 appears as another compatibility condition.Here the value of γ 0 is obtained directly as a by-product of the perturbative analysis.
As mentioned above, our approach is very much inspired, yet different to most of the current literature about asymptotic analysis of asexual models, where the limiting problem is a Hamilton-Jacobi equation, see [Per07] for a comprehensive introduction, and references above.To draw a parallel with our problem, let us consider the case where B ε (f ) is replaced with the (linear) convolution operator K ε * f , where the kernel has the scaling property K ε = (1/ε)K (•/ε), and K is a probability distribution kernel.There, the small parameter ε measures the typical size of the deviation between the offspring trait and the sole parental trait.In this context, it is natural to introduce the Hopf-Cole transform U ε = −ε log F ε .Then, the problem is equivalent to the asymptotic analysis of the following equation as ε → 0: For this model, it is known that U ε converges towards the viscosity solution of a Hamilton-Jacobi equation [BMP09]: (1.15) Note that the limiting equation on U 0 (1.15) can be derived formally from (1.14) by a first order Taylor expansion on U ε .There are two noticeable discrepancies between the asexual case (1.14)-(1.15)and our problem involving the infinitesimal model with small variance.Firstly, ε plays a similar role in both cases, i.e., measuring typical deviations between offspring and parental traits.However, the appropriate normalization differs by a factor ε: it is −ε log F ε in the asexual case, whereas it is −ε 2 log F ε in our context, see (1.3).This scaling difference is the signature of major differences between the two problems (asexual vs. sexual).Secondly, the two limiting problems (1.15) and (P U 0 ) have completely different natures: a Hamilton-Jacobi PDE in the asexual case, vs. a finite difference equation in the sexual case.Moreover, due to the lack of a comparison principle in the original problem (P F ε ), we could not envision a similar notion of viscosity solutions for (P U 0 ).Instead, we use rigid contraction properties and a suitable perturbative analysis to construct a unique strong solution near the limiting problem, as depicted in Figure 1.2.[MR13] observed that the infinitesimal operator B ε alone enjoys a uniform contraction property with respect to the quadratic Wasserstein distance, with a factor of contraction 1/2.Recently, this was used by [MR15] to perform a hydrodynamic limit in a different regime than the one under consideration here.However, the combination of B ε with a zeroth-order heterogeneous mortality m(z) seems to destroy this nice structure (details not shown).
The next section is devoted to the reformulation of problem (P U ε ) into a fixed point problem, introducing a set of notation and the strategy to prove Theorem 1.4.The organization of the paper is postponed to the end of the next section.
Up until the last part of the article we implicitly work in dimension d = 1, for the readers' convenience.In Section 7 we pinpoint the few elements of the proof that are specific to the one-dimensional case and give an extension to the higher-dimensional case in order to complete the proof of Theorem 1.4.
Acknowledgements.-The authors are grateful to Laure Saint-Raymond for stimulating discussions at the early stage of this work.They are thankful to Sepideh Mirrahimi for pointing out the extension of the result to local maxima of the selection function, see Remark 1.3.

Reformulation of the problem as a fixed point
2.1.Looking for problem (P U ε ).-The equivalence between problem (P F ε ) and problem (P U ε ) through the transform (1.3) is not immediate.It is detailed in [BBC + 18], but we recall here the key steps for the sake of completeness.Plugging (1.3) into problem (P F ε ) yields When ε → 0, we expect the numerator integral to concentrate around the minimum of the principal term that is: We introduce the notation Using the change of variable (z 1 , z 2 ) = (z + εy 1 , z + εy 2 ), we obtain the following equation: Definition 2.1.-We denote by Q the following quadratic form: J.É.P. -M., 2019, tome 6 It is the residual quadratic form after our change of variable.We notice that 1 √ 2π exp(−Q) is the density of a bivariate normal random variable with covariance matrix At the denominator of (2.1) naturally arises N the density function of a N (0, 1) random variable.Finally, (2.1) is equivalent to problem (P U ε ): simply by conjuring 2U ε (z/2) at the numerator and U ε (0) at the denominator, resulting into the definition of the remainder I ε (U ε ) (1.4) that will be controlled uniformly close to 1 in all our analysis.
In the next section we explain how we reformulate the problem (P U ε ) into a fixed point argument in order to use a Banach-Picard fixed point theorem which prove our results rigorously.
2.2.Some auxiliary functionals and the fixed point mapping.-This section is devoted to the derivation of an alternative formulation for problem The first step is to dissociate the study of λ ε and U ε .We first evaluate the problem (P U ε ) at z = 0.It yields the following condition on λ ε , since m(0) = 0: Considering the terms I ε as a perturbation, we divide problem (P F ε ) by which is positive, and we take the logarithm on each side.Then we obtain the following equation, considering (2.3): It would be tempting to transform (2.4) into a fixed point problem by inverting the linear operator in the left-hand-side.However, the latter is not invertible as it contains linear functions in its kernel.Therefore we are led to consider linear contributions separately.
Our main strategy is to decompose the unknown U ε under the form This is consistent with the analytic shape of our statement in (1.10), where γ 0 and V 0 have quite different features with respect to the function m.
Next, it is natural to differentiate (2.4).One ends up with the following recursive equation for every z ∈ R One simply deduces that, if U ε exists and is regular, then we must have: One can formally integrate back the previous equation to obtain At this stage we formally identify: -U ε (0) = 0, since U ε ∈ E α .This is not a loss of generality by homogeneity since F ε is itself defined up to a multiplicative constant in problem (P F ε ). - The real number γ ε is unknown at this stage, but it needs to verify some compatibility condition to make the series converging in (2.6)-(2.7).In particular, if we evaluate (2.6) at z = 0 we obtain that γ ε must satisfy We will solve (2.8) using an implicit function theorem in order to recover the value γ ε associated with a given V .Beforehand, we introduce the following notation: Definition 2.2 (Finite differences operator D ε ).-We define the finite differences functional D ε as We introduce the following auxiliary functional which makes the link between γ ε and V .
The implicit relationship (2.8) is equivalent to J ε (γ ε , V e ) = 0. From this perspective, the following result is an important preliminary step.
Proposition 2.4 (Existence and uniqueness of γ ε ).-For any ball K ⊂ E α 0 , there exists ε K , such that for all ε ε K and for any V ∈ K, there exists a unique solution γ ε (V ) to the equation: where the bound |γ ε (V )| R K is defined as Next we define the main quantity we will work with: the double integral I ε which is the rescaled infinitesimal operator.For convenience we define it as a mapping on E α 0 .It is compatible with (1.4) because of the decomposition (2.5).
Definition 2.5 (Auxiliary functional I ε ).-We define the functional Finally, in view of (2.7) and (2.5), we see that V ε must be a solution of this implicit equation: (2.12) This justifies the introduction of our central mapping, upon which our fixed point argument will be based.
Definition 2.6 (Fixed point mapping).-We define the mapping 2.3.Reformulation of the problem.-We are now in position to write our main result for this Section: Theorem 2.7 (Existence and uniqueness of the fixed point) There is a ball K 0 ⊂ E α 0 and a positive constant ε 0 such that for every ε ε 0 , the mapping H ε admits a unique fixed point in K 0 .
To conclude, it is sufficient to check that solving problem (P U ε ), on the ball K 0 , and seeking a fixed point for H ε in K 0 are equivalent problems for ε ε 0 small enough.
Proposition 2.8 (Reformulation of the problem (P U ε )).-There is a ball K 0 of E α , and a positive constant ε 0 such that for every ε ε 0 , the following statements are equivalent:

Moreover, the statement of Theorem 2.7 holds true in the set
The main mathematical difficulties are stacked into Theorem 2.7.The rest of the article is organized as follows: -In section 3, we justify why the function γ ε is well defined in Proposition 2.4.-Then in section 4, we provide the main properties and the key estimates of the nonlocal operator I ε .We point out why this term plays the role of a perturbation between problem (P U ε ) and problem (P U 0 ).In section 4.2 we prove crucial contraction estimates.
-Those estimates are the main ingredients of the proof of properties of H ε in Section 5: most notably the finiteness of H ε (V ), and the fact that H ε is a contraction mapping.
-This allows us to establish the proof of Theorem 2.7 and Proposition 2.8, and finally to come back to the proof of our main result Theorem 1.4 in the sections 6.1 and 6.2.
-Section 7 is devoted to those specific arguments that require an extension to the higher dimensional case d > 1.
3. Well-posedness of the implicit function γ ε 3.1.Heuristics on finding γ ε .-We consider V ∈ E α 0 , and we look for solutions γ ε of J ε (γ ε , V ) = 0, or equivalently: in accordance with (2.9).We will see here how a Taylor expansion of the right-handside around ε = 0 helps to understand why it defines a unique γ ε in a given interval for small ε.We will show formally why J ε (•, V ) can be uniformly approximated by a non-degenerate linear function for small ε.
We expand the right-hand-side with respect to ε: Then solving we get the expression: .
These heuristics are consistent with the statement in Theorem 1.4, up to the relation between V 0 and m that can be easily read out from (1.11).Note that the denominator involves ∂ 2 z V (0), so that the local convexity of V should be controlled uniformly during our construction.This is the purpose of the restriction in E α 0 (1.7).In the following, we provide estimates that turn these heuristics into a rigorous proof.
3.2.Proof of Proposition 2.4.-The aim of this section is to prove the existence and uniqueness of γ ε (V ) stated in Proposition 2.4.We first start with a Lemma providing some useful estimates on the function J ε .Combining these estimates with a continuity and monotonicity arguments, we will be able to prove the Proposition 2.4.Lemma 3.1 (Estimates of J ε ).-For any ball K ⊂ E α 0 , there exists ε K > 0, such that for all ε ε K and V ∈ K, the following estimate holds true for all g in the interval (−R K , R K ): where, in the former expansion, the variable y i is a by-product of Taylor expansions and is such that Remark 3.2.-We prove the uniqueness of γ ε on a uniformly bounded interval.One may think it is a strong restriction not to look at large γ ε .It is in fact a natural restriction as we have by definition We postpone the proof of the technical Lemma 3.1 at the end of this section and we first use it to prove the Proposition 2.4: for ε small enough.Integrating (3.4) with respect to g, we obtain where it is important to notice that the perturbation O(ε) is uniform with respect to ε Therefore, J ε is uniformly increasing with respect to g on (−R K , R K ).Moreover, the choice of R K is such that for ε small enough, and similarly, J ε (−R K , V ) < 0. Finally, there exists a unique Proof of Lemma 3.1.-Let K be a ball of E α 0 of radius K α .In section 3.1, we have used formal Taylor expansions to get a formula for γ ε (V ), morally valid when ε = 0.The idea here is to write exact rests to broaden the formula for small but positive ε.
J.É.P. -M., 2019, tome 6 So, it remains Clearly the last two contributions are uniform O(ε) for V ∈ K and ε ε K small enough.Indeed, the term P is at most quadratic with respect to y i (3.6), so Q + θε 2 P is uniformly bounded below by a positive quadratic form for ε small enough.

Proof of expansion (3.4).
-The first step is to compute the derivative of J with respect to g: Similar Taylor expansions as above yields: where P = g(y 1 + y 2 ) + y 1 ∂ z V (ε y 1 ) + y 2 ∂ z V (ε y 2 ).Interestingly, the leading order term does not cancel anymore, and it remains: The justification that the remainder is a uniform O(ε) is similar as above, except that now P has a linear part depending on g, but the latter is assumed to be bounded a priory by R K .
4. Analysis of the perturbative term I ε 4.1.Lipschitz continuity of some auxiliary functionals.-The function I ε is crucially involved in the definition of the mapping H ε .Thus to prove any contraction property on this mapping we will need Lipschitz estimates about I ε and the three first derivatives of its logarithm.But first we show that I ε really plays the role of a perturbative term between problem (P U ε ) and problem (P U 0 ) that converges to 1 uniformly as ε → 0.
Proposition 4.1 (Estimation of I ε ).-For every K ball of E α 0 , for every δ > 0, there exists a constant ε δ that depends only on K and δ, such that for every ε ε δ and for every V ∈ K: We deduce from this lower and upper estimates that the whole I ε (V ) converges uniformly to 1 as ε → 0.
Next, we show Lipschitz continuity of various quantities of interest.
Proposition 4.2 (Lipschitz continuity of γ ε ).-For every ball K ⊂ E α 0 , there exist constants L K (γ), and ε K , depending only on K, such that for all We argue by means of Fréchet derivatives: let s ∈ (0, 1), and consider the following computation: where the Fréchet derivative of J with respect to V is: We perform similar Taylor expansions as in (3.5), We proceed as in the previous section for the exponential term: there exists θ = θ(y 1 , y 2 ) ∈ (0, 1) such that exp(−εP ) = 1 − εP exp(−θεP ), where P = γ(y Again, the crucial point is the cancellation of the O(ε −1 ) contribution in (4.2), as in (3.7)What remains is of order one or below, and one can easily show that there exists C K such that On the other hand, we have already established that ∂ γ J ε (γ, V ) = ∂ 2 z V (0)/2 + O(ε) in Lemma 3.1.Consequently, integrating (4.1) from s = 0 to 1, we find: We deduce from the previous estimates and the local convexity condition in (1.7) that for some C K and ε ε K small enough.
In turn, Proposition 4.2 implies the Lipschitz continuity of I ε as a function of V .
Proposition 4.3 (Lipschitz continuity of I ε ).-For every ball K of E α 0 , there exist constants ε K , C K depending only on K, such that for all ε ε Proof.-The Lipschitz continuity of I ε with respect to V can be proven by composition of Lipschitz functions.With the same notations as in the proof of Proposition 4.2, and with the shortcut notation I ε = A ε /B ε to separate the numerator from the denominator in (2.11) we have, where we have simply written dz , and where G V ε denotes the exponential weight: We deduce that A ε is such that: As the weight G V ε is uniformly close to a positive quadratic form for small ε, we find that the numerator has a Lipschitz constant of order ε uniformly with respect to z: The same holds true for the denominator B ε .In addition, a direct by-product of the proof of Proposition 4.1 is that A ε and B ε are uniformly bounded above and below by positive constants for ε small enough.Consequently, the quotient It is useful to introduce the probability measure dG V ε induced by the exponential weight G V ε : As a consequence of the previous estimates, we obtain the following one.
Lemma 4.4 (Lipschitz continuity of dG V ε ).-For every ball K of E α 0 , there exist constants ε K , C K depending only on K, such that for all ε ε J.É.P. -M., 2019, tome 6 Furthermore, under the same conditions, we have the following bound, uniform with respect to z ∈ R: Proof.-We first prove (4.5): the function Therefore, its integral over (y 1 , y 2 ) ∈ R 2 converges to 1 as ε → 0, and there exists ε K depending on K such that G V ε (y 1 , y 2 , z) dy 1 dy 2 4/ √ 2π for ε ε K .This leads to (4.5).
In order to obtain (4.4), we proceed as in the proof of Proposition 4.3, as the denominator of dG V ε is the numerator A ε of I ε .For the Lipschitz continuity of the numerator of dG V ε , we find: This concludes the proof of (4.4).
To conclude, we have established in this section that I ε is a perturbative term, both in the uniform sense I ε (V ) → 1, and in the Lipschitz sense: In addition, we have proven a similar Lipschitz smallness property for a probability distribution dG V ε that will appear frequently in our contraction estimates.

Contraction properties (first part).
-On the way to estimating the fixed point mapping H ε (2.13), we need good estimates on the logarithmic derivatives of I ε .For that purpose, we introduce the following quantities for i = 1, 2, 3: For the sake of conciseness, we omit sometimes the dependency with respect to y 1 , y 2 in the notations, as for instance: The following notation with a duality bracket is useful: Indeed, for any V ∈ E α 0 , we have: (4.7) Similarly: And finally: In order to obtain estimates on W (i) it seems natural from the previous pattern of differentiation to begin with estimates on the symmetric difference of the derivatives of V .
Lemma 4.5.-For any V ∈ E α , and (y 1 , y 2 ) ∈ R 2 , we have: It is important to notice that the first two right-hand-sides (resp.first and second derivatives) are of order ε.The third one is larger but controlled by 2 α−1 < 1.This is the first occurrence of the contraction property we are seeking.This is the main reason why we make the analysis up to the third derivatives.
Proof of (4.9).-By Taylor expansions, we have: where Using the definition of • α (1.5), we obtain Since we chose α < 1, | • | α is sub-additive.Thus, we get By symmetry of the role played by y 1 and y 2 , we have proven equation (4.9).

J.É.P. -M., 2019, tome 6
Proof of (4.10).-The second estimate is a consequence of the first one, applied to the derivative of V .Notice that it is allowed as E α 0 enables control of derivatives up to the third order.
Proof of (4.11).-We must be a little more careful in the estimations of the third estimate (4.11), because we cannot go up to the fourth derivative in the Taylor expansions.This is why we do not have an ε bound, but we gain a contraction factor instead.We have We bound separately each term using again the sub-additivity of |•| α .For ε 1: Summing it all up, one ends up with:

This is precisely equation (4.11).
The following proposition is a first step towards contraction properties that will be established in section 5.For convenience, we introduce the following notation: There exists constants ε K , C K depending only on K such that for all ε ε K , we have: It is also possible to get estimates on W (i) ε (V ) itself, with the same hypotheses.This is useful to prove the invariance of certain subsets of E α 0 .
Proposition 4.7.-With the same setting as in Proposition 4.6, we also have: We do not give the details of the proof of the latter Proposition, since it is a straightforward adaptation of Proposition 4.6.Actually, we cannot readily apply Proposition 4.6 to (V 1 , V 2 ) = (V, 0) as 0 / ∈ E α 0 , because of the additional condition (1.7) on ∂ 2 z V (0) which is required to prove boundedness and Lipschitz continuity of γ ε .
Proof of Proposition 4.6.-The proof of theses inequalities is quite tedious because of the numerous non-linear calculations.However, the technique is similar for each inequality, and consists in separating the fully non linear behavior from the quasilinear parts of the left-hand-sides of equations (4.12) to (4.14).
Proof of (4.12).-This is the easiest part, because it is quasi-linear with respect to V .Indeed, we have We reformulate it in two parts, one involving V 1 − V 2 , and the other involving For the first contribution in (4.15), we apply directly Lemma 4.5 to V 1 − V 2 : For the last inequality we used equation (4.5), which enables to bound uniformly the measure dG V ε with respect to z.From Lemmas 4.4 and 4.5, there exists ε K and C K such that for ε ε K , the second contribution in the right-hand-side (4.15) satisfies The last integral is uniformly bounded for ε small enough, involving moments of a Gaussian distribution.Therefore, the whole quantity is bounded by ε 2 C K V α , uniformly with respect to z.This concludes the proof of equation (4.12).
Proof of (4.13).-To begin with, we have We split the difference into two, as in the previous part, The first contribution can be rearranged as follows, by factorizing the difference of squares: The term involving V 1 + V 2 is bounded uniformly in a crude way: (in fact it is bounded by a O(ε) uniformly with respect to z, but this detail is omitted here).Then, we apply Lemma 4.5 twice with V 1 − V 2 to obtain: To estimate B, the term involving the difference of measures dG V ε , we apply (4.4) and Lemma 4.5: . We find, exactly as above, that the quantity (1+|z|) α |B| is bounded by ε 2 C K V α .Combining both estimates on A, B, we deduce equation (4.13).
Proof of (4.14).-The full expression for W (3) ε is as follows: We split again in two pieces, one involving V 1 − V 2 , and the other one involving with We shall estimate all the contributions separately.Firstly, A 1 yields the contraction factor: The latter is the main contribution in (4.14).The remaining terms are lower-order contributions with respect to ε.For A 2 , we have For A 3 , we have similarly It remains to control the term involving B. We argue as in (4.16) and (4.17).We set Then The latter is controlled by εC K V α for the same reasons as usual.Combining all the pieces together, we obtain finally (4.14).

Analysis of the fixed point mapping H ε
In this section we focus on the fixed point mapping H ε (2.13), which is defined through an infinite series.We are first concerned with the convergence of the series for V ∈ E α 0 .
5.1.Well-posedness of H ε on balls.-Consider the following decomposition of each term of the series (2.13) in two parts, with the corresponding notations: They have the following properties: Lemma 5.1.-For every ball K ⊂ E α 0 , there exists ε K such that for any ε ε K , and The proof of Lemma 5.1 is a straightforward consequence of Proposition 4.1 and the assumptions on m made in definition 1.2, particularly (1.6).
Lemma 5.2.-For every ball K ⊂ E α 0 , there exists ε K such that for any ε ε K , and Proof.-We begin by verifying the condition ∂ z Γ I ε (0) = 0.This is in fact equivalent to the choice of γ ε (V ), as can be seen on the following computation: Now, comparing (3.1) with (4.7), we see that ∂ z Γ I ε (0) = 0 is equivalent to J(γ ε (V ), V ) = 0, provided ε is small enough (for the quantities to be well defined).
Secondly, we need to get uniform bounds on the derivatives of Γ I ε to prove that it belongs to E α .The following formulas relate the successive logarithmic derivatives of We can use directly the weighted estimates in Proposition 4.7, which include the algebraic decay of the first order derivative.Algebraic combinations are compatible with those estimates because A fortiori those terms are all uniformly bounded and so we obtain that The main result of this section is the following one: -For every ball K ⊂ E α 0 , there exists ε K such that for any ε ε K , and V ∈ K, the sum H ε (V ) is finite.
Before proving this statement, we first establish an auxiliary technical lemma about the following summation operator S: Proof.-We perform a Taylor expansion: there exists h k , such that Λ(2 . Therefore, we have immediately One can now proceed to the proof of the finiteness of the sum of H ε in definition 2.6.
Proof of Proposition 5.3.-Let K be the ball of E α 0 of radius K α and take V ∈ K, z ∈ R. To use the previous lemma, we first notice the identity by definition: There are two conditions to verify in order to apply Lemma 5.4: Those properties are verified thanks to Lemmas 5.1 and 5.2.The Proposition 5.3 immediately follows.
So far, we have not used the algebraic decay condition which is part of the definition of E α .In the following lemma, we refine the estimate on S (Λ) ∈ E α .This foreshadows the same result for the function H ε (V ), as stated in the next section.Lemma 5.5 (Better control of the series).-Assume that Λ ∈ E α , that ∂ z Λ(0) = 0, and that (1 + |z|) α ∂ z Λ ∈ L ∞ .Then, S (Λ) belongs to E α , with a uniform estimate: There is some subtlety hidden here.In fact, we were not able to propagate the algebraic decay at first order from Λ to S (Λ).What saves the day is that we gain some algebraic decay of the first order derivatives somewhere in our procedure (see e.g.Proposition 4.7).
Proof.-Recall the notation ϕ α (h) = (1 + |h|) α .We begin with the uniform bound on the first derivative, which is the main reason why we have to impose algebraic decay in our functional spaces.
Step 1: ∂ z S (Λ) is uniformly bounded.-We split the sum in two parts.Let h ∈ R, and let N h ∈ N be the lowest integer such that |h| 2 N h .We consider the two regimes: k > N h and k N h .In the former regime, a simple Taylor expansion yields by definition of N h .In the regime k N h , we use the algebraic decay which is encoded in the space E α .If |h| > 1, we have N h 1, and By definition of N h , we have 2 N h −1 < |h|, so that the right-hand-side above is bounded by a constant that get arbitrarily large as α → 0 (hence, the restriction on α > 0): The case |h| 1 is trivial as the sum is reduced to a single term ∂ z Λ(h) since N h = 0.
Step 2: ϕ α ∂ 2 z S (Λ) is uniformly bounded.-This bound and the next one are easier.For any h ∈ R, we have Since 1 2 −k , one obtains The latter sum is finite since α < 1.
Step 3: -The proof is similar to the previous argument.

Contraction properties (second part).
-In this section we prove that H ε stabilizes some subset of E α 0 .We first show that H ε maps balls into balls with incremental radius that do not depends on the initial ball (Proposition 5.8).This property immediately implies the existence of an invariant subset for H ε (Corollary 5.9).Finally, we prove that the mapping H ε is a contraction mapping for ε small enough (Theorem 5.10).To completely justify the definition of H ε , it remains to show that H ε (V ) ∈ E α 0 .We begin with the lower bound on the second derivative, which is for free.
Lemma 5.6 (Lower bound on ∂ 2 z H ε (V )(0)).-For every ball K ⊂ E α 0 , there exists ε K such that for any ε ε K , and V ∈ K, we have: Proof.-The identity ∂ z H ε (V )(0) = 0, and more particularly ∂ z Γ I ε (0) = 0 is a consequence of the choice of γ e (V ) in Proposition 2.4.Indeed, we have, by (5.4), For the second estimate, a simple computation yields, using m(0) = ∂ z m(0) = 0: But since V ∈ E α 0 , one can use again the uniform estimates of Proposition 4.7 to write that for ε ε K , that depends only on the ball K: where O(ε) that depends only on the ball K.Then, we use Proposition 4.1 with δ = 1/3 to deduce that for ε small enough, we have I ε (V ) 4/3.Then (5.7) can be simplified into Recall that ∂ 2 z m(0) > 0 by assumption.Therefore, for ε small enough, we get as claimed Remark 5.7.-Considering the proof, another way to interpret the result is that automatically for any function V ∈ E α such that ∂ z V (0) = 0, the function H ε prescribes a lower bound on ∂ 2 z V (0).Since we are seeking a fixed point H ε (V ) = V , we may as well put this condition in the subspace E α 0 without loss of generality.Finally, we can establish a first useful estimate on H ε (V ) α , showing more than just its finiteness: Proposition 5.8 (Contraction in the large).-For every ball K ∈ E α 0 , there exists an explicit constant κ(α) < 1, as well as C m , C K and ε K that depend only on K such that, for all ε ε K , and for every V ∈ K, Proof.-Let K be the ball of E α 0 , and take V ∈ K.For clarity we write respectively I ε (h) and Combining various estimates derived in Section 5.1, and particularly Lemma 5.5 together with Lemmas 5.1 and 5.2, we find that H ε (V ) = S (Γ ε ) = S (Γ m ε ) − S (Γ I ε ) belongs to E α .However, the associated estimate (5.5) is not satisfactory, at least for the S (Γ I ε ) and we need to re-examine the dependency of the constants upon ε and α.
We are now in position to state the more important result of this section: Theorem 5.10 (Contraction mapping).There exists a constant C K0 such that for any ε ε 0 , and every function V 1 , V 2 ∈ K 0 , the following estimate holds true Proof.-We denote by V the difference V 1 − V 2 , again.The proof is analogous to Proposition 5.8.For clarity we write respectively We decompose H ε (V ) as above: We deal with H m ε in the following lemma: Lemma 5.11.-There exists a constant C 0 such that for any ε ε 0 , and every function V 1 , V 2 ∈ K 0 , we have Proof.-Recall the following definition: . The first derivative has the following expression, (5.12) Clearly, I 2 ε (0) + m is bounded below, uniformly for ε small enough.Therefore, we can repeat the arguments of Lemma 5.5, with Λ = log(I 1 ε (0) + m) in order to get The next order derivatives can be handled similarly.Indeed, the following quantities must be bounded uniformly by εC 0 V α : The first and the third items are handled similarly as for the first derivative.The three other items are handled analogously.For the sake of concision, we focus on the second line: We have, .
The first factor is uniformly bounded by assumption (1.6), for ε small enough.The second factor is the same as above, so we can conclude directly.
It remains to handle H I ε .We have the following formulas for the two first derivatives (5.1)-(5.3): Finally the formula for the third derivative is: The combination of Proposition 4.6 and Lemma 5.5 yields In the same way, we get the bound for the second derivative, using the factorization ) ε (z) , together with the uniform bound in Proposition 4.7.
Since we assume U ε ∈ E α , we can differentiate the previous equation, and evaluate it at z = 0 to get: As in Section 2, a direct computation shows that γ U and V U are linked by the following relation: In order to invert this relationship, and deduce that γ , the other conditions being clearly verified.
Differentiating the problem (P U ε ) twice, and evaluating at z = 0, we get: Then, using straightforward adaptations of Propositions 4.1 and 4.7, where V should be replaced with V U ∈ E α and γ ε (V ) should be replaced by γ U , we find that for ε sufficiently small.We deduce that the missing condition is in fact a consequence of the formulation (P U ε ): 0), so we have established that V ε ∈ E α 0 .Hence, we can legitimately invert (6.3), so as to find γ U = γ ε (V U ), where the function γ ε is defined in Proposition 2.4.Since U ε ∈ K 0 by assumption, we have in particular V U α R 0 .Of course, V U is the candidate of being the unique fixed point of H ε in K 0 (but also in K 0 ).The proof of this claim follows the lines of section 2.2, checking that all manipulations are justified.
First, we divide (6.2) by According to Proposition 4.1, this quantity is uniformly close to 1 for ε small, so it does not vanish.Taking the logarithm on both sides, we get for all z ∈ R: We differentiate the last equation to end up with the following recursive equation for every One simply deduces, that for all z ∈ R, we necessarily have: Note that the C 1 continuity at z = 0 is used here.Moreover, ∂ z V U (0) = 0 by definition of V U .The analysis performed in Proposition 5.3 guarantees that this sum is indeed finite.Finally, integrating back the previous identity yields The last expression is nothing but H ε (V U ), by definition (2.13).Therefore, As previously, we decompose U ε = γ ε • +V ε , where γ ε stands for γ ε (V ε ).Firstly, we have λ ε = I ε (V ε )(0) → 1, using Proposition 4.1.Secondly, using an argument of diagonal extraction, there exists a subsequence ε n , and a limit function V 0 such that We have used the Arzelà-Ascoli theorem and the uniform C 3 bound in order to get the convergence up to the second derivative.However, there is no reason why the convergence should hold for the third derivative, due to the lack of compactness.
Looking at (P U ε ), we see that I ε (U ε ) converges uniformly to 1, and, for every given z ∈ R, (6.4) implies that Passing to the pointwise limit in problem (P U ε ), we get that V 0 solves the following problem: Then, we have necessarily: This completes the proof of Theorem 1.4(ii), up to the identification of the limit of γ ε , if it exists.In our approach, this goes through the characterization of the functional J ε (2.9).This was indeed the purpose of Lemma 3.1.Here comes an important difficulty, as compactness estimates are not sufficient to pass to the limit in J ε (0, V ε ) as ε → 0 (3.3), as it would formally involve the pointwise value ∂ 3 z V 0 (0) which is beyond what our compactness estimates can provide.Note that passing to the limit in ∂ g J ε (g, V ε ) as ε → 0 is not an issue, as it can be encompassed by (6.5), see (3.4).
The remaining term, W (3) ε (z) is more delicate to handle.In fact, it will result in a contraction estimate, exactly as in section 5. We recall the expression of W (3) ε (4.8): As in the proof of equations (4.12) to (4.14), we get that the last two contributions involving the non-linear and lower order terms Let δ > 0. We can choose R 1 (δ) large enough so that, for all R R 1 (δ), we have In the region where y 1 and y 2 are both below R/2, we introduce the difference with ∂ 3 z m, as in (6.9): where By construction, we have |z + εy i | εR/2 + εR/2 εR.Therefore, we have As for B we find: J.É.P. -M., 2019, tome 6

Extension to higher dimensions
Our methodology can be extended to higher dimension, without too much effort.This section is devoted to the generalization of the elements of proof that were specific to the one-dimensional case.
All the estimates on the operator H ε and its constitutive pieces are still operational in higher dimension.The only part of our proof that requires some specific attention is the construction of the linear part γ ε (V ε ) which was performed in Section 3. Indeed, we used a monotonicity argument to show that γ ε (V ε ) can be defined in a unique way.
We proceed as in Section 3. First we show formally how to obtain the expression of the vector γ 0 (1.11) via suitable Taylor expansions.Then, we justify these Taylor expansions, and we exhibit a monotonic function that enables to conclude, exactly as in dimension 1.
The quadratic form Q yields the multivariate centered Gaussian distribution associated with the following covariance matrix Σ ∈ M 2d (R): 3 Id − Id − Id 3 Id .
The Kronecker product y 1 ⊗ y 1 yields a matrix of moments, and so the relation (7.1) can be simplified, similarly to the one dimensional case, so as to obtain: The right hand side is a tensor applied to a matrix yields a vector that can be simplified even further using tensorial properties: D 3 V (0) Id = D(∆V )(0).Then, provided that D 2 V (0) is non degenerate, we obtain the limited expected value of γ 0 in dimension higher than 1, that is a generalization of (3.2): In the case where V 0 is given by (1.11) through the fixed point procedure, we obtain γ 0 (V 0 ) = 1 2 D 2 m(0) −1 D(∆m)(0).
J.É.P. -M., 2019, tome 6 7.2.Extension of the proof of Proposition 2.4 (section 3.2).-We now fix V ∈ K, where K is a ball of E α 0 .The purpose is to prove that there is a unique solution in R d of the following problem: (7.2) J ε (γ, V ) = 0.
We insist upon the fact that the variable g belongs to R d and the function J ε (•, V ) is now defined as a vector field on R d , J ε : R d × E α → R d .As in section 3.2, we can obtain the following estimate by means of refined Taylor expansions, where J ε (0, V ) is bounded a priori, independently upon ε > 0 for V ∈ K. To prove the existence of a root γ ε , we used the mean value theorem in the proof of Proposition 2.4.The analogous statement in higher dimension is the Brouwer fixed point theorem.Indeed, (7.2) can be recast as follows: Thus, we are led to finding a fixed point of a continuous function.As in the onedimensional case, thanks to the lower bounded D 2 V (0) µ 0 Id encoded in the definition of E α 0 (1.7), we can show easily that there exists R K such that the ball of radius R K in R d is left invariant by T .Brouwer's fixed point theorem guarantees that there exists a fixed point γ ε to T , which is also a root of (7.2).
For the uniqueness part, we can use strict monotonicity, similarly as in the one dimensional case.This is possible, thanks to (3.4): (7.4) D g J ε (g, V ) = 1 2 D 2 V (0) + O(ε).
We deduce from this strong estimate that the vector field J ε (•, V ) is locally uniformly monotonic, in the sense that there exists µ K such that the following inequality holds true for all ε sufficient small, and every g 1 , g 2 ∈ B(0, R K ): This monotonicity condition is clearly satisfied, as it is equivalent to the following first order condition, (7.6) It is immediate that any strictly monotonic vector field admits at most one root.This completes the proof of uniqueness of γ ε (V ).