WELL-POSEDNESS OF SOME NON-LINEAR STABLE DRIVEN SDES

NOUFEL FRIKHA, VALENTIN KONAKOV, AND STÉPHANE MENOZZI

Abstract. We prove the well-posedness of some non-linear stochastic differential equations in the sense of McKean-
Vlasov driven by non-degenerate symmetric α-stable Lévy processes with values in Rd under some mild Hölder
regularity assumptions on the drift and diffusion coefficients with respect to both space and measure variables. The
methodology developed here allows to consider unbounded drift terms even in the so-called super-critical case, i.e.
arXiv:1910.05945v1 [math.AP] 14 Oct 2019

when the stability index α ∈ (0, 1). New strong well-posedness results are also derived from the previous analysis.

1. Introduction
In this article, we are interested in some non-linear stochastic differential equations (SDEs for short) with
dynamics
Z t Z t
(1.1) Xt = ξ + b(s, Xs , [Xs ])ds + σ(s, Xs− , [Xs ])dZs , [ξ] = µ ∈ P(Rd )
0 0

driven by a d-dimensional α-stable process Z, α ∈ (0, 2), with coefficients b : R+ × Rd × P(Rd ) → Rd and
σ : R+ × Rd × P(Rd ) → Rd ⊗ Rd . Here and throughout the article, we will denote by P(Rd ) the space of probability
measures on Rd and by [θ] the law of a random variable θ.
This type of dynamics commonly referred to in the literature as McKean-Vlasov, distribution dependent or
mean-field SDEs naturally appears as the limit equation of an individual particle evolving within a large system
of particles interacting through its empirical measure, as the size of the population grows to infinity, following the
well-known propagation of chaos phenomenon. We refer to the original works by Kac [Kac56] in kinetic theory, by
McKean [McK66], [McK67] in non-linear parabolic partial differential equations and by Sznitman [Szn91] in the
diffusive case α = 2.
The well-posedness theory in the weak or strong sense of such dynamics has been an active research topic during
the last decades, see e.g. Funaki [Fun84], Oelschlaeger [Oel84], Gärtner [Gär88], [Szn91], Jourdain [Jou97] and
more recently, Li and Min [LM16], Chaudru de Raynal [CdR19], Mishura and Veretennikov [MV18], Lacker [Lac18],
Chaudru de Raynal and Frikha [CdRF18b], Röckner and Zhang [RZ18] for a short sample in the diffusive setting.
Let us also mention the book of Kolokolstov [Kol10] for some related works on kinetic equations. In the current
strictly α-stable context α ∈ (0, 2), we mention the work of Jourdain et al. [JMW08] with applications to the
fractional porous media equation on the real-line and of Jourdain et al. [JMW05b], [JMW05a] for a probabilistic
approach to some non-linear equations involving the fractional Laplacian operator. Let us also mention the work
of Graham [Gra92] who derives existence and uniqueness results for non-linear diffusions with Lipschitz coefficients
and bounded jump rates.
The classical well-posedness theory of such dynamics is now well-understood in the standard Cauchy-Lipschitz
framework on the space Rd × Pp (Rd ), p ≥ 1, Pp (Rd ) being the space of probability measures with finite p moments
equipped with the Wasserstein distance of order p.
Trying to go beyond the previous setting by weakening the regularity assumptions on the coefficients with
respect to both space and measure variables remains an interesting and challenging question that has attracted
the attention of the research community. In the diffusive setting α = 2, let us mention e.g. the work [Sch87]
by Scheutzow where Ra counterexample to uniqueness for the SDE (1.1) is exhibited. Namely, when σ ≡ 0 and
b(t, x, m) = b(m) = R b̄(y)m(dy) for some bounded and locally Lipschitz function b̄, the non-linear SDE (1.1)
with random initial condition has several solutions. However, when σ ≡ 1, the noise helps to restore existence
Rand uniqueness. Indeed, the well-posedness of the corresponding martingale problem when b(t, x, m) = b(x, m) =
Rd
b̄(x, y) m(dy) has been established by Shiga and Tanaka [ST85]. Such a result has been extended by Jourdain
[Jou97] where uniqueness is proved under the more general assumption that b is bounded measurable and Lipschitz
with respect to the total variation metric on the space of probability measures, the diffusion coefficient being

Date: October 15, 2019.

2000 Mathematics Subject Classification. Primary 60H10, 60G46; Secondary 60H30, 35K65.
Key words and phrases. McKean-Vlasov stable SDEs; non-linear martingale problem; weak uniqueness; strong uniqueness.
1
2 NOUFEL FRIKHA, VALENTIN KONAKOV, AND STÉPHANE MENOZZI

Lipschitz in space and independent of the measure argument. Still in the non degenerate diffusive framework, some
extensions have been recently obtained in [MV18] who established some weak and strong well-posedness results for
a diffusion coefficient depending only on time and space variables, in [Lac18] and also in [RZ18] for dynamics with
a singular convolution type interaction kernel in the drift coefficient. We finally mention the recent contribution
[CdRF18b] where well-posedness results in the weak and strong sense are established through a fixed point approach
on a suitable complete metric space of probability measures under mild Hölder type regularity assumptions on the
coefficients with respect to both space and measure arguments. In particular, weak well-posedness is established
when the diffusion coefficient is not Lipschitz with respect to the underlying Wasserstein distance. Some new
propagation of chaos results for the approximation of the dynamics (1.1) by a system of particles have recently
been obtained by the same authors in their companion paper [CdRF18a]. In [JMW08], the authors extend the
classical strong well-posedness results of [Szn91] for non-linear SDEs when both initial condition and Lévy driving
process are square integrable. When the Lévy process is not square integrable, as it is the case in our current
α-stable framework, existence without uniqueness is proved when both coefficients are Lipschitz with respect to
some bounded modification of the Wasserstein metric of order 1.
We here revisit the problem of the unique solvability of the SDE (1.1) by tackling the corresponding formulation
of the non-linear martingale problem under some rather mild assumptions on the coefficients in both the spatial and
measure arguments. Namely, we will consider Hölder continuous in space coefficients with possibly unbounded drift
term b and a Lipschitz regularity condition w.r.t. a suitable Hölder type metric (which actually heavily depends
on the spatial Hölder exponents of the coefficients) for the measure arguments. We will also assume some usual
non-degeneracy and boundedness conditions on the Lévy measure ν of the driving symmetric stable process Z and
the diffusion coefficient σ appearing in (1.1).
In the current setting, we thus aim at finding a unique probability measure P on D(R+ , Rd ) (Skorokhod space of
càdlàg functions) such that denoting by (P(t))t≥0 its time marginals, (yt )t≥0 the associated canonical process and
by (Ft )t≥0 its filtration, P(0) = µ and for any ϕ ∈ C01,2 (R+ × Rd , R) the process
Z t


(1.2) ϕ(t, yt ) − ϕ(0, y0 ) − ∂r + ArP ϕ(r, yr )dr
0

is an ((Ft )t≥0 , P)-martingale starting from 0 at time 0. In the above formulation, ArP is the following integro-
differential operator
(1.3) Z


ArP ϕ(r, x) = hb(r, x, P(r)), Dx ϕ(r, x)i+ ϕ(r, x+σ(r, x, P(r))z)−ϕ(r, x)−h∇ϕ(r, x), σ(r, x, P(r))zi1|z|≤1 ν(dz)
Rd \{0}

acting on smooth test functions and in (1.3) ν is a symmetric stable Lévy measure on Rd , i.e. for α ∈ (0, 2), writing
in polar coordinates ζ = rξ, (r, ξ) ∈ R∗+ × Sd−1 :
dr
(1.4) ν(dζ) = ω(dξ),
r1+α
where ω is a symmetric measure on the sphere Sd−1 satisfying some non-degeneracy assumptions.
Our main idea consists in applying a fixed point argument to a family of proxy linearized martingale problems
on a suitable complete metric space of probability measures. In the current strictly stable setting α ∈ (0, 2), for
the well-posedness in the weak and strong sense of the corresponding linear martingale problem, i.e. when the
dependence with respect to the measure argument is frozen, we can mention among others Tanaka et al. [TTW74],
Bass [Bas03], Bass et al. [BBC04], Mikulevicius and Pragarauskas [MP14], Priola [Pri18] or Zhang et al. [HWZ19]
for the specific (possibly degenerate) stable case. In our previously mentionned Hölder setting, the main idea to
complete our fixed point approach, consists in exploiting some parabolic Schauder estimates but under some rough
regularity conditions on the final condition, whose regularity will be somehow related to the spatial one of the
coefficients. Such estimates are obtained through a forward parametrix perturbation argument in the same spirit
as those established by Chaudru de Raynal et al [CdRHM18], [CdRMP19] in the Kolmogorov degenerate diffusive
setting or in the α-stable supercritical case.
As a by-product of the usual strong uniqueness results on linear stable driven SDEs, the strong well-posedness
of the corresponding non-linear SDE is also derived under suitable regularity assumptions on the coefficients and
a non-degeneracy assumption on the Lévy measure. Again, such results will be obtained for a bounded non-
degenerate σ, with b, σ being Hölder continuous in space and for potentially unbounded in space drift terms b.
Moreover, concerning the measure argument, the maps P(Rd ) ∋ m 7→ b(t, x, m), σ(t, x, m) are also assumed to have
a bounded linear functional (or flat) derivative denoted by [δb/δm](t, x, m)(y) and [δσ/δm](t, x, m)(y) (see Section
2 for a precise definition) which is Hölder-continuous w.r.t. both variables x and y.
 WELL-POSEDNESS OF SOME NON-LINEAR STABLE DRIVEN SDES 3

Compared to the aforementioned references in the strictly α-stable context, our methodology allows to consider-
ably weaken the regularity assumptions on the coefficients b and σ with respect to both space and measure variables.
It also covers a large class of interaction in the measure argument with Hölder type regularity. In particular, no
Lipschitz regularity with respect to Wasserstein distance is needed which, to the best of our knowledge, appears
to be new. Let us indicate as well that, even in the stable Gaussian case, i.e. α = 2 for which Z = B (standard
Brownian motion), our approach applies and would provide well-posedness for non-linear SDEs with unbounded
drift terms and full dependence on the measure in the diffusion coefficient (see Remark 2.4 below).
The article is organized as follows. The basic definitions together with the assumptions, the main results and the
strategy of proof are described in Section 2. The sensitivity analysis w.r.t the measure argument of the semigroups
generated by the linearized α-stable driven SDE is carried out in Section 3. The proof of some useful technical
results are given in Appendix.

2. Overview: definitions, assumptions and main results

In order to tackle the non-linear martingale problem as introduced in (1.2), we will proceed through a fixed
point procedure on a suitable complete metric space of probability measure valued flows. Let us now describe this
functional space.

The metric space of measure valued flows. We first endow P(Rd ) with the following metric. Fix β ∈ (0, 1]
and define for µ, ν ∈ P(Rd ),
Z
(2.1) dβ (µ, ν) := sup f (z)(µ − ν)(dz),
f :kf kCβ ≤1

where kf kC β := |f |∞ + sup(x,y)∈(Rd)2 , x6=y |f (x)−f

|x−y|β
(y)|
denotes the usual Hölder norm on C β (Rd , R), where | · | stands
for the usual Euclidean norm of Rd (see e.g. Krylov [Kry96]). For β = 1, dβ is also known as the Fortet-Mourier
distance. Let us introduce now the cost function cβ (x, y) = |x − y|β ∧ 1, (x, y) ∈ Rd . Denote by Π(µ, ν) the set of
all transport plans from µ to ν w.r.t cβ . Then, for any π ∈ Π(µ, ν) and any f satisfying kf kC β ≤ 1, it is readily
seen Z Z Z
| f (z)(µ − ν)(dz)| = | (f (x) − f (y))π(dx, dy)| ≤ 2 cβ (x, y)π(dx, dy)

so that optimizing w.r.t π and f

Z
fβ (µ, ν) := 2 inf
dβ (µ, ν) ≤ 2W cβ (x, y)π(dx, dy).
Π(µ,ν)

Let us note that (P(Rd ), dβ ) is a complete metric space1 and that dβ metrizes the weak convergence of probability
measures.
For fixed 0 ≤ s < t < ∞, we then introduce the set of continuous probability measure valued flows C([s, t], P(Rd )).
We will from now on denote the elements of C([s, t], P(Rd )) with a bold capital letter and equip it with the following
uniform metric. For all P, P′ ∈ C([s, t], P(Rd )),
(2.2) dβ,s,t (P, P′ ) := sup {dβ (P(r), P′ (r))} .
r∈[s,t]

Since (P(Rd ), dβ ) is a complete metric space, it follows that C([s, t], P(Rd )) equipped with dβ,s,t is also complete.
We refer to Villani [Vil09] for related issues concerning the aforementioned points. There are many topologies
with which we can equip the space C([s, t], P(Rd )). The above choice of metric strongly derives from the fact that
a Hölder-type regularity assumption on both coefficients b and σ appears to be sufficient to establish the well-
posedness in the weak sense of standard linear stable driven SDEs (see e.g. [MP14]). Let us also mention that,
beyond our Hölder setting, the choice of the metric is very much related to the spatial smoothness of the coefficients.
In particular, when little or no regularity is available in space for the coefficients (bounded or Lq − Lp , see e.g.
[MV18] or [RZ18]) the natural metric for the measure appears to be the total variation one.

1If (µ ) d
k k∈N is a Cauchy sequence in P(R ) equipped with dβ , then a slight modification of the proof of Lemma 6.12 in [Vil09]
allows to conclude that it is tight so that it admits a subsequence (still denoted by (µk )k∈N ) which weakly converges to some measure
µ. For each k, introduce an optimal transport plan πk between µk and µ (see e.g. Theorem 4.1 in [Vil09] for the existence of such
optimal coupling) for the induced related cost function cβ . By Lemma 4.4 in [Vil09], (πk )k∈N is tight in P(Rd × Rd ). Hence, up to an
extraction, we may assume that πk → π weakly as k → ∞. Since each πk is optimal, Theorem R 5.19 of [Vil09] guarantees that π is an
optimal (trivial) coupling between µ and µ so that lim supk→∞ dβ (µk , µ) ≤ 2 lim supk→∞ (Rd )2 cβ (x, y)πk (dx, dy) = 0. Since (µk )k∈N
is a Cauchy sequence with a converging subsequence (w.r.t the metric dβ ), it eventually converges to µ.
4 NOUFEL FRIKHA, VALENTIN KONAKOV, AND STÉPHANE MENOZZI

Now, for a given µ ∈ P(Rd ), we introduce the subset As,t,µ of C([s, t], P(Rd )) defined by

(2.3) As,t,µ := P ∈ C([s, t], P(Rd )) : P(s) = µ .

The subspace As,t,µ equipped with the metric defined in (2.2) can also be viewed as a complete metric space.

In order to rigorously formulate our regularity assumption on the coefficients w.r.t the measure argument, we
introduce the notion of flat derivative of a continuous map U : P(Rd ) → R, P(Rd ) being equipped with the distance
dβ for some β ∈ (0, 1]. This notion will play a key role in our analysis. We refer to Chapter 5 of the monograph
by Carmona and Delarue [CD18] for a more detailed discussion on the notion of differentiability of functions of
probability measures.
Definition 2.1 (Flat derivative of U ). The continuous map U : P(Rd ) → R is said to have a continuous linear
functional derivative if there exists a continuous function δU/δm : P(Rd ) × Rd → R such that y 7→ [δU/δm](m)(y)
is bounded, uniformly in m for m ∈ K, K being any compact subset of P(Rd ) and such that for any m, m′ ∈ P(Rd ),
Z
U ((1 − ε)m + εm′ ) − U (m) δU
lim = (m)(y) (m′ − m)(dy).
ε↓0 ε R d δm
The mapR y 7→ [δU/δm](m)(y) being defined up to an additive constant, we will follow the usual normalization
convention Rd [δU/δm](m)(y) m(dy) = 0.
Let us illustrate the structure of this notion of differentiation with a couple of commonly encountered examples.
R
Example 2.2. (1) If the function U is of scalar form, namely, U (m) = Rd h(y)m(dy) for some real-valued
bounded and continuous function h defined on Rd then it is readily seen that
δU
(y) = h(y).
δm
(2) If the function U isR of quadratic type w.r.t
R the measure, namely, denoting by ⋆ the usual convolution
operator, U (m) = Rd [h ⋆ m](x)m(dx) = (Rd )2 h(x − y) m(dx)m(dy) for some bounded and continuous
function h defined on Rd then one has
Z Z
δU
(y) = h(x − y) m(dx) + h(y − x) m(dx).
δm Rd Rd

Remark that if U has a linear functional derivative then for any m, m′ ∈ P(Rd ), it holds
Z 1Z
′ δU
(2.4) U (m) − U (m ) = (λm + (1 − λ)m′ )(y) (m − m′ )(dy) dλ.
0 Rd δm

The family of inhomogeneous integro-differential operators. In order to solve locally in time the non-
linear martingale problem (1.2), we will perform a fixed point argument in the metric space As,t,µ . We thoroughly
need to consider some inhomogeneous linearized integro-differential operators associated with a given measure flow
µ ∈ C([s, t], P(Rd )). Namely, for (r, x) ∈ [s, t]×Rd and all ϕ ∈ C01,2 ([s, t]×Rd , R), we abuse the definition introduced
in (1.3)2 and write

 
(2.5) Arµ ϕ(r, x) ≡ A (r, ·, µ; Dx ) ϕ(r, x) = Lµ,α
r ϕ(r, x) + hb(r, x, µ(r)), Dx ϕ(r, x)i,

where
Z


(2.6) Lµ,α
r ϕ(r, x) = p.v. ϕ(x + σ(r, x, µ(r))ζ) − ϕ(x) ν(dζ),
Rd

and ν is the Lévy measure given by (1.4).

We now introduce some non-degeneracy and regularity assumptions on the Lévy measure and coefficients.

2Observe indeed that P in (1.3) is a probability measure on D(R+ , R) whereas µ ∈ C([s, t], P(Rd )) is a flow of Rd -valued probability
measures.
 WELL-POSEDNESS OF SOME NON-LINEAR STABLE DRIVEN SDES 5

Non-degeneracy assumptions.
(ND) There exists κ ≥ 1 s.t. for all z ∈ Rd :
Z
−1 α
(2.7) κ |z| ≤ |hz, ξi|α ω(dξ) ≤ κ|z|α .
Sd−1

Note that a wide family of spectral measures satisfy condition (2.7), from absolutely continuous ones to (very)
singular ones. For instance, taking ω = ΛSd−1 (Lebesgue measure on the sphere) readily yields the previous control.
Pd
On the other hand, any measure ω = i=1 ci (δei + δ−ei ) with positive coefficients (ci )i∈[[1,d]] , where the (ei )i∈[[1,d]]
stand for the canonical vectors, also satisfies (2.7).
In order to solve the non-linear martingale problem, we will also assume that:
(AC) The jump measure writes: ν(dy) = f (y)dy for a continuously differentiable function f : Rd → R with bounded
and Lipschitz first order derivative. Since ν is a symmetric stable Lévy measure (see (1.4)), we also have that:
y
g( |y| )
(2.8) f (y) = ,
|y|d+α
where g is an even and continuously differentiable spherical function with a bounded and Lipschitz first order
derivative.
We also assume that the diffusion coefficient σ in (2.6) satisfies the following uniform ellipticity condition.
(UE) There exists Λ ≥ 1 s.t. for all (t, z, µ) ∈ R+ × Rd × P(Rd ) and all ξ ∈ Rd :
(2.9) Λ−1 |ξ|2 ≤ h(σ ∗ σ)(t, z, µ)ξ, ξi ≤ Λ|ξ|2 .

Regularity assumptions on the coefficients. We will also need some regularity assumptions on the coefficients σ and
b w.r.t both the spatial and measure arguments.
We assume that the diffusion coefficient σ satisfies the following conditions:
(DH ) For any (t, µ) ∈ R+ × P(Rd ), the mapping x 7→ σ(t, x, µ) is bounded and 2η-Hölder continuous, for some
η ∈ (0, 1/2], uniformly in (t, µ) ∈ R+ × P(Rd ). Namely, there exists C ≥ 1 s.t., for all (t, µ) ∈ R+ × P(Rd ):
(2.10) kσ(t, ·, µ)kC 2η ≤ C,
where C 2η (Rd , Rm ) := {ϕ ∈ C(Rd , Rm ) : kϕk∞ + supx6=y |ϕ(x)−ϕ(y)|
|x−y|2η < +∞} with m ∈ {1, d × d} denotes the usual
Hölder space. When there is no ambiguity, as in (2.10), we simply denote C 2η := C 2η (Rd , Rm ) to ease the reading.
For any (t, x) ∈ R+ × Rd , the continuous map P(Rd ) ∋ m 7→ σ(t, x, m) admits a bounded and continuous flat
derivative denoted by the map [δσ/δm](t, x, m)(.) such that (x, y) 7→ [δσ/δm](t, x, m)(y) is 2η-Hölder continuous,
uniformly w.r.t the variables t and m.
We suppose that the drift coefficient b fulfills the following conditions:
(BH ) For any (t, µ) ∈ R+ × P(Rd ), the mapping x 7→ b(t, x, µ) belongs to the homogeneous Hölder space
|ϕ(x) − ϕ(y)| 1
C˙2η (Rd , Rd ) := {ϕ ∈ C(Rd , Rd ) : sup 2η
< +∞}, η ∈ (0, ]
x6=y |x − y| 2
uniformly w.r.t. the variables (t, µ). In particular, the drift b may be unbounded in space.
For any (t, x) ∈ R+ × Rd , the continuous map P(Rd ) ∋ m 7→ b(t, x, m) admits a bounded and continuous flat
derivative denoted by the map [δb/δm](t, x, m)(.) such that (x, y) 7→ [δb/δm](t, x, m)(y) is 2η-Hölder continuous,
uniformly w.r.t the variables t and m.
We eventually assume that the following condition, linking the spatial Hölder exponent 2η in assumptions (DH )
and (BH ) and the stability index α is fulfilled:
(L) It holds that:
2η + α > 1.
We importantly point out that the former condition appearing in (BH ) is the natural constraint arising in
[CdRMP19] to derive Schauder estimates for a general, and potentially singular, non-degenerate spherical measure
ω in the sense of (2.7) in the super-critical case.
We will say that assumption (AS ) is satisfied if assumptions (ND), (AC), (UE), (DH ), (BH ) and (L) hold.

Before going further, we first derive an important control that will play a key role in our analysis of the well-
posedness of the non-linear martingale problem related to the SDE (1.1). Under the regularity assumptions (DH )
6 NOUFEL FRIKHA, VALENTIN KONAKOV, AND STÉPHANE MENOZZI

and (BH
), denote by U one of the coefficients σ or b and introduce the map x 7→ δU t, x, µ, ν := U t, x, µ −
U t, x, ν , for any fixed µ, ν ∈ P(Rd ). From (2.4) and the very definition (2.1) of d2η , one readily gets

δ δ
(2.11) ∀β ∈ [0, 1], |δU t, ., µ, ν |∞ ≤ sup k U (t, ., m)kC 2η d2η (µ, ν) ≤ sup k U (t, ., m)kC 2η d(1−β)2η (µ, ν).
t,m δm t,m δm

Similarly,
Z 1 Z h δ i

′

δ
δU t, x, µ, ν − δU t, x , µ, ν = U (t, x, µλ ) − U (t, x′ , µλ ) (y)(µ − ν)(dy) dλ
0 Rd δm δm
δ
where we introduced the notation µλ := λµ + (1 − λ)ν for λ ∈ [0, 1]. Using the fact that (x, y) 7→ δm U (t, x, m)(y)
is 2η-Hölder uniformly w.r.t. the variables t and m, for any β ∈ [0, 1], we obtain
h δ δ i h δ δ i
U (t, x, µλ )− U (t, x′ , µλ ) (y) − U (t, x, µλ ) − U (t, x′ , µλ ) (y ′ )
δm δm δm δm
δ
≤ 2 sup k U (t, ., m)(.)kC 2η |x − x | ∧ |y − y ′ |2η
′ 2η
t,m δm
δ
≤ 2 sup k U (t, ., m)(.)kC 2η |x − x′ |β2η |y − y ′ |(1−β)2η .
t,m δm
δ δ
Now, from the boundedness of δm U and the uniform 2η-Hölder regularity of x 7→ δm U (t, x, µ), we similarly get
 δ δ  δ
| U (t, x, µλ ) − U (t, x′ , µλ ) (.)|∞ ≤ 2 sup k U (t, ., m)(.)kC 2η |x − x′ |β2η
δm δm t,m δm

for any β ∈ [0, 1]. From the preceding estimates, one thus deduces

δ
(2.12) ∀β ∈ [0, 1], kδU t, ., µ, ν kC β2η ≤ 2 sup k U (t, ., m)(.)kC 2η d(1−β)2η (µ, ν).
t,m δm

Under (AS ), for any fixed measure flow Q in C([0, ∞), P(Rd )), with time marginal Q(t), and any initial conditions
(s, µ) ∈ [0, ∞) × P(Rd ), there exists a unique solution to the
linearized martingale problem, that is, the martingale
problem associated with the operator AtQ ≡ A (t, ·, Q; Dx ) t≥s introduced in (2.5) is well-posed. We can refer to
[CdRMP19] which gives the corresponding Schauder estimates which in turn allows to derive the aforementioned
well-posedness following the procedure described in e.g. Mikulevicus and Pragarauskas [MP14] or Priola [Pri12].
Moreover, denoting by PQ s,x the unique solution to the martingale problem starting from δx at time s, we have
that the associated canonical process is strong Markov. The corresponding time-inhomogeneous Markov semigroup
Q
will be denoted by (Ps,t )t≥s .

Theorem 2.3 (Well-posedness of the non-linear martingale problem). Under assumption (AS ), for any
s ≥ 0 and any initial distribution µ ∈ P(Rd ), the non-linear martingale problem starting from µ at time s is
well-posed for any η ∈ (0, 1/2] if α ≥ 1 and under the condition α > 2η ∨ (1 − η) if α < 1.
Observe that, in the super-critical case α < 1 the lowest possible stable index α we can attain is strictly greater
than 2/3.
Corollary 2.1 (Strong uniqueness). Under (AS ), if 2η ≥ 1− α2 , then equation (1.1) admits a unique strong solution
for any α ≥ 1 and when α < 1 under the additional condition α > 2η ∨ (1 − η).
Proof of Theorem 2.3. Let T > s be fixed and such that T − s is small. As already mentioned, the central idea is
to use the Banach fixed point theorem on the complete metric space As,T,µ equipped with d.
We consider the map T : As,T,µ → As,T,µ which to the measure flow Q ∈ As,T,µ associates the measure flow
T (Q) ∈ As,T,µ induced by the probability measure

PQs given by the unique solution to the linear martingale problem
associated with the operator A (t, ·, Q; Dx given by (2.5) starting from the initial distribution µ at time s. Namely,


one has T (Q) (t) = [XtQ ], for any t ∈ [s, T ], where (XtQ )t∈[s,T ] is the unique weak solution to the linear SDE with
dynamics
Z t Z t
Q
Xt = ξ + Q
b(r, Xr , Q(r))dr + σ(r, XrQ− , Q(r))dZs , [ξ] = µ.
s s
In particular, one has T (Q)(s) = µ and
Z Z
Q
(2.13) h(y)T (Q)(t)(dy) = Ps,t h(y)µ(dy),
 WELL-POSEDNESS OF SOME NON-LINEAR STABLE DRIVEN SDES 7

Q

recalling that (Ps,t )t≥s denotes the strong Markov semigroup associated with the operator A (t, ·, Q; Dx ) t≥s .
Let now Pi , i = 1, 2, be two measure flows on As,T,µ . Write from (2.13) and (2.2):
Z



dη,s,T T(P1 ), T(P2 ) = sup sup h(y) T(P1 )(t) − T(P2 )(t) (dy)
t∈[s,T ] h:khkCη ≤1
Z  
P1 P2
(2.14) = sup sup Ps,t − Ps,t h(y)µ(dy).
t∈[s,T ] h:khkCη ≤1
Pi i
Let us now write for notational convenience Ps,t ≡ Ps,t , t ∈ [s, T ] for i = 1, 2. The following Proposition, whose
proof is postponed to Section 3 is the key to prove that T generates a contraction on As,T,µ in small time.
Proposition 2.1 (Sensitivity Analysis of the semi-groups w.r.t. the measure argument). Under (AS ) and the
additional assumption α ≥ 2η∨(1− η2 ) if α < 1, for any fixed T > 0, there exists a constant C := C(T, (AS ), α, η) ≥ 1
(T 7→ C(T, (AS ), α, η) being non-decreasing) and an exponent ζ̃ > 0 s.t. for all h ∈ C η (Rd , R), khkC η ≤ 1 and for
any (t, x) ∈ [s, T ] × Rd :
1 2
(2.15) |(Ps,t − Ps,t )h(x)| ≤ C(t − s)ζ̃ dη,s,t (P1 , P2 ).
From (2.15) and (2.14), recalling as well from (2.2) that [s, T ] ∈ t 7→ dη,s,t (P1 , P2 ) is a non-decreasing function of
t, we eventually derive:
dη,s,T (T(P1 ), T(P2 )) ≤ C(T − s)ζ̃ dη,s,T (P1 , P2 ).
Hence, we see that for T sufficiently close to s, so that C(T − s)ζ̃ < 1, the map T is a contraction on the complete
metric space As,T,µ . It thus admits a unique fixed point that we denote by P0 . For n ≥ 1, if Pn−1 is constructed,
denote by Pn the unique fixed-point of T on AnT,(n+1)T,Pn−1 (nT ) . Now set P(t) := Pn (t) if t ∈ [nT, (n + 1)T ]
and denote by PP s the unique solution to the linear martingale problem starting from the distribution µ at time s
in which the measure dependence is given by P. By the characterization of the martingale problem, it is readily
−1
seen that for any integer n and any t ∈ [nT, (n + 1)T ], one has (PP s ) ◦ ytP = Pn (t) = P(t). By the uniqueness of
the fixed point, it turns out that Ps := PP s is also the unique solution to the non-linear martingale problem. This
completes the proof.

At this step, we importantly emphasize that Proposition 2.1 allows to conclude that Theorem 2.3 is valid under
the additional stronger condition α > 2η ∨ (1 − η2 ) which appears to be quite stringent in the supercritical case
α < 1. Indeed, in this context, the minimal attainable stability index α is strictly greater than 4/5. In order to
establish the well-posedness of the non-linear martingale problem under the sole condition α > 2η ∨ (1 − η), one has
to suitably modify the above argument by performing the fixed point argument using the map T2 = T ◦ T still on
the complete metric space As,T,µ equipped with the distance dβ,s,T for some well-chosen β ∈ (0, 2η). For the sake
of clarity, the main modifications of the proof of Proposition 2.1 needed to establish that the iterated procedure
provides a contraction are presented in Appendix B.
Proof of Corollary 2.1. Theorem 2.3 provides weak uniqueness for the McKean-Vlasov SDE (1.1). The dependence
in the law can then classically be viewed as an inhomogeneous time dependence (see e.g. [CdRF18b]). Namely,
(1.1) can be rewritten as:
Z t Z t
Xt = ξ + eb(s, Xs )ds + σ
e (s, Xs− )dZs ,
0 0
where eb(s, Xs ) = b(s, Xs , [Xs ]), σ
e(s, Xs ) = σ(s, Xs , [Xs ]). To derive strong well-posedness, we can therefore now
rely on the classical inhomogeneous linear setting.
If the coefficients are bounded, the result then readily follows from Chen et al. [CZZ17] who proved the complete
strong well-posedness for stable driven SDEs, including the super-critical case up to α ∈ (0, 2). It can also be easily
checked that the procedure developed in the homogeneous case by [Pri12] for α ≥ 1 can readily be extended to the
current time dependent framework. The methodology therein indeed only depends on a priori smoothness controls
that have been established under the current assumptions in [CdRMP19]. Hence, the methodology of [Pri12] to
derive strong uniqueness still applies and extends to the current framework with unbounded coefficients and some
super-critical cases. 
8 NOUFEL FRIKHA, VALENTIN KONAKOV, AND STÉPHANE MENOZZI

Remark 2.4 (On possible extensions). Before going to the proofs of our main results, let us discuss some possible
extensions of our main existence and uniqueness results.
• Let us first indicate that the global spatial Hölder continuity of the drift assumed in (BH ) could be weakened
to a local Hölder continuity condition. Namely, we could only impose that there exists K0 s.t. uniformly in
(t, µ) ∈ R+ × P(Rd ),
|b(t, x, µ) − b(t, y, µ)| ≤ K0 |x − y|2η , |x − y| ≤ 1.
This is precisely the condition appearing in [CdRMP19]. The drawback of considering the above condition
is that it would induce a localization of our arguments and therefore additional technicality. We preferred
to avoid it to focus on the specific non-linear aspects. Anyhow, such an assumption would allow to consider
drifts which have linear growth in space, like e.g. in the work by Mishura and Veretennikov [MV18].
• In the subcritical case α ∈ (1, 2), we believe that the approach of [CdRF18b] could be adapted in order to
consider for the drift term a Lipschitz continuity condition w.r.t the total variation metric on the space of
probability measures uniformly w.r.t. the time and space variables. In this context, one still has to perform
the above fixed point procedure on the space As,T,µ now equipped with the total variation distance. This
will concern future investigations.
• Let us point out that the condition α > 2η appears in Theorem 2.3 only for integrability purposes in order to
establish our Schauder estimates in the context of unbounded drift coefficients. In particular, it could also
be removed by employing a localization technique3 in the same spirit of the one employed in [CdRMP19]
or by assuming that the drift coefficient is globally bounded.
• Even though we only addressed the pure jump case, the methodology developed below (relying mainly
on the control of the explosions of the gradient of the solution of a PDE with Hölder terminal condition
and coefficients) would also apply in the diffusive stable setting corresponding to α = 2. Coupled to the
previously mentioned localization procedure, this could allow to consider non-linear drifts, locally Hölder
continuous in space and with spatial linear growth as well as a diffusion coefficient also depending on the
measure argument.

3. Sensitivity of the semi-groups w.r.t. the measure argument

The main purpose of this section is to prove Proposition 2.1. To derive the key control (2.15) the main point is
formally to expand one semi-group around the other and to exploit some Schauder like controls in the same spirit of
those established in [CdRMP19] under the current assumption (AS ). With the previous notations, defining for any
Pi
fixed t ∈ [0, T ] with T ≤ 1 small enough to be specified later on, (s, x) ∈ [0, t]×Rd and i ∈ {1, 2}, ui (s, x) := Ps,t h(x)
it is clear that ui is a mild solution of equation:
( 
∂s + AsPi ui (s, x) = 0, (s, x) ∈ [0, t) × Rd ,
(3.1)
ui (t, x) = h(x), x ∈ Rd .
with AsPi as in (2.5) and h ∈ C γ (Rd , R) where γ ∈ (0, 1) corresponds to the index of the underlying distance
employed for the fixed point procedure described in the proof of Theorem 2.3. Namely, for the proof of Proposition
2.1, that is, under the assumption α > 2η ∨ (1 − η2 ), we will take γ = η. This choice notably allows to derive the
announced result in the subcritical case α ≥ 1 and in the super-critical one up to α > 54 . In order to handle the
supercritical case α < 1 under the sole condition α > 2η ∨ (1 − η), we will take γ = 2η for the first iterate of the
map T and then γ = (1 − λ)2η for a suitable choice of λ ∈ [0, 1] for the second iterate of T. The main required
modification for the latter procedure is precisely described in Appendix B.
It now follows from the Schauder estimates established in the above reference that, if h ∈ C α+2η (Rd , R) then
ui ∈ L∞ ([0, t], C α+2η (Rd , R)). We carefully indicate that the results in [CdRMP19] are stated in the super-critical
case α < 1 but still hold for α ∈ [1, 2) (See Remark 6 therein). Also, the estimates of [CdRMP19] are established
for a constant diffusion coefficient σ. It is anyhow also indicated in Remark 5 of that work that those estimates
would extend to dynamics with a variable Hölder in space diffusion coefficient following the usual perturbation
approach in that case (see e.g. [MP14] or [ZZ18]). Hence, the estimates are still valid under our current assumption
(AS ). Furthermore, the techniques we develop below also emphasize how the dependence on the diffusion coefficient
must be handled to derive Schauder estimates (see e.g. Lemmas 3.3 and A.3 below which precisely investigate the
sensitivity of the non-local term with respect to a diffusion coefficient and the proof of Lemma 3.1).
The point now is that h is here only assumed to be in C γ (Rd , R) (and not in C α+2η (Rd , R)). For the analysis
let us now consider a mollification hδ ∈ C α+2η (Rd , R) of h ∈ C γ (Rd , R), i.e. hδ = h ⋆ ρδ with ⋆ standing for the
3which would actually be the same that the one allowing to consider the local Hölder condition for the drift described in the first
point
 WELL-POSEDNESS OF SOME NON-LINEAR STABLE DRIVEN SDES 9

convolution and for ρδ (·) = δ −d ρ( δ· ) where ρ : Rd → R+ is a smooth compactly supported function. In particular,
Pi δ
there exists C ≥ 1 s.t. khδ kC γ ≤ C and khδ − hkC γ −→ 0. Denoting by uδi (s, x) := Ps,t h (x), it is then clear from
δ→0
Pi
the uniform Hölder continuity of h that (uδi − ui )(s, x) = Ps,t (hδ − h)(x) −→ 0, uniformly in (s, x) ∈ [0, t] × Rd .
δ→0
Hence, it suffices to establish (2.15) for the mollified final functions uniformly w.r.t. the smoothing parameter δ.
To this end, observe that uδi (s, x) satisfies
( 
∂s + AsPi uδi (s, x) = 0, (s, x) ∈ [0, t) × Rd ,
(3.2)
uδi (t, x) = hδ (x), x ∈ Rd ,

and uδi ∈ L∞ ([0, t], C α+γ (Rd , R)).

To compare both semigroups, we now write the PDE satisfied by w(s, x) := wδ (s, x) = (uδ1 − uδ2 )(s, x). Namely,
  

 ∂s + AsP1 )w(s, x) = − b(s, x, P1 ) − b(s, x, P2 ) · Dx uδ2 (s, x)

 
(3.3) + (L P1 ,α
− L P2 ,α δ
)u (s, x) =: −HP1 ,P2 uδ2 (s, x), (s, x) ∈ [0, t) × Rd ,

 s s 2


w(t, x) = 0, x ∈ Rd .

The important point is now to observe that, proceeding this way we have isolated the terms with the difference of
the reference measures in the source (term HP1 ,P2 uδ2 in (3.3)). Hence, the above left hand side depends only on one
reference measure, here P1 , which will serve in order to introduce a proxy associated with the unbounded coefficients
of AsP1 similarly to what was previously performed in [CdRMP19] to derive
quantitative estimates. Observe that,
for the PDE (3.2) with mollified terminal condition, Duδ1 and LP s
i ,α δ
u1 i∈{1,2} are indeed well defined pointwise.
The following results will play a central role in our analysis. The first lemma controls the explosions of the
gradient and its Hölder modulus of continuity for the solution of (3.2). The second one gives controls of the source
HP1 ,P2 uδ2 in (3.3) in terms of the distance between the two reference measures, up to multiplicative and integrable
time singularities.

Lemma 3.1 (Regularity of the semi-groups for a Hölder source). Assume that α + γ > 1 and γ ∈ (0, 2η]. With the
notations of (3.2), there exists a constant C ≥ 1 s.t. for any Pi ∈ C([0, t], P(Rd )), i = 1, 2, for all (s, x) ∈ [0, t) × Rd
and all δ > 0:
1 γ
(3.4) |uδi (s, x)| + (t − s) α − α |Dx uδi (s, x)| ≤ C.

Also, for all ε ∈ (0, 1 − ( α1 − αγ )), there exists Cε s.t. for all (s, x, x′ ) ∈ [0, t) × (Rd )2 :

|Dx uδi (s, x) − Dx uδi (s, x′ )| h 1 γ
i

(3.5) (t − s)1−ε ≤ Cε , ϑ := ϑ(α, γ, ε) = α 1 − ε − − = α(1 − ε) − 1 + γ.
|x − x′ |ϑ α α
Remark 3.1 (About the explosion rates of the gradient). Observe from equation (3.4) that the gradient explodes
at an integrable rate in time, precisely from the condition α + γ > 1 (of course this is meaningful only in the
super-critical case α < 1, since when α ≥ 1 it is readily fulfilled by any γ > 0). Equation (3.5) then specifies how
far we can obtain Hölder moduli in space that explodes at an integrable rate. Namely, the gradient can be seen as a
function of regularity order α + γ − 1. We manage to control the corresponding Hölder modulus with an integrable
singularity in time up to an additional small factor εα (exponent ϑ above), where in turn ε quantifies the distance
to explosion.

Lemma 3.2 (Controls for the source involving the different measures). Let uδi solve (3.2). Taking then, with the
notation introduced in (3.5), ϑ > α − 1, it holds:
 γ


1
(3.6) |HP1 ,P2 uδi (s, x)| ≤ Cd2η (P1 (s), P2 (s)) (t − s)−1+ε 1{α≥1} + (t − s)− α − α ,

and for all (x, x′ ) ∈ (Rd )2 , for any λ ∈ [0, 1]

|HP1 ,P2 uδi (s, x) − HP1 ,P2 uδi (s, x′ )|

 γ


1
(3.7) ≤ Cd(1−λ)2η (P1 (s), P2 (s))|x − x′ |2λη (t − s)−1+ε 1{α≥1} + (t − s)− α−α

+ Cd2η (P1 (s), P2 (s))(|x − x′ |ϑ + |x − x′ |1+ϑ−α )(t − s)−1+ε .

10 NOUFEL FRIKHA, VALENTIN KONAKOV, AND STÉPHANE MENOZZI

3.1. Proxy and a priori controls. For given freezing parameters (τ, ξ) ∈ [0, T ] × Rd to be specified later on,
introduce the flow:
Z s
(3.8) θs,τ (ξ) = ξ + b(v, θv,τ (ξ), P1 (v))dv, s ≥ τ,
τ
θs,τ (ξ) = ξ for s < τ . In the current setting (Hölder drift in space), this object exists from the Peano theorem but
is not necessarily unique. Rewrite then (3.3) as follows
  
 P1 ,α,(τ,ξ)

 ∂v + L̃ v w(v, x) + b(v, θv,τ (ξ), P1 (v)) · Dx w(v, x)
 
(3.9) δ α,(τ,ξ)
 = − H u
P1 ,P2 2 (v, x) + RP1 (v, x) , (v, x) ∈ [0, t) × Rd ,


w(t, x) = 0, on Rd ,
where for all ϕ ∈ C02 (Rd , R)
Z


(3.10) L̃vP1 ,α,(τ,ξ) ϕ(x) := p.v. ϕ(x + σ(v, θv,τ (ξ), P1 (v))ζ) − ϕ(x) ν(dζ)
Rd
and

 
α,(τ,ξ)
(3.11) RP1 (v, x) := b(v, x, P1 (v)) − b(v, θv,τ (ξ), P1 (v)) · Dx w(v, x) + LsP1 ,α − L̃sP1 ,α,(τ,ξ) w(v, x).
α,(τ,ξ)
Precisely, the contribution RP1 corresponds, for the measure argument fixed to P1 , to the difference of the
generator with variable coefficients and the proxy one which is frozen along the deterministic flow (3.8). The
following Lemma proved in Appendix A.4 gives quantitative bounds on this remainder term.
Lemma 3.3 (Bounds for the difference of the generators for a fixed measure argument). There exists a constant
C s.t. for all (τ, ξ), (v, y) ∈ [0, T ] × Rd :
 
α,(τ,ξ)
RP1 (v, y) ≤ C kb(v, ·, P1 (v))kC˙2η |Dy w(v, y)| + kσ(v, ·, P1 (v))kC 2η (kwk∞ + kDx w(v, ·)k∞ )1{α<1}

(3.12) +kDx w(v, ·)kC ϑ 1{α>1} + kw(v, ·)kC 1+ϑ 1{α=1} |y − θv,τ (ξ)|2η .
P ,α,(τ,ξ)
Also, under (AS ), it is clear that the time-dependent operator L̃v 1 + b(v, θv,τ (ξ), P1 ) · Dx generates a
P ,α,(τ,ξ)
family of transition probability (or two parameter transition semi-group) P̃s,v1 )0≤s≤v≤t . Because of the non-
degeneracy conditions of (2.7) and (2.9) in (ND) and (UE), the associated heat-kernel exists (see e.g. [Kol00],
[Wat07] or Appendix A below) and, for fixed 0 ≤ s < v ≤ t, it writes:
 
(3.13) p̃P1 ,α,(τ,ξ) (s, v, x, y) = pΘs,v,(τ,ξ),P1 y − m(τ,ξ)
v,s (x) ,

where we denoted:
Z v
(3.14) m(τ,ξ)
v,s (x) := x + b(r, θr,τ (ξ), P1 (r))dr,
s
Z v
(3.15) Θs,v,(τ,ξ),P1 := σ(r, θr,τ (ξ), P1 (r))dZr ,
s
and where Z is a stable process with Lévy measure ν defined on some probability space (Ω, A, P).
(τ,ξ)
Observe that from the definition (3.14) of the shift mv,s (x) we have the important property
(3.16) m(τ,ξ)
v,s (x)|(τ,ξ)=(s,x) = θv,s (x).

Let us first give some bounds on the density function of the stochastic integral (3.15) and its derivatives. They
are somehow classic under the absolute continuity condition (AC) for the Lévy measure (see e.g. [Kol00]). For the
sake of completeness, the proof of the following result is postponed to Appendix A. .
Lemma 3.4 (Controls on the proxy density and its derivatives). Assume (ND) and (UE) hold. Then, there exists
a constant C and a probability density q̄ : R+ d
∗ × R → R satisfying for all γ < α the integrability property
Z
γ
(3.17) |z|γ q̄(t − s, z)dz ≤ C(t − s) α ,
Rd

s.t. for all multi-index β, |β| ≤ 3 and for any 0 ≤ s < t, (τ, ξ) ∈ R+ × Rd , µ ∈ C(R+ , P(Rd )), y ∈ Rd :
C
(3.18) |Dyβ pΘs,t,(τ,ξ),µ (y)| ≤ |β|
q̄(t − s, y).
(t − s) α
 WELL-POSEDNESS OF SOME NON-LINEAR STABLE DRIVEN SDES 11

1
Also, any diagonal perturbation does not affect the previous estimate. Namely, for z ∈ Rd such that |z| ≤ (t − s) α ,
it holds:
C
(3.19) |Dyβ pΘs,t,(τ,ξ),µ (y + z)| ≤ |β|
q̄(t − s, y).
(t − s) α
As a very important consequence of this lemma, we also derive from (3.13)-(3.14) that for all (x, y) ∈ (Rd )2 :

C

(3.20) |Dxβ p̃P1 ,α,(τ,ξ) (s, v, x, y)| := |Dxβ pΘs,v,(τ,ξ),P1 y − m(τ,ξ)
v,s (x) | ≤ |β|
q̄ v − s, y − m(τ,ξ)
v,s (x) .
(v − s) α
3.2. Proof of Proposition 2.1. Throughout this subsection, we set γ = η, that is, h ∈ C η (Rd , R) with khkC η ≤ 1.
With the notations and controls of the previous subsection at hand, we can now restart from (3.9) to write for all
(s, x) ∈ [0, t] × Rd ,
Z t 
P1 ,α,(τ,ξ) α,(τ,ξ)

w(s, x) = dv P̃s,v HP1 ,P2 uδ2 (v, ·) + RP1 (v, ·) (x)
s
Z t Z
α,(τ,ξ)

(3.21) = dv p̃P1 ,α,(τ,ξ) (s, v, x, y) HP1 ,P2 uδ2 (v, y) + RP1 (v, y) dy.
s Rd
α,(τ,ξ)
From (3.21) and the definition of in (3.11), we see that w appears in both sides of the equality. The point
RP1
is now to perform a circular argument to control w. To this end, we also point out that the a priori controls of
Lemma 3.1 will be useful to control HP1 ,P2 uδ2 introduced in (3.3). To complete the proof of Proposition 2.1 we aim
at showing the following:
(3.22) |w(s, x)| ≤ Cds,t,η (P1 , P2 )(t − s)ζ̃ ,
for some ζ̃ > 0, which precisely gives (2.15) for h replaced by hδ .
Write first that, from (BH ), the definition of HP1 ,P2 uδ2 and Lemma 3.1 (see equations (3.5) and (3.6) with
γ = η), it holds that there exists a constant C s.t. for all (v, y) ∈ [s, t) × Rd :
 η


1
(3.23) |HP1 ,P2 uδ2 (v, y)| ≤ Cdη (P1 (v), P2 (v)) (t − v)−1+ε 1{α≥1} + (t − v)− α − α ,
using as well for the above inequality that, from the definitions in (2.1)-(2.2),
1
(3.24) ∀η ∈ (0, ], d2η (P1 (v), P2 (v)) ≤ dη (P1 (v), P2 (v)).
2
This bound will be frequently used in the sequel.
On the other hand, from Lemma 3.3 and (3.20):
Z t Z
α,(τ,ξ)
(3.25) dv p̃P1 ,α,(τ,ξ) (s, v, x, y)RP1 (v, y)dy
s R d
Z t Z  
≤C dv dy kb(v, ·, P1 (v))kĊ 2η |Dy w(v, y)| + kσ(v, ·, P1 (v))kC 2η (kwk∞ + kDx w(v, ·)k∞ )1{α<1}
s R d


+kDx w(v, ·)kC ϑ 1{α>1} + kw(v, ·)kC 1+ϑ 1{α=1} |y − θv,τ (ξ)|2η q̄ v − s, y − m(τ,ξ) v,s (x) ,

denoting here and for the rest of the section with a slight abuse of notation kwk∞ := supv∈[s,t] kw(v, ·)k∞ .
Hence, choosing (τ, ξ) = (s, x) and exploiting (3.16) and Lemma 3.4, we derive:
Z t Z
α,(τ,ξ)
dv p̃P1 ,α,(τ,ξ) (s, v, x, y)RP1 (v, y)dy
s Rd
(τ,ξ)=(s,x)
Z t  
2η
(3.26) ≤C dv(v − s) α kwk∞ 1{α≤1} + kDx w(v, ·)k∞ + kDx w(v, ·)kC˙ϑ 1{α≥1} .
s
From (3.26), (3.23) and (3.21), one derives:

 1 η 1 η 
|w(s, x)| ≤ C dη,s,t (P1 , P2 ) (t − s)ε∧(1−( α − α )) 1{α≥1} + (t − s)(1−( α − α )) + (t − s)kwk∞ 1{α≤1}
Z t
2η
 
(3.27) + dv(v − s) α kDx w(v, ·)k∞ + kDx w(v, ·)kC˙ϑ 1{α≥1} .
s
It is now clear that we actually have to estimate the gradient of w, i.e. we need to differentiate (3.21). We will now
split the rest of the analysis in function of the index α. Indeed, for α > 1, differentiating (3.21) yields an integrable
12 NOUFEL FRIKHA, VALENTIN KONAKOV, AND STÉPHANE MENOZZI

singularity in time and the previous controls do not change much. The counterpart is that we also need to give
a bound on the Hölder modulus of continuity. On the other hand, for α ≤ 1, the induced time singularity is not
integrable and it is crucial to equilibrate it through adequate cancellation techniques.
The case α ≥ 1. From (3.21), we can differentiate w.r.t x and also use a cancellation technique for the source term
HP1 ,P2 uδ2 . Namely,
Z t Z  
 α,(τ,ξ)
(3.28) Dx w(s, x) = dv Dx p̃P1 ,α,(τ,ξ) (s, v, x, y) HP1 ,P2 uδ2 (v, y) − HP1 ,P2 uδ2 (v, θv,τ (ξ)) + RP1 (v, y) dy.
s Rd
Similarly to (3.23), we derive from (BH ), equation (3.7) (with λ = 1/2) of Lemma 3.1 and (3.24) that:
h  η


1
|HP1 ,P2 uδ2 (v, y) − HP1 ,P2 uδ2 (v, θv,τ (ξ))| ≤ Cdη (P1 (v), P2 (v)) (|y − θv,τ (ξ)|η ∧ 1) (t − v)−1+ε + (t − v)− α−α

i
(3.29) +(t − v)−1+ε (|y − θv,τ (ξ)|ϑ ∧ 1) + (|y − θv,τ (ξ)|1+ϑ−α ∧ 1)
h  η


1
≤ Cdη (P1 (v), P2 (v)) (|y − θv,τ (ξ)|η ∧ 1) (t − v)−1+ε + (t − v)− α−α

i
+(t − v)−1+ε (|y − θv,τ (ξ)|1+ϑ−α ∧ 1) ,
recalling that since α ≥ 1, 1 + ϑ − α ≤ ϑ for the last inequality. From the definition in (3.5) we thus get that, for ε
meant to be small,
h 1 η
i
(3.30) 1+ϑ−α =1−α+α 1−ε− − = η − εα.
α α
We thus obtain the following control
(3.31) |HP1 ,P2 uδ2 (v, y) − HP1 ,P2 uδ2 (v, θv,τ (ξ))| ≤ Cdη (P1 (v), P2 (v))|y − θv,τ (ξ)|η−εα (t − v)−1+ε .
Hence, from (3.28), (3.31), Lemma 3.3 and (3.18), we get:
(3.32) |Dx w(s, x)|
 Z t Z
dv

≤ C dη,s,t (P1 , P2 ) 1 (t − v)−1+ε dy|y − θv,τ (ξ)|η−εα q̄ v − s, y − m(τ,ξ)
v,s (x)
s (v − s) α Rd
Z t Z 
dv
+ 1 dy kb(v, ·, P1 (v))kĊ 2η |Dy w(v, y)|
s (v − s) α Rd
 

+kσ(v, ·, P1 (v))kC 2η kDx w(v, ·)kC ϑ 1α>1 + kw(v, ·)kC 1+ϑ 1α=1 |y − θv,τ (ξ)|2η q̄ v − s, y − m(τ,ξ)
v,s (x) .
Taking (τ, ξ) = (s, x), it follows again from (3.16) and Lemma 3.4 that:
 Z t
1 η
|Dx w(s, x)| ≤ C dη,s,t (P1 , P2 ) dv(v − s)− α + α −ε (t − v)−1+ε
s
Z t
1 2η
 
+ dv(v − s)− α + α kDx w(v, ·)k∞ + kDx w(v, ·)kC˙ϑ + kw(v, ·)k∞ 1α=1
s
1 η
≤ C dη,s,t (P1 , P2 )(t − s)− α + α
Z t  
1
−α + 2η
(3.33) + dv(v − s) α kDx w(v, ·)k∞ + kDx w(v, ·)kC˙ϑ + kw(v, ·)k∞ 1α=1 .
s
From (3.33), we also need to estimate the gradient through a circular argument.
Let us now make a short comment before proceeding further. We cannot expect for the gradient of w a better
behaviour than the one provided by the main term of the perturbative expansion (3.21), i.e. the one associated
with HP1 ,P2 uδ2 . Observe now that we get: − α1 + αη =: ζ < 0. This exponent being negative, we cannot expect to
have a pointwise bound for the gradient of w. Having in mind that we want to keep track in the above r.h.s. of
the product of the distance dη,s,t (P1 , P2 ) and a contribution of type (t − s)ζ̄ , ζ̄ > 0, which provides a smoothing
in time effect, we will therefore investigate the behavior of
1 η ε
(3.34) Φ(v) := (t − v)Ξ sup |Dw(v, x)|, Ξ := − + , v ∈ [s, t],
x∈R d α α 2
To investigate the explosion of the corresponding Hölder modulus, let us introduce as well
ϑ
(3.35) Ψ(v) := (t − v)Ξ+ α kDw(v, ·)kC ϑ , v ∈ [s, t].
 WELL-POSEDNESS OF SOME NON-LINEAR STABLE DRIVEN SDES 13

Let us note that supv∈[s,t] Φ(v) < ∞ and supv∈[s,t] Ψ(v) < ∞. Indeed, from the definition of HP1 ,P2 uδ2 in (3.3), since


from [CdRMP19] uδ2 ∈ L∞ [0, T ], C α+2η (Rd , R) (with corresponding norm which a priori explodes with δ), it holds


that, if b is bounded in space, HP1 ,P2 uδ2 ∈ L∞ [0, T ], C 2η∧(α+2η−1) (Rd , R) . Hence, the previously invoked Schauder


estimates still apply for w solving (3.3) (again with corresponding L∞ [0, T ], C α+2η∧(α+2η−1) (Rd , R) norm explod-
ing with δ). If now b is not bounded in space, recall that a priori,
we assumed in (BH ) that b(t, ·, µ) ∈ Ċ 2η (Rd , Rd ),
we would then have HP1 ,P2 uδ2 ∈ L∞ [0, T ], Ċ 2η∧(α+2η−1) (Rd , R) . In that case, from the integrability constraint
α > 2η, the arguments of [CdRMP19] could be reproduced to derive the finiteness of supv∈[s,t] Φ(v), supv∈[s,t] Ψ(v)
which involve the gradient which is in that case bounded whereas the function itself is not (see e.g. Krylov and
Priola [KP10] for similar issues in the diffusive setting).
From the previous definitions, a key point, for our circular argument to work is that Ξ + αϑ < 1. Indeed, from
(3.33), this is necessary in order to make the time singularities generated by Φ and Ψ integrable. We now recall
from the definition of ϑ in (3.5) that:

ϑ 1 η ε 1 η ε
(3.36) Ξ+ = − + + 1 − ε − ( − ) = 1 − < 1.
α α α 2 α α 2
With these notations at hand, we can from (3.27) get rid of the supremum norm kwk∞ appearing in the case
α = 1 and upper bound it in terms of the functions Φ, Ψ. Namely, for ε and T small enough (so that t − s < 1/2),

kwk∞ ≤ C dη,s,t (P1 , P2 )(t − s)ε + (t − s)kwk∞ 1{α=1}
Z t
2η
 ϑ ϑ

+ dv(v − s) α (t − v)− α Φ(v) + (t − v)−(Ξ+ α ) Ψ(v) .
s
 !
ε
(3.37) ≤ C dη,s,t (P1 , P2 )(t − s) + sup Φ(v) + sup Ψ(v) .
v∈[s,t] v∈[s,t]

up to a modification of the constant C for the last inequality.

We now have to distinguish the diagonal and off-diagonal regimes w.r.t. the current considered times. Namely,
1
for given (x, x′ ) ∈ (Rd )2 , if |x − x′ | ≥ (t − s) α we say that the off-diagonal regime holds. In this case, we readily
get from (3.33) and (3.37):

ϑ |Dx w(s, x) − Dx w(s, x′ )|

(t − s) α ≤ |Dw(s, x) − Dw(s, x′ )| ≤ |Dw(s, x)| + |Dw(s, x′ )|
|x − x′ |ϑ

≤ C dη,s,t (P1 , P2 )(t − s)ζ
Z t
1 2η
 ϑ

+ dv(v − s)− α + α (t − v)−Ξ Φ(v) + (t − v)−(Ξ+ α ) Ψ(v) ,
s
ϑ
Ξ+ α |Dx w(s, x) − Dx w(s, x′ )| 1 2η
(t − s) ≤ C dη,s,t (P1 , P2 )(t − s)Ξ+ζ + (t − s)1− α + α sup Φ(v)
|x − x′ |ϑ v∈[s,t]

1 ϑ 2η
(3.38) +(t − s)1−( α + α )+ α sup Ψ(v) .
v∈[s,t]

1
Let us now turn to the global diagonal regime, |x − x′ | ≤ (t − s) α . This case is more subtle. Indeed, due to the
time integration, i.e. for v ∈ [s, t], there is a local off-diagonal regime, namely for v ∈ [s, s + c0 |x − x′ |α ] for a
constant c0 ∈ (0, 1) to be specified, and then a true diagonal regime w.r.t. the integration variable v which is the
one providing the smoothing effects through the heat kernel. The delicate point is that, in order to exploit the
off-diagonal bounds, the natural choice of freezing spatial point consists in considering precisely the spatial point
in argument of the function. On the other hand, in the true diagonal regime the freezing spatial point must be the
same for the two quantities to expand. This naturally induces to consider an intermediate quantity for the regime
change.
Introduce for a couple (τ, ξ ′ ) ∈ R+ × Rd of freezing points, the following Green kernel. :
Z r Z
′ ′
∀0 ≤ s < r ≤ t, x ∈ Rd , G̃s,r
P1 ,α,(τ,ξ )
f (s, x) := dv dy p̃P1 ,α,(τ,ξ ) (s, v, x, y)f (v, y).
s Rd
14 NOUFEL FRIKHA, VALENTIN KONAKOV, AND STÉPHANE MENOZZI

For x, x′ ∈ Rd s.t. |x − x′ |α ≤ (t − s) we can write, similarly to the proof of Proposition 12 in [CdRMP19] (Duhamel
formula with change of freezing point), that:

Dx w(s, x′ )
P1 ,α,(τ,ξ ′ )

P ,α,(τ,ξ̃ ′ )

= Dx G̃s,τ 0
(HP1 ,P2 uδ2 (s, x′ ) (τ ,τ,ξ′ )=(t ,s,x′ ) + Dx G̃τ01,t (HP1 ,P2 uδ2 )(s, x′ ) (τ ,τ,ξ̃′ )=(t ,s,x)
0 0 0 0

P1 α,(τ,ξ ′ ) ′

P1 ,α,(τ,ξ̃ ′ ) ′


+ Dx P̃s,τ0 w(τ0 , x ) (τ0 ,τ,ξ′ )=(t0 ,s,x′ ) − Dx P̃s,τ0 w(τ0 , x ) (τ0 ,τ,ξ̃′ )=(t0 ,s,x)
Z t Z
′ α,(τ,ξ ′ )
+ dv dy 1{v≤τ0 } Dx p̃P1 ,α,(τ,ξ ) (s, v, x′ , y)RP1 (v, y)
s Rd
!
P1 ,α,(τ,ξ̃ ′ ) ′ α,(τ,ξ̃ ′ )
(3.39) +1{v>τ0 } Dx p̃ (s, v, x , y)RP1 (v, y) ,
(τ0 ,τ,ξ ′ ,ξ̃ ′ )=(t0 ,s,x′ ,x)

where

(3.40) t0 = s + c0 |x − x′ |α .

The point is now precisely to use the two expansions of the gradient in (3.21) (that we first differentiate w.r.t. x
choosing then (τ, ξ) = (s, x)) and (3.39), keeping in mind that the additional term

P1 α,(τ,ξ ′ )

P1 ,α,(τ,ξ̃ ′ )


Dx P̃s,τ 0
w(τ0 , x′ ) (τ0 ,τ,ξ′ )=(t0 ,s,x′ ) − Dx P̃s,τ 0
w(τ0 , x′ ) (τ0 ,τ,ξ̃′ )=(t0 ,s,x) ,

arising from the change of freezing points precisely needs to be analyzed.

We get:

Dx w(s, x) − Dx w(s, x′ )
Z t Z


α,(τ,ξ)
= dv1{v≤τ0 } Dx p̃P1 ,α,(τ,ξ) (s, v, x, y) HP1 ,P2 uδ2 (v, y) + RP1 (v, y) dy
s Rd
Z t Z
′ α,(τ,ξ ′ )

− dv1{v≤τ0 } Dx p̃P1 ,α,(τ,ξ ) (s, v, x′ , y) HP1 ,P2 uδ2 (v, y) + RP1 (v, y) dy
Rd (τ0 ,τ,ξ,ξ ′ )=(t0 ,s,x,x′ )
s


P1 α,(τ,ξ ′ ) ′ P1 ,α,(τ,ξ̃ ′ ) ′
− Dx P̃s,τ0 w(τ0 , x ) (τ0 ,τ,ξ′ )=(t0 ,s,x′ ) − Dx P̃s,τ0 w(τ0 , x ) (τ0 ,τ,ξ̃′ )=(t0 ,s,x)
Z t Z


+ dv1{v>τ0 } Dx p̃P1 ,α,(τ,ξ) (s, v, x, y) − Dx p̃P1 ,α,(τ,ξ) (s, v, x′ , y)
s Rd
α,(τ,ξ)

× HP1 ,P2 uδ2 (v, y) + RP1 (v, y) dy
(τ0 ,τ,ξ)=(t0 ,s,x)
=: ∆w1 (s, x, x′ ) + ∆w2 (s, x, x′ ) + ∆w3 (s, x, x′ ).
(3.41)

The term ∆w1 (s, x, x′ ), corresponding to the local off-diagonal regime within the global diagonal one, can be
analyzed as above. Proceeding as in (3.31)-(3.33), and with the bound of (3.37) we get:

|∆w1 (s, x, x′ )|
Z τ0 Z
α,(τ,ξ)

≤ dv Dx p̃P1 ,α,(τ,ξ) (s, v, x, y) [HP1 ,P2 uδ2 (v, y) − HP1 ,P2 uδ2 (v, θv,τ (ξ))] + RP1 (v, y) dy
s Rd (τ0 ,τ,ξ)=(t0 ,s,x)
Z τ0 Z
′ α,(τ,ξ )

′
+ dv Dx p̃P1 ,α,(τ,ξ ) (s, v, x′ , y) [HP1 ,P2 uδ2 (v, y) − HP1 ,P2 uδ2 (v, θv,τ (ξ ′ ))] + RP1 (v, y) dy
s Rd (τ0 ,τ,ξ ′ )=(t0 ,s,x′ )
Z t0
dv
≤ C 1 η dη (P1 (v), P2 (v))(t − v)−1+ε
s (v − s) α −( α −ε)
Z t0 
−α1
+ 2η
 −Ξ ϑ
−(Ξ+ α )
 |x − x′ |ϑ
+ dv(v − s) α (t − v) Φ(v) + (t − v) Ψ(v) ϑ .
s (v − s) α
 WELL-POSEDNESS OF SOME NON-LINEAR STABLE DRIVEN SDES 15

From (3.40), we then get that for v ∈ [s, t0 ], (t − v) ≥ (1 − c0 )(t − s) and therefore, recalling from (3.40) that
ϑ
|x − x′ |ϑ = (c−1
0 (t0 − s)) :
α


|∆w1 (s, x, x′ )| η 1 ϑ
≤ C (t − s)−1+ε dη,s,t (P1 , P2 )(t0 − s)1+( α −ε)−( α + α )
|x − x′ |ϑ
Z t0
−Ξ 1 2η 1
+(t − s) dv(v − s)− α + α Φ(v) ϑ
s (v − s) α
Z t0 
ϑ 1 2η 1
+(t − s)−(Ξ+ α ) dv(v − s)− α + α Ψ(v) ϑ
s (v − s) α

≤ C (t − s)−1+ε dη,s,t (P1 , P2 )

−Ξ+1− α1
+ 2η ϑ
−α −(Ξ+ α ϑ 1
)+1− α + 2η ϑ
−α
+(t − s) α sup Φ(v) + (t − s) α sup Ψ(v) ,
v∈[s,t] v∈[s,t]

recalling for the last inequality that:

η 1 ϑ η 1 1 η

(3.42) 1+( − ε) − ( + ) = 1 + ( − ε) − + 1 − ε − ( − ) = 0.
α α α α α α α
This eventually gives:
ϑ |∆w1 (s, x, x′ )|  ϑ
(t − s)Ξ+ α ≤ C (t − s)Ξ+ α −1+ε dη,s,t (P1 , P2 )
|x − x′ |ϑ
1 2η 1 ϑ 2η

+(t − s)1− α + α sup Φ(v) + (t − s)1−( α + α )+ α sup Ψ(v) .
v∈[s,t] v∈[s,t]
1 η ε
Recalling from (3.34) that Ξ := α − α + 2, from (3.42) we get:
ϑ 1 ϑ η 3 η η 3 ε
(3.43) Ξ+ − 1 + ε = ( + ) − − 1 + ε = 1 − ε + − − 1 + ε = =: ζ̄.
α α α α 2 α α 2 2
Hence
ϑ |∆w1 (s, x, x′ )| 
(t − s)Ξ+ α ≤ C (t − s)ζ̄ dη,s,t (P1 , P2 )
|x − x′ |ϑ
1 2η 1 ϑ 2η

(3.44) +(t − s)1− α + α sup Φ(v) + (t − s)1−( α + α )+ α sup Ψ(v) .
v∈[s,t] v∈[s,t]

Turning now to ∆w3 in (3.41), expanding the frozen densities, exploiting as well (3.19) which gives that a diagonal
perturbation of the density does not affect the related estimates, we write:
Z t Z 1 Z
|∆w3 (s, x, x′ )| ≤ dv dλ Dx2 p̃P1 ,α,(τ,ξ) (s, v, x + λ(x′ − x), y)(x′ − x)
τ0 0 Rd

α,(τ,ξ)

× [HP1 ,P2 uδ2 (v, y) − HP1 ,P2 uδ2 (v, θv,τ (ξ))] + RP1 (v, y) dy
(τ0 ,τ,ξ)=(t0 ,s,x)
 Z t
′ dv
≤ C |x − x |dη,s,t (P1 , P2 ) 2 η (t − v)−1+ε
(v − s) α −( α −ε)
t0
Z t   
2 2η ϑ
+|x − x′ |ϑ dv(v − s)− α + α (t − v)−Ξ Φ(v) + (t − v)−(Ξ+ α ) Ψ(v) + |x − x′ |1−ϑ ,
t0

where we also used (3.7), for γ = η, λ = 21 , and (3.24) for the last inequality. Splitting the first time integral in
the right-hand side of the above inequality into the two disjoint intervals [t0 , t+s t+s
2 ] and ( 2 , t], noting that in the
1 ′ α
current diagonal regime, for c0 ∈ (0, 2 ), t0 = s + c0 |x − x | ≤ (t + s)/2, we get (recall that we are far from the
singularity):
Z t+s
2 2 η 2 η
dv(v − s)− α +( α −ε) (t − v)−1+ε ≤ C(t − s)−1+ε (t0 − s)− α +( α −ε)+1
t0
and Z t
2 η 2 η 2 η
dv(v − s)− α +( α −ε) (t − v)−1+ε ≤ C(t − s)ε (t − s)− α +( α −ε) = C(t − s)− α + α .
t+s
2
16 NOUFEL FRIKHA, VALENTIN KONAKOV, AND STÉPHANE MENOZZI

1
Recalling that, for v ∈ [t0 , t], |x − x′ | ≤ (v − s)/c0 α , we derive:

2 η 2 η
|∆w3 (s, x, x′ )| ≤ C dη,s,t (P1 , P2 )|x − x′ |[(t − s)−1+ε (t0 − s)− α +( α −ε)+1 + (t − s)− α + α ]
Z t 
′ ϑ
ϑ−1 1
−α ϑ
−α + 2η
 −Ξ ϑ
−(Ξ+ α )

+|x − x | c0 α
dv(v − s) α (t − v) Φ(v) + (t − v) Ψ(v) .
t0

2 η 1
Recall, from (3.42), − α2 + ( αη − ε) + 1 = αϑ − α1 . This yields (t0 − s)− α +( α −ε)+1 = (c0α |x − x′ |)−1+ϑ . Also in the
2 η 1−ϑ 2 η ϑ 1 η
current diagonal regime |x − x′ |1−ϑ (t − s)− α + α ≤ (t − s) α − α + α = (t − s)− α − α + α . Eventually, since from (3.42)
and (3.43) the other time singularities are integrable, we get:
′

ϑ |∆w3 (s, x, x )| ϑ−1 ϑ 1 η
Ξ+ α
(t − s) ≤ C d η,s,t (P1 , P2 )[c 0
α
(t − s)Ξ+ α −1+ε + (t − s)Ξ− α + α ]
|x − x′ |ϑ
ϑ−1 h i
1− α1
+ 2η 1−( α1 ϑ
+α )+ 2η
(3.45) +c0 α
(t − s) α sup Φ(v) + (t − s) α sup Ψ(v) .
v∈[s,t] v∈[s,t]

The term ∆w2 (s, x, x′ ) in (3.41), corresponding to the change of freezing point remains to be analyzed. A key
(τ,ξ ′ )
point for this contribution is to note from (3.13)-(3.14) (affine structure in x′ of mτ0 ,s (x′ )) that:
′
 ′

Dx p̃P1 ,α,(τ,ξ ) (s, τ0 , x′ , y) = Dx pΘs,τ0 ,(τ,ξ′ ),P1 τ0 − s, y − mτ(τ,ξ
0 ,s
) ′
(x )
 Z τ0


= Dx pΘs,τ0 ,(τ,ξ′ ),P1 τ0 − s, y − x′ + b(r, θr,τ (ξ ′ ), P1 (r))dr
s
 Z τ0
′


= −Dy pΘs,τ0 ,(τ,ξ′ ),P1 τ0 − s, y − x + b(r, θr,τ (ξ ′ ), P1 (r))dr
s
P1 ,α,(τ,ξ ′ ) ′
(3.46) = −Dy p̃ (s, τ0 , x , y).

Hence, with the notation of (3.14), we get:

∆w2 (s, x, x′ )
Z ′
Z
′

= dy p̃P1 ,α,(τ,ξ ) (s, τ0 , x′ , y)Dy w(τ0 , y) − dy p̃P1 ,α,(τ,ξ̃ ) (s, τ0 , x′ , y)Dy w(τ0 , y)
Rd Rd (τ,τ0 ,ξ ′ ,ξ̃ ′ )=(s,t0 ,x′ ,x)
Z ′
h ′
i
= dy p̃P1 ,α,(τ,ξ ) (s, τ0 , x′ , y) Dy w(τ0 , y) − Dy w(τ0 , m̃τ,ξ s,τ0 (x ))
′
Rd (τ,τ0 ,ξ ′ )=(s,t0 ,x′ )
 ′ ′

+ Dy w(τ0 , m̃τ,ξ ′ τ,ξ̃
s,τ0 (x )) − Dy w(τ0 , m̃s,τ0 (x ))
′
(τ,τ0 ,ξ ′ ,ξ̃ ′ )=(s,t0 ,x′ ,x)
Z ′
h i
ξ̃ ′
− dy p̃P1 ,α,(τ,ξ̃ ) (s, τ0 , x′ , y) Dy w(τ0 , y) − Dy w(τ0 , m̃τ, s,τ0 (x ))
′
,
Rd (τ,τ0 ,ξ̃ ′ )=(s,t0 ,x)

R ′ R ′
recalling that Rd dy p̃P1 ,α,(τ,ξ ) (s, τ0 , x′ , y) = 1 and Rd dy p̃P1 ,α,(τ,ξ̃ ) (s, τ0 , x′ , y) = 1 for the last equality.
From (3.13), Lemma 3.4 and the definitions in (3.35), we again derive:
X Z ϑ
′
|∆w2 (s, x, x )| ≤ C dy q̄(t0 − s, y − m̃s,z ′ s,z ′ ϑ
s,t0 (x ))|y − m̃s,t0 (x )| Ψ(t0 )(t − t0 )
−(Ξ+ α )

z∈{x,x′ } Rd
′ ϑ
+|m̃s,x ′ s,x ′ ϑ
s,t0 (x ) − m̃s,t0 (x )| Ψ(t0 )(t − t0 )
−(Ξ+ α )

 Z t0 ϑ
ϑ ϑ
≤ Ψ(t0 )(t − t0 )−(Ξ+ α ) C|t0 − s| α + (b(v, θv,s (x′ ), P1 (v)) − b(v, θv,s (x), P1 (v)))dv .
s
1
Now, from Lemma 11 in [CdRMP19] we have |θv,s (x′ ) − θv,s (x)| ≤ C(|x − x′ | + (v − s) α ) and therefore:
Z t0 Z t0
′
(b(v, θv,s (x ), P1 (v)) − b(v, θv,s (x), P1 (v)))dv ≤ C |θv,s (x′ ) − θv,s (x)|2η dv
s s
1
≤ C(t0 − s)(|x − x′ | + (t0 − s) α )2η
≤ Cc0 |x − x′ |α+2η ≤ Cc0 |x − x′ |,
 WELL-POSEDNESS OF SOME NON-LINEAR STABLE DRIVEN SDES 17

recalling that, α + 2η > 1 and since we are in the diagonal regime and that the time t − s ≤ 1 we indeed have
|x − x′ | ≤ 1. This finally gives:
ϑ ϑ
|∆w2 (s, x, x′ )| ≤ CΨ(t0 )(t − t0 )−(Ξ+ α ) c0α |x − x′ |ϑ .
Recalling that t − t0 = t − s − (t0 − s) ≥ t − s − c0 |x − x′ |α ≥ (1 − c0 )(t − s), this finally yields:
ϑ |∆w2 (s, x, x′ )| ϑ ϑ
(3.47) (t − s)Ξ+ α ′ ϑ
≤ Cc0α (1 − c0 )−(Ξ+ α ) Ψ(t0 ).
|x − x |
1
Plugging (3.44), (3.45) and (3.47) into (3.41), we get that in the diagonal case |x − x′ | ≤ (t − s) α and with the
definition of ζ̄ = 2ε in (3.43):
ϑ |Dx w(s, x) − Dx w(s, x′ )|  ϑ−1 h 1 2η
(t − s)Ξ+ α ≤ C (1 + c0 α ) dη,s,t (P1 , P2 )(t − s)ζ̄ + (t − s)1− α + α sup Φ(v)
|x − x′ |ϑ v∈[s,t]
i ϑ

1+ 2η 1 ϑ
α −( α + α )
ϑ
−(Ξ+ α )
(3.48) +(t − s) sup Ψ(v) + + c0 (1 − c0 )
α
Ψ(t0 ) .
v∈[s,t]

Putting together (3.48) and (3.38), corresponding respectively to the control of the normalized Hölder modulus of
the gradient in the off-diagonal and diagonal regimes, we eventually derive:
 ϑ−1 ϑ−1 1 2η
Ψ(s) ≤ C dη,s,t (P1 , P2 )(1 + c0 α )(t − s)ζ̄ + (1 + c0 α )(t − s)1− α + α sup Φ(v)
v∈[s,t]
ϑ−1
1− 1+ϑ 2η ϑ ϑ

(3.49) + sup Ψ(r) (1 + c0 α
)(t − s) α + α + c0 (1 − c0 )
α −(Ξ+ α )
.
r∈[s,t]

Equation (3.49) could be established similarly for s replaced by any v ∈ [s, t] in the above l.h.s. This in turn yields,
ϑ−1
taking T ≥t − s and c0 small enough s.t. (1 + c0 α )T ζ̄ is also small , up to a modification of C:
ϑ−1  1 2η

(3.50) sup Ψ(v) ≤ C(1 + c0 α ) dη,s,t (P1 , P2 )(t − s)ζ̄ + (t − s)1− α + α sup Φ(v) .
v∈[s,t] v∈[s,t]

The point is now to plug (3.50) into (3.33) to complete the circular argument. Recalling as well that 1−( α1 + αϑ
)+ 2η
α =
η
α − ε, this yields:
 
−α1 η
+α 1
1− α + 2η −Ξ 1− α1
+ 2η ϑ
−(Ξ+ α )
|Dx w(s, x)| ≤ C dη,s,t (P1 , P2 )(t − s) + sup Φ(v)(t − s) α + sup Ψ(v)(t − s) α
(3.33) v∈[s,t] v∈[s,t]
 h i
1 η ϑ−1 η
≤ C dη,s,t (P1 , P2 ) (t − s)− α + α +(1 + c0 α )(t − s)ζ̄+ α −ε−Ξ
(3.50)
h ϑ−1 i
1
1− α + 2η −Ξ 1
(1− α + 2η η
)+ α −ε−Ξ
+ sup Φ(v) (t − s) α + (1 + c0 )(t − s)
α α .
v∈[s,t]

Again the previous equation would also hold for any v ∈ [s, t] instead of s in the previous l.h.s. Thus, from the
definition on Ξ in (3.34) and ζ̄ in (3.43),

ϑ−1
sup Φ(v) := sup (v − s)Ξ |Dx w(s, x)| ≤ C(1 + c0 α ) dη,s,t (P1 , P2 )(t − s)ζ̄
v∈[s,t] v∈[s,t]

1
1− α + 2η
+ sup Φ(v)(t − s) α .
v∈[s,t]

ϑ−1
Taking again T ≥t − s and c0 small enough and s.t. (1 + c0 α )T ζ̄ is sufficiently small , we derive:
ϑ−1
(3.51) sup Φ(v) ≤ C(1 + c0 α )dη,s,t (P1 , P2 )(t − s)ζ̄ .
v∈[s,t]

Plugging first (3.51) into (3.50), we get:

ϑ−1
(3.52) sup Ψ(v) ≤ C(1 + c0 α )2 dη,s,t (P1 , P2 )(t − s)ζ̄ .
v∈[s,t]
18 NOUFEL FRIKHA, VALENTIN KONAKOV, AND STÉPHANE MENOZZI

It now remains to plug (3.51) and (3.52) into (3.27)-(3.37)


|w(s, x)| ≤ kwk∞ ≤ C dη,s,t (P1 , P2 )(t − s)ε
Z t
2η
 ϑ

+ dv(v − s) α (t − v)−Ξ Φ(v) + (t − v)−(Ξ+ α ) Ψ(v)
s
ϑ−1
(3.53) ≤ C(1 + c0 α )2 dη,s,t (P1 , P2 )(t − s)ε .
Observe indeed carefully that, even though the time contribution in front of the distance in (3.51) and (3.49) appears
ε
at a coarser rate, namely (t − s)ζ̄ = (t − s) 2 , since we take out the supremum of Φ, Ψ and integrate once again in
time in the above first inequality, we eventually get a final contribution in (t−s)ε for the distance (from the previous
definitions of Ξ, ϑ in (3.34) and (3.5) respectively). This also illustrates the intuitive fact that the remainder terms
in the perturbative expansion, those associated with Φ, Ψ, yield somehow negligible contributions.
Equation (3.53) precisely gives the expected bound (3.22) (with ζ̃ = ε) and completes the proof of Proposition
2.1 in the case α ≥ 1.

The case α < 1. Restarting from (3.27), we write:

 Z t 
1 η 2η
1−( α −α )
(3.54) |w(s, x)| ≤ C dη,s,t (P1 , P2 )(t − s) + (t − s)kwk∞ + dv (v − s) kDw(v, ·)k∞
α

so that
 Z t 
1 η 2η
(3.55) kwk∞ ≤ C dη,s,t (P1 , P2 )(t − s)1−( α − α ) + dv (v − s) α kDw(v, ·)k∞ ,
s

assuming that T ≥ t − s small enough. Observe first that, for the first term of the above r.h.s., we have assumed
that ( α1 − αη ) < 1 ⇐⇒ α + η > 1. This indeed allows to derive the smoothing effect for the contribution HP1 ,P2 uδ2 .
Let us now proceed from the cancellation argument as in (3.28). We get from equations (3.6) and (3.7) with
λ = 1/2 (recalling that γ = η) and (3.24) that in the current case:
h η


1
δ δ η − α −α
|HP1 ,P2 u2 (v, y) − HP1 ,P2 u2 (v, θv,τ (ξ))| ≤ Cdη (P1 (v), P2 (v)) |y − θv,τ (ξ)| (t − v)
i
+|y − θv,τ (ξ)|ϑ (t − v)−1+ε
(3.56) ≤ Cdη (P1 (v), P2 (v))|y − θv,τ (ξ)|α−1+η−αε (t − v)−1+ε ,
h
i
as soon as ε ≤ 1 − ( α1 − αη ), recalling from (3.5) that ϑ = α 1 − ε − α1 − αη and α < 1 for the last inequality.
Now, similarly to (3.33), using (3.56) and again Lemma 3.4, we also get:
Z t 1 1 η
|Dx w(s, x)| ≤ C dv (v − s)− α +1− α + α −ε dη (P1 (v), P2 (v))(t − v)−1+ε
s
Z t
1 2η

+ dv (v − s)− α + α kwk∞ + kDw(v, ·)k∞
s
 Z t 
2 η 1 2η
(3.57) ≤ C dη,s,t (P1 , P2 )(t − s)1−( α − α ) + dv (v − s)− α + α kDw(v, ·)k∞ ,
(3.55) s

provided that
2 η η
(3.58) + > 0 ⇐⇒ α > 1 −
2−
α α 2
in order to obtain integrable time singularities in the first integral (taking as well ε small enough). Recalling that
η ∈ (0, 1/2], and that we have also assumed

that α > 2η for integrability purposes, this means that α > 2η ∨ (1 − η2 ).
η
With the notation (3.34) for Φ(v) v∈[s,t] , taking in our current supercritical case Ξ = α1 − 2α < 1 (from (3.58)),
we derive from (3.55) and (3.57):
 Z t 
1 η 2η
1−( α −α ) −Ξ
|w(s, x)| ≤ kwk∞ ≤ C dη,s,t (P1 , P2 )(t − s) + sup Φ(v) dv (v − s) α (t − v)
v∈[s,t] s
 
1 η η
(3.59) ≤ C dη,s,t (P1 , P2 )(t − s)1−( α − α ) + sup Φ(v)(t − s)1+2 α −Ξ ,
v∈[s,t]
 WELL-POSEDNESS OF SOME NON-LINEAR STABLE DRIVEN SDES 19

and
 Z t 
2 η 1 2η
1−( α −α )
|Dx w(s, x)| ≤ C dη,s,t (P1 , P2 )(t − s) + sup Φ(v) dv (v − s)− α + α (t − v)−Ξ
v∈[s,t] s
 2 η 1 2η

1−( α −α )
≤ C dη,s,t (P1 , P2 )(t − s) + sup Φ(v)(t − s)1−( α − α )−Ξ ,
v∈[s,t]

which in turn gives:

 2 η 1 2η

(t − s)Ξ |Dx w(s, x)| ≤ C dη,s,t (P1 , P2 )(t − s)1−( α − α )+Ξ + sup Φ(v)(t − s)1−( α − α ) .
v∈[s,t]

Observe now that for the previous choice of Ξ and from (3.58), 1 − ( α2 − αη ) + Ξ = 1 − ( α1 − 2α
η
) > 0, so that we indeed
have a regularizing effect. Furthermore, the above equation would still be valid with s replaced by any ṽ ∈ [s, t] for
the l.h.s. This yields that for all ṽ ∈ [s, t]
 1 η 1 2η

Φ(ṽ) ≤ C dη,s,t (P1 , P2 )(t − s)1−( α − 2α ) + sup Φ(v)(t − s)1−( α − α )
v∈[s,t]

which yields
1 η
(3.60) sup Φ(v) ≤ Cdη,s,t (P1 , P2 )(t − s)1−( α − 2α )
v∈[s,t]

Plugging (3.60) into (3.59) eventually yields:

1 η
kwk∞ ≤ Cdη,s,t (P1 , P2 )(t − s)1−( α − 2α ) .
This concludes the proof of Proposition 2.1 in the super-critical case under the condition α > 2η ∨ (1 − η2 ). This
naturally gives the constraint on the stability index α ∈ (4/5, 1). As already mentioned, in Appendix B, we remove
this assumption and establish our main result, namely, Theorem 2.3 under the sole condition α > 2η ∨ (1 − η).

Appendix A. Additional controls on the density in the stable case

To derive the first main results of Lemma 3.4, the key idea consists in separating the small and the large jumps in
the Fourier transform of the proxy Θs,v,(τ,ξ),µ introduced in (3.15). This trick, which is often known as the Itô-Lévy
decomposition, has been used in many contexts (see e.g. [Wat07], [Szt10] or [HM16] for density estimates; [HMP19]
for Lp bounds related to degenerate stably driven Ornstein-Uhlenbeck processes).
Fix now a final parameter t. Rewrite from (3.15), for any ζ ∈ Rd :
h i  Z t Z +∞ Z  
dr

ϕΘs,t,(τ,ξ),µ (ζ) := E exp(ihζ, Θs,t,(τ,ξ),µ i) = exp dv ω(dλ) cos(hζ, σ v, θv,τ (ξ), µ(v) λri) − 1
s 0 r1+α Sd−1
 Z t Z (t−s) α1 Z  
dr

= exp dv ω(dλ) cos(hζ, σ v, θ v,τ (ξ), µ(v) λri) − 1
s 0 r1+α Sd−1
 Z t Z +∞ Z  
dr

× exp dv 1+α
ω(dλ) cos(hζ, σ v, θ v,τ (ξ), µ(v) λri) − 1
(t−s) α r
1
s Sd−1

(A.1) =: pbMs,t,(τ,ξ),µ (ζ)PbNs,t,(τ,ξ),µ (ζ).

It precisely allows to write: Θs,t,(τ,ξ),µ := Ms,t,(τ,ξ),µ + Ns,t,(τ,ξ),µ where Ms,t,(τ,ξ),µ and Ns,t,(τ,ξ),µ are independent
random variables corresponding respectively to the small and large jumps part of the stochastic integral Θs,t,(τ,ξ),µ .
The density of Θs,t,(τ,ξ),µ writes for all y ∈ Rd as
Z
(A.2) pΘs,t,(τ,ξ),µ (y) = pMs,t,(τ,ξ),µ (y − z)PNs,t,(τ,ξ),µ (dz).
Rd

It is known, see e.g. Lemma B.1 in [HM16], Lemma A.2 in [HMP19], that the density pMs,t,(τ,ξ),µ associated with
the small jumps of the stochastic integral is smooth and has a polynomial decay of arbitrary order (i.e. it belongs
to the Schwartz class).On the other hand, PNs,t,(τ,ξ),µ is a Poisson measure s.t. for all β ∈ (0, α), there exists
R β
Cβ,α ≥ 1 s.t. E[|Ns,t,(τ,ξ),µ |β ] = Rd |z|β PNs,t,(τ,ξ),µ (dz) ≤ Cα,β (t − s) α . Importantly, both components pMs,t,(τ,ξ),µ
and PNs,t,(τ,ξ),µ (when the jump measure ν is absolutely continuous) can be dominated uniformly w.r.t. the freezing
parameters (τ, ξ), µ. For the sake of clarity, and since these steps are also crucial to study further the sensitivities
of the quantities pMs,t,(τ,ξ),µ , PNs,t,(τ,ξ),µ w.r.t. the measure argument, we prove these facts in the next subsection.
20 NOUFEL FRIKHA, VALENTIN KONAKOV, AND STÉPHANE MENOZZI

A.1. First controls for the two parts of the frozen density. We first give here some useful lemmas con-
cerning the behavior of the laws of the independent random variables Ms,t,(τ,ξ),µ and Ns,t,(τ,ξ),µ s.t. Θs,t,(τ,ξ),µ =
Ms,t,(τ,ξ),µ + Ns,t,(τ,ξ),µ .
Lemma A.1 (Density estimate on the Martingale part and associated derivatives.). For all m ≥ 1, there exists
Cm ≥ 1 s.t. for all 0 ≤ s < t ≤ T, y ∈ Rd , (τ, ξ) ∈ [0, T ] × Rd ,
 −m
Cm |y|
(A.3) pMs,t,(τ,ξ),µ (y) ≤ d 1+ 1 =: C̃m pM̄s,t ,m (y),
(t − s) α (t − s) α
R
where Rd pM̄s,t ,m (y)dy = 1.
Also, for all m ≥ 1 and all multi-index β, |β| ≤ 3,
 −m
β Cm |y| C̃m
|Dy pMs,t,(τ,ξ),µ (y)| ≤ d+|β|
1+ 1 = |β|
pM̄s,t ,m (y).
(t − s) α (t − s) α (t − s) α
Proof. Inverting the Fourier transform in (A.1) write:
Z  Z t Z (t−s) α1 Z  
1 −ihζ,yi dr

pMs,t,(τ,ξ),µ (y) = dζe exp dv ω(dλ) cos(hζ, σ v, θ v,τ (ξ), µ(v) λri) − 1 .
(2π)d Rd s 0 r1+α Sd−1
Setting (t − s)1/α ζ = ζ̃ yields:
Z −ihζ̃, y
i
1 d 1
pMs,t,(τ,ξ),µ (y) = (t − s)− α dζ̃e (t−s) α
(2π)d Rd
Z t Z 1
(t−s) α Z  
dr ζ̃

(A.4) × exp dv ω(dλ) cos(h 1 , σ v, θv,τ (ξ), µ(v) λri) − 1 .
s 0 r1+α Sd−1 (t − s) α
Let us now denote
Z t Z (t−s) α
1
Z  
dr ζ̃

fˆs,t (ζ̃) := exp dv ω(dλ) cos(h 1 , σ v, θv,τ (ξ), µ(v) λri) − 1 .
s 0 r1+α Sd−1 (t − s) α
Since the Lévy measure in the above expression has finite support, we get from Theorem 3.7.13 in Jacob [Jac05]
that fˆs,t is infinitely differentiable as a function of ζ̃. Moreover,
Z t Z (t−s) α1 Z

dr |σ v, θv,τ (ξ), µ(v) λ|r ζ̃

|Dfˆs,t (ζ̃)| ≤ dv 1+α
ω(dλ) 1 sin(h 1 , σ v, θv,τ (ξ), µ(v) λri)
s 0 r Sd−1 (t − s) α (t − s) α
 Z t Z (t−s) α 1
Z  
dr ζ̃

× exp dv ω(dλ) cos(h 1 , σ v, θ v,τ (ξ), µ(v) λri) − 1 .
s 0 r1+α Sd−1 (t − s) α
Write now:
Z t Z (t−s) α1 Z

dr |σ v, θv,τ (ξ), µ(v) λ|r ζ̃

dv 1+α
ω(dλ) 1 sin(h 1 , σ v, θv,τ (ξ), µ(v) λri)
s 0 r Sd−1 (t − s) α (t − s) α
Z
dr r r
≤ C(t − s) 1+α 1 (1α<1 + 1α≥1 |ζ̃| 1 ) ≤ C(1 + |ζ̃|).
r≤(t−s) α r
1
(t − s) α (t − s) α
Thus:
Z t Z !

ζ̃
|Dfˆs,t (ζ̃)| ≤ C(1 + |ζ̃|) exp dv {cos(hσ v, θv,τ (ξ), µ(v) q, 1 i) − 1}ν(dq)
s Rd (t − s) α
1
× exp(2(t − s)ν(B(0, (t − s) α )c )) ≤ C(1 + |ζ̃|) exp(−C −1 |ζ̃|α ), C ≥ 1,
using for the last inequality that from (2.7)
Z t Z !

ζ̃
exp dv {cos(hσ v, θv,τ (ξ), µ(v) q, 1 i) − 1}ν(dq)
s Rd (t − s) α
Z Z !
t
ζ̃ α
= exp −Cα,d dv |hσ v, θv,τ (ξ), µ(v) λ, 1 i| ω(dλ)
s Sd−1 (t − s) α
≤ C exp(−C −1 |ζ̃|α ),
 WELL-POSEDNESS OF SOME NON-LINEAR STABLE DRIVEN SDES 21

and also that, from the decomposition of ν, ν(B(0, (t − s)1/α )c ) ≤ C/(t − s).
Similarly, for all multi-index β of length |β| = l ∈ N:

|Dβ fˆs,t (ζ̃)| ≤ Cl (1 + |ζ̃|l ) exp(−C −1 |ζ̃|α ), Cl ≥ 1.

Thus, fˆs,t belongs the Schwartz space. Denoting by fs,t its inverse Fourier transform, we have:

∀m ≥ 0, ∀y ∈ Rd , ∃Cm ≥ 1 s.t. : |fs,t (y)| ≤ Cm (1 + |y|)−m .

d 1
Now, since pM (t − s, y) = (t − s)− α fs,t (y/(t − s) α ), the announced bound follows. The control concerning the
derivatives is derived similarly.


Lemma A.2 (Controls on the Poisson measure). Let ν be any symmetric stable jump measure satisfying (2.7).
Then, for all β ∈ [0, α), there exists Cα,β ≥ 1 s.t. for all 0 ≤ s < t, z ∈ Rd ,
Z
β
(A.5) |y|β PNs,t,(τ,ξ),µ (dy) ≤ Cα,β (t − s) α .
Rd

Assume now the jump measure ν is absolutely continuous, i.e. Assumption (AC) holds. Then, there exists a
P ρ⋆n
s,t dr
Poisson measure PN̄ρs,t := exp(−1) n≥0 n! where ρs,t (dy) = κ(t − s) r1+α 1 1 ΛSd−1 (dθ), y = rθ, (r, θ) ∈
α |r|≥c(t−s)
R+
∗ ×S
d−1
, with c > 0 depending on the non-degeneracy constants in (2.7), (2.9), is a probability measure and ρ⋆n
s,t
denotes its nth fold convolution, for which it holds that:

(A.6) PNs,t,(τ,ξ),µ ≤ CPN̄ρs,t .

d −1
Proof. Introduce ν̄s,t (dz) = ν(dz)1 1 which is a finite measure and s.t. ν̄s,t (R ) ≤ C(t − s) . With this
|z|≥(t−s) α
notation at hand, write:
 Z t Z +∞ Z  
dr

b
PNs,t,(τ,ξ),µ (ζ) = exp dv ω(dλ) cos(hζ, σ v, θv,τ (ξ), µ(v) λri) − 1 =
1+α
(t−s) α r
1
s Sd−1
Z t

∗
= exp( ν s,t (σ v, θv,τ (ξ), µ(v) ζ) − (t − s)ν̄s,t (Rd ))
dv b̄
s
Z t
= exp( ν s,t,(v,(τ,ξ),µ) (ζ) − (t − s)ν̄s,t (Rd )),
dv b̄
s


defining ν̄s,t,(v,(τ,ξ),µ) (A) := ν̄s,t ({y ∈ Rd : σ v, θv,τ (ξ), µ(v) y ∈ A}) and where νb̄s,t , νb̄s,t,(v,(τ,ξ),µ) denote the
Rt
Fourier-Stieltjes transform of the considered measure. Hence, introducing the measure ζs,t,(τ,ξ),µ := s dvν̄s,t,(v,(τ,ξ),µ)
and expanding the previous exponential and by termwise Fourier inversion, we get:


X ζs,t,(τ,ξ),µ ⋆n
d d
(A.7) PNs,t,(τ,ξ),µ = exp(ζs,t,(τ,ξ),µ − (t − s)ν̄s,t (R )) = exp(−(t − s)ν̄s,t (R )) ,
n!
n≥0

where in the above equation we again used that, for a finite measure ρ on Rd , (ρ)⋆n := ρ ⋆ · · · ⋆ ρ denotes its nth fold
n times
convolution. Observe now from the definition of ν̄s,t and the non-degeneracy conditions on σ (see eq. (2.9) in (UE))
that there exists C ≥ 1 s.t. for all 0 ≤ s < t, z ∈ Rd , µ ∈ C(R+ , P(Rd )), C −1 ≤ ζs,t,(τ,ξ),µ (Rd ) = (t−s)ν̄s,t (Rd ) ≤ C.
ζs,t,(τ,ξ),µ ζs,t,(τ,ξ),µ
Introducing the normalized (probability) measure ζ̃s,t,(τ,ξ),µ := ζs,t,(τ,ξ),µ (Rd )
= (t−s)ν̄s,t (Rd )
, one then rewrites:
Z ∞
n Z
X (t − s)ν̄s,t (Rd )
β d
|y| PNs,t,(τ,ξ),µ (dy) ≤ exp(−(t − s)ν̄s,t (R )) |y|β (ζ̃s,t,(τ,ξ),µ )⋆n (dy).
Rd n=0
n! R d

Introducing now on a some probability space (Ω, F , P) a sequence (Xi )i≥0 of i.i.d. random variables with law
ζ̃s,t,(τ,ξ),µ , we can rewrite from the above control:
Z X∞ Xn ∞
Cn −1 X C
−1
n
(A.8) |y|β PNs,t,(τ,ξ),µ (dy) ≤ e−C E[| Xi |β ] ≤ e−C (n + 1)1∨β E[|X1 |β ].
Rd n=0
n! i=0 n=0
n!
22 NOUFEL FRIKHA, VALENTIN KONAKOV, AND STÉPHANE MENOZZI

Now,
Z Z t Z
β β
E[|X1 | ] = |y| ζ̃s,t,(τ,ξ),µ (dy) ≤ C dv |y|β ν̄s,t,(v,(τ,ξ),µ) (dy)
Rd s Rd
Z t Z
= C dv 1
|σ(v, θv,τ (ξ), µ(v))y|β ν(dy)
s |y|≥(t−s) α
Z t Z Z Z t Z
dr dr
≤ C dv ω(dλ)|σ(v, θv,τ (ξ), µ(v))λr|β ≤ C dv
s
1
r≥(t−s) α r1+α Sd−1 s
1
r≥(t−s) α r1+α−β
β
(A.9) ≤ Cα,β (t − s) , α

recalling that α > β for the last inequality. Observe as well that the last constant Cα,β depends on α and β and
explodes when β ↑ α. Plugging the above bound into (A.8) we derive the first statement (A.5) without any specific
assumption on the jump measure.
R t Assume now that ν is absolutely continuous. Then, still with the dprevious notations, we can bound the measure
s dvν̄s,t,(v,(τ,ξ),µ) uniformly in (v, (τ, ξ), µ). Namely, for all A ∈ B(R ) and with the notations of assumption (AC):
Z t  Z t


Z t


dvν̄s,t,(v,(τ,ξ),µ) (A) = dv ν̄s,t,(v,(τ,ξ),µ) (A) = dv ν̄s,t ({y ∈ Rd : σ(v, θv,τ (ξ), µ(v))y ∈ A})
s s s
Z t Z
= dv 1
1σ(v,θv,τ (ξ),µ(v))y∈A ν(dy)
s |y|≥(t−s) α
Z t Z
f (σ(v, θv,τ (ξ), µ(v))−1 ỹ) dỹ
= dv 1
1 ỹ∈A −1 d+α
s |σ(v,θv,τ (ξ),µ(v))−1 ỹ|≥(t−s) α |σ(v, θv,τ (ξ), µ(v)) ỹ| det(σ(v, θv,τ (ξ), µ(v)))
Z t Z  σ(v, θ (ξ), µ(v))−1 ỹ  dỹ
v,τ
≤ C dv 1ỹ∈A g −1 ỹ| |ỹ|d+α
.
s
1
|ỹ|≥c(t−s) α |σ(v, θ v,τ (ξ), µ(v))

Since g is bounded on Sd−1 (recall g is Lipschitz continuous on Sd−1 ), writing ỹ = rθ, (r, θ) ∈ R+
∗ ×S
d−1
, we derive
dr
that there exists a constant C and a probability measure ρs,t (dy) = κ(t − s) r1+α 1 1 ΛSd−1 (dθ), s.t. for all
α |r|≥c(t−s)
(τ, ξ, µ) ∈ [0, T ] × Rd × C(R+ , P(Rd )), A ∈ B(Rd ):
Z t  Z
dr
Z
dvν̄s,t,(v,(τ,ξ),µ) (A) ≤ C(t − s) ΛSd−1 (dθ)1rθ∈A = Cρs,t (A),
s
1
|r|≥c(t−s) α r1+α Sd−1

up to a modification of C. The above equation and (A.7) eventually give (A.6).


Remark A.1 (Integrability of the stochastic integral). As a direct corollary of Lemma A.1 and A.2 we derive that
for any non-degenerate jump measure ν satisfying (2.7) and all β ∈ [0, α) there exists Cβ ≥ 1 (increasing with β)
s.t. for all 0 ≤ s < t, z ∈ Rd , µ ∈ C(R+ , P(Rd )):
Z
β
(A.10) |y|β pΘs,t,(τ,ξ),µ (y)dy ≤ Cβ (t − s) α .
Rd

More generally,
R under (AC), Lemma 3.4 readily follows from Lemmas A.1 and A.2 setting for all z ∈ Rd ,
q̄(t − s, z) := Rd pM̄s,t ,m (z − ζ̄)PN̄ρs,t (dζ̃), for any arbitrary integer m ≥ d + 1.

A.2. Proof of Lemma 3.1. Quantitative bounds for the solution uδ1 of the PDE (3.2). To derive the
announced estimates, we rewrite the solution uδi ∈ C α+η (Rd , R) of (3.2) as follows:
  
Pi ,α,(τ,ξ)

 ∂v + L̃v uδi (v, x) + b(v, θv,τ (ξ), Pi (v)) · Dx uδi (v, x)
(A.11) = −RPi
α,(τ,ξ)
(v, x), (v, x) ∈ [0, t) × Rd ,

 δ
ui (t, x) = h (x), on Rd ,
δ

α,(τ,ξ)
with the notations of (3.10) and (3.11) (replacing w by uδi in RPi ), i.e.
Z
i ,α,(τ,ξ)


L̃P
v ϕ(x) := p.v. ϕ(x + σ(v, θv,τ (ξ), Pi (v))ζ) − ϕ(x) ν(dζ)
Rd
and
 
α,(τ,ξ)
RPi (v, x) := b(v, x, Pi (v)) − b(v, θv,τ (ξ), Pi (v)) · Duδi (v, x) + LP
s
i ,α
− L̃P
s
i ,α,(τ,ξ)
uδi (v, x).
 WELL-POSEDNESS OF SOME NON-LINEAR STABLE DRIVEN SDES 23

The PDE (A.11) suggests to rewrite uδi through a Duhamel type expansion (which has been fully justified in
Proposition 8 of [CdRMP19]). This yields for all (s, x) ∈ [0, t] × Rd ,
Z t
δ Pi ,α,(τ,ξ) δ Pi ,α,(τ,ξ) α,(τ,ξ)


ui (s, x) = P̃s,t h (x) + dv P̃s,v RPi (v, ·) (x)
s
Z Z t Z
Pi ,α,(τ,ξ) δ α,(τ,ξ)
(A.12) = p̃ (s, t, x, y)h (y)dy + dv p̃Pi ,α,(τ,ξ) (s, v, x, y)RPi (v, y)dy.
Rd s Rd

Now, from (3.11) and (3.20), similarly to (3.25)-(3.26), choosing (τ, ξ) = (s, x) and exploiting (3.16) and Lemma
3.4, we derive:
Z t Z
α,(τ,ξ)
dv p̃Pi ,α,(τ,ξ) (s, v, x, y)RPi (v, y)dy
s Rd
(τ,ξ)=(s,x)
Z t  
2η
(A.13) ≤C dv(v − s) α kuδi k∞ 1{α≤1} + kDuδi (v, ·)k∞ + kDuδi (v, ·)kĊ ϑ 1{α≥1} .
s

Thus, recalling from the definition of hδ = h ⋆ ρδ (for a smooth mollifying kernel ρδ ) that khδ k∞ ≤ khk∞ , we
get:
 Z t  
2η
δ δ δ δ
(A.14) |ui (s, x)| ≤ C khk∞ + (t − s)kui k∞ 1{α≤1} + dv(v − s) α kDui (v, ·)k∞ + kDui (v, ·)kĊ ϑ 1{α≥1} .
s

From (A.12), we can differentiate w.r.t x and also use a cancellation technique for the final condition. Namely,
(A.15)
Z Z t Z
δ Pi ,α,(τ,ξ) δ δ α,(τ,ξ)
Dx ui (s, x) = Dx p̃ (s, t, x, y)[h (y) − h (θt,τ (ξ))]dy + dv Dx p̃Pi ,α,(τ,ξ) (s, v, x, y)RPi (v, y)dy.
Rd s Rd
δ
Hence, from (A.15), recalling that h is γ-Hölder continuous in space uniformly in the regularization parameter δ,
with in particular khδ kĊ γ ≤ khkĊ γ , similarly to (A.14), we get:

(A.16) |Dx uδi (s, x)|

 khk Z


≤C Ċ γ
1 dy|y − θv,τ (ξ)|γ q̄ v − s, y − m(τ,ξ)
v,s (x)
(t − s) α Rd
Z t Z 
dv 
+ 1 dy kb(v, ·, Pi (v))kC 2η |Dy uδi (v, y)| + kσ(v, ·, Pi (v))kC 2η (kuδi k∞ + kDx uδi (v, ·)k∞ )1{α<1}
s (v − s) α Rd


+kDx uδi (v, ·)kC ϑ 1{α>1} + kuδi (v, ·)kC 1+ϑ 1{α=1} |y − θv,τ (ξ)|2η q̄ v − s, y − m(τ,ξ)
v,s (x) .

Taking (τ, ξ) = (s, x), it follows again from (3.16) and Lemma 3.4 that:
(A.17)|Dx uδi (s, x)|
 Z t  
1
−α γ
+α 1
−α + 2η δ δ δ
≤ C khkĊ γ (t − s) + dv(v − s) α kui (v, ·)k∞ 1{α≤1} + kDx ui (v, ·)k∞ + kDx ui (v, ·)kĊ ϑ 1{α≥1} .
s

We now use the notations of (3.35) and (3.34) with the exponents modified, namely we consider here
1 γ
(A.18) Φ(v) := (t − v)Ξ sup |Dx uδi (v, x)|, Ξ := − , v ∈ [s, t],
x∈Rd α α
and
ϑ
(A.19) Ψ(v) := (t − v)Ξ+ α kDx uδi (v, ·)kC ϑ , v ∈ [s, t].

ϑ 1 γ 1 γ
(A.20) Ξ+ = − + 1 − ε − ( − ) = 1 − ε < 1.
α α α α α
In particular, with these notations at hand, since t − s is small, it follows from (A.14) that
 Z t  
2η ϑ
−Ξ −(Ξ+ )
(A.21) kui k∞ ≤ C khk∞ + dv(v − s) α (t − v) Φ(v) + (t − v) α Ψ(v) .
s
24 NOUFEL FRIKHA, VALENTIN KONAKOV, AND STÉPHANE MENOZZI

As above, we distinguish the diagonal and off-diagonal regimes w.r.t. the current considered times. In the
off-diagonal regime holds, we readily get from (A.18):

ϑ |Dx uδi (s, x) − Dx uδi (s, x′ )|

(t − s) α ≤ |Dx uδi (s, x) − Dx uδi (s, x′ )| ≤ |Dx uδi (s, x)| + |Dx uδi (s, x′ )|
|x − x′ |ϑ

1 γ
≤ C khkĊ γ (t − s)− α + α
Z t
1 2η
 ϑ

+ dv(v − s)− α + α kuδi k∞ 1{α≤1} + (t − v)−Ξ Φ(v) + (t − v)−(Ξ+ α ) Ψ(v) ,
s

ϑ
Ξ+ α |Dx uδi (s, x) − Dx uδi (s, x′ )| 1 2η
(t − s) ≤ C khkĊ γ + (t − s)1− α + α sup Φ(v)
|x − x′ |ϑ v∈[s,t]

1
1−( α ϑ
+α )+ 2η
(A.22) +(t − s) α sup Ψ(v) ,
v∈[s,t]

using as well (A.21) for the last inequality.

1
Let us now turn to the global diagonal regime: |x − x′ | ≤ (t − s) α . We again need to split as above the analysis
in function of the considered running time introducing a change of freezing point. Namely, from Proposition 12 in
[CdRMP19], we derive:

Dx uδi (s, x) − Dx uδi (s, x′ )

Z t Z
α,(τ,ξ)
= dv1v≤τ0 Dx p̃Pi ,α,(τ,ξ) (s, v, x, y)RPi (v, y)dy
s Rd
Z t Z 
′ α,(τ,ξ ′ )
− dv1v≤τ0 Dx p̃Pi ,α,(τ,ξ ) (s, v, x′ , y)RPi (v, y)dy
Rd (τ0 ,τ,ξ,ξ ′ )=(t0 ,s,x,x′ )
s ′
′

Pi α,(τ,ξ ) δ ′ Pi ,α,(τ,ξ̃ ) δ ′
− Dx P̃s,τ 0
u i (τ0 , x ) (τ0 ,τ,ξ ′ )=(t0 ,s,x′ )
− D x P̃s,τ0 u i (τ 0 , x ) (τ0 ,τ,ξ̃ ′ )=(t0 ,s,x)
Z

+ dyhδ (y) Dx p̃Pi ,α,(τ,ξ) (s, t, x, y) − Dx p̃Pi ,α,(τ,ξ) (s, t, x′ , y)
Rd
Z t Z 

α,(τ,ξ)
+ dv1v>τ0 Dx p̃Pi ,α,(τ,ξ) (s, v, x, y) − Dx p̃Pi ,α,(τ,ξ) (s, v, x′ , y) RPi (v, y)dy
s Rd (τ0 ,τ,ξ)=(t0 ,s,x)

=: ∆ui1 (s, x, x′ )
δ
+ ∆uδi2 (s, x, x′ ) + ∆uδi3 (s, x, x′ ).
(A.23)

The term ∆uδ11 (s, x, x′ ), corresponding to the local off-diagonal regime within the global diagonal one, can be
analyzed as above (see the analysis of the contribution ∆w1 (s, x, x′ ) introduced in (3.41)). This yields:
Z τ0 Z
δ ′ α,(τ,ξ)
|∆ui1 (s, x, x )| ≤ dv Dx p̃Pi ,α,(τ,ξ) (s, v, x, y)RPi (v, y)dy
s Rd (τ0 ,τ,ξ)=(t0 ,s,x)
Z τ0 Z
′ ′
α,(τ,ξ )
+ dv Dx p̃Pi ,α,(τ,ξ ) (s, v, x′ , y)RPi (v, y)dy
s Rd (τ0 ,τ,ξ ′ )=(t0 ,s,x′ )
Z t0 
1 2η  ϑ  |x − x′ |ϑ
≤ C dv(v − s)− α + α (t − v)−Ξ Φ(v) + (t − v)−(Ξ+ α ) Ψ(v) ϑ .
s (v − s) α

For v ∈ [s, t0 ], (t − v) ≥ (1 − c0 )(t − s) and therefore:

 
|∆uδi1 (s, x, x′ )| 1
−Ξ+1− α + 2η ϑ
−α ϑ
−(Ξ+ α 1
)+1− α + 2η ϑ
−α
≤ C (t − s) α sup Φ(v) + (t − s) α sup Ψ(v) ,
|x − x′ |ϑ v∈[s,t] v∈[s,t]

recalling for the last inequality that − α1 + 2η ϑ η 1 ϑ
α − α = 2 α −( α + α ) =
2η
α − 1 1 γ
α +1−ε−( α − α ) = −1+ 2η γ
α − α +ε > −1.
This eventually gives:

ϑ |∆uδi1 (s, x, x′ )|  1 2η 1 ϑ 2η

(A.24) (t − s)Ξ+ α ≤ C (t − s)1− α + α sup Φ(v) + (t − s)1−( α + α )+ α sup Ψ(v) .
|x − x′ |ϑ v∈[s,t] v∈[s,t]
 WELL-POSEDNESS OF SOME NON-LINEAR STABLE DRIVEN SDES 25

Turning now to ∆uδ13 in (A.23), expanding the frozen densities, exploiting as well (3.19) which gives that a
diagonal perturbation of the density does not affect the related estimates, we write:

|∆uδi3 (s, x, x′ )|
Z 1 Z
≤ dλ Dx2 p̃Pi ,α,(τ,ξ) (s, t, x + λ(x′ − x), y)(x′ − x)[hδ (y) − hδ (mτ,ξ ′
s,t (x + λ(x − x))]dy
0 Rd (τ0 ,τ,ξ)=(t0 ,s,x)
Z t Z 1 Z
α,(τ,ξ)
+ dv dλ Dx2 p̃Pi ,α,(τ,ξ) (s, v, x + λ(x′ − x), y)(x′ − x) × RPi (v, y)dy
τ0 0 Rd (τ0 ,τ,ξ)=(t0 ,s,x)

2 γ
≤ C khkĊ γ |x − x′ |(t − s)−( α − α )
Z t   
2 2η ϑ
+|x − x′ |ϑ dv(v − s)− α + α (t − v)−Ξ Φ(v) + (t − v)−(Ξ+ α ) Ψ(v) |x − x′ |1−ϑ .
t0

α1
Recalling that, for v ∈ [t0 , t], |x − x′ | ≤ (v − s)/c0 , we derive:

γ 1 ϑ
|∆uδi3 (s, x, x′ )| ≤ C khkĊ γ |x − x′ |ϑ (t − s) α −( α + α )
Z t 
1 ϑ 2η  ϑ 
+|x − x′ |ϑ dv(v − s)− α − α + α (t − v)−Ξ Φ(v) + (t − v)−(Ξ+ α ) Ψ(v) .
t0

ϑ 1−γ+ϑ
Therefore, recalling from (A.18)-(A.19) that Ξ + α = α , we get:
 
ϑ |∆uδi3 (s, x, x′ )| 1
+ 2η 1 ϑ
)+ 2η
(A.25) (t − s)Ξ+ α ≤ C khk Ċ γ + (t − s)1− α α sup Φ(v) + (t − s)1−( α +α α sup Ψ(v) .
|x − x′ |ϑ v∈[s,t] v∈[s,t]

The remaining term ∆uδi2 in (A.23) can be handled exactly as the previous ∆w2 introduced in (3.41), replacing w
by uδi . This eventually gives:

ϑ |∆uδi2 (s, x, x′ )| ϑ ϑ
(A.26) (t − s)Ξ+ α ≤ Cc0α (1 − c0 )−(Ξ+ α ) Ψ(t0 ).
|x − x′ |ϑ
1
Plugging (A.24), (A.25) and (A.26) into (A.23), we get that in the diagonal case |x − x′ | ≤ (t − s) α :

ϑ |Dx uδi (s, x) − Duδi (s, x′ )|  1 2η

(t − s)Ξ+ α ≤ C khkĊ γ + (t − s)1− α + α sup Φ(v)
|x − x′ |ϑ v∈[s,t]
ϑ

1+ 2η 1 ϑ
α −( α + α )
ϑ
(A.27) +(t − s) sup Ψ(v) + c0α (1 − c0 )−(Ξ+ α ) Ψ(t0 ) .
v∈[s,t]

Putting together (A.27) and (A.22), we can conclude the proof of Lemma 3.1 following exactly the procedure for
w described in Section 3.1 (using as well (A.14) and (A.18)). This yields the controls in equations (3.4) and (3.5).
The proof is now complete. 

A.3. Proof of Lemma 3.2. We first analyze the sensitivities of the generators w.r.t. the measure arguments or
w.r.t to both the spatial points and the measure arguments. The results of Lemma 3.2, i.e. equations (3.6) and
(3.7) then readily follow from Lemma A.3 and the a priori bounds of Lemma 3.1.

Lemma A.3 (Sensitivity analysis of the generators w.r.t. the measure and the spatial parameters). Let ϕ be a
function in C 1+ϑ (Rd , R) with ϑ > α − 1 if α ≥ 1. Under (AS ) it then holds that, there exists a positive constant
C := C((AS )) s.t. for all v ∈ [s, t]:

(A.28) |(b(v, x, P1 (v)) − b(v, x, P2 (v))) · Dϕ(x)| ≤ Cd2η,s,t (P1 , P2 )kDϕk∞ ,

(A.29) (LvP1 ,α − LvP2 ,α )ϕ(x) ≤ Cd2η,s,t (P1 , P2 ) (kDϕk∞ + kϕk∞ )1{α<1} + kDϕkĊ ϑ 1{α>1} + kϕkC 1+ϑ 1{α=1} .
26 NOUFEL FRIKHA, VALENTIN KONAKOV, AND STÉPHANE MENOZZI

Furthermore, the following estimates also hold for the differences: there exists a positive constant C := C((AS ))
such that for any λ ∈ [0, 1]

(b(v, x, P1 (v)) − b(v, x, P2 (v))) · Dϕ(x) − (b(v, x′ , P1 (v)) − b(s, x′ , P2 (v))) · Dϕ(x′ )


(A.30) ≤ C d(1−λ)2η,s,t (P1 , P2 )|x − x′ |λ2η kDϕk∞ + d2η,s,t (P1 , P2 )|x − x′ |ϑ kDϕkĊ ϑ ,
(LvP1 ,α − LvP2 ,α )ϕ(x) − (LvP1 ,α − LvP2 ,α )ϕ(x′ ))


≤ C d(1−λ)2η,s,t (P1 , P2 )|x − x′ |λ2η (kDϕk∞ + kϕk∞ )1{α<1} + kDϕkĊ ϑ 1{α>1} + kϕkC 1+ϑ 1{α=1}

(A.31) +d2η,s,t (P1 , P2 )|x − x′ |1+ϑ−α kDϕkĊ ϑ .

Proof. Write first using (BH ) and (2.11):

(b(v, x, P1 (v)) − b(v, x, P2 (v))) · Dϕ(x) ≤ Cd2η,s,t (P1 , P2 )kDϕk∞ .

This gives (A.28).

Write now
(LvP1 ,α − LvP2 ,α )ϕ(x)
Z Z
   
= p.v. ϕ(x + σ(v, x, P1 (v))z) − ϕ(x) ν(dz) − p.v. ϕ(x + σ(v, x, P2 (v))z) − ϕ(x) ν(dz)
Rd Rd
v,P1 −1 v,P2 −1
Z Z (σx ) θ (σx ) θ
+∞
dr   g( v,P1 −1
|(σx ) θ|
) g( v,P2 −1
|(σx ) θ|
) 
= p.v. ΛSd−1 (dθ) ϕ(x + rθ) − ϕ(x) v,P1

−
,
0 r1+α Sd−1 |(σxv,P1 )−1 θ|d+α det σx |(σxv,P2 )−1 θ|d+α det σx v,P2

(A.32)
where we used the absolute continuity condition on the Lévy measure, assumption (AC), and the specific structure
(2.8) of the density w.r.t. the Lebesgue measure. For simplicity we have also denoted in the above equation for
i ∈ {1, 2}:
(A.33) σxv,Pi := σ(v, x, Pi (v)).
From now on and for notational convenience, we introduce the following map:
−1
(σ(v,x,m)) θ
g( |(σ(v,x,m)) −1 θ| )
d

(A.34) P(R ) ∋ m 7→ D(v, x, θ, m) :=
|(σ(v, x, m))−1 θ|d+α det σ(v, x, m)
so that, with our notations
δD(v, x, θ, P1 (v), P2 (v)) = D(v, x, θ, P1 (v)) − D(v, x, θ, P2 (v))
v,P1 −1 v,P2 −1
(σx ) θ (σx ) θ
 g( v,P ) g( v,P ) 
|(σx 1 )−1 θ| |(σx 2 )−1 θ|
(A.35) = v,P1

−
.
|(σxv,P1 )−1 θ|d+α det σx |(σxv,P2 )−1 θ|d+α det σx v,P2

Under (DH ) and (AC), using the Lipschitz regularity of g and (2.11) with β = 0 applied to the diffusion coefficient
σ, it is readily seen that:
(A.36) |δD(v, x, θ, P1 (v), P2 (v))| ≤ Cd2η,s,t (P1 , P2 ).
Equation (A.29) then readily follows from (A.32)-(A.36) by usual Taylor expansions and the symmetry condition
which allows to use cancellation techniques for the small jumps when α ≥ 1.

Let us now turn to the differences. Write first,

(b(v, x, P1 (v)) − b(v, x, P2 (v))) · Dϕ(x) − (b(v, x′ , P1 (v)) − b(s, x′ , P2 (v))) · Dϕ(x′ )
≤ |b(v, x, P1 (v)) − b(v, x, P2 (v))||Dϕ(x) − Dϕ(x′ )|
+ (b(v, x, P1 (v)) − b(v, x, P2 (v))) − (b(v, x′ , P1 (v)) − b(v, x′ , P2 (v))) |Dϕ(x′ )|
≤ C(d2η,s,t (P1 , P2 )|x − x′ |ϑ kDϕkĊ ϑ + d(1−λ)2η,s,t (P1 , P2 )|x − x′ |λ2η kDϕk∞ ),
 WELL-POSEDNESS OF SOME NON-LINEAR STABLE DRIVEN SDES 27

using (BH ), (2.11) and (2.12) applied to the drift coefficient b, with β = 0 and β = λ respectively, for the last
inequality. This gives (A.30).

Let us now turn to the difference of the non-local operators. From (A.32) we get with the notation of (A.34):

(LvP1 ,α − LPv
2 ,α
)ϕ(x) − (LvP1 ,α − LvP2 ,α )ϕ(x′ ))
Z +∞ Z
dr  
≤ p.v. 1+α
ΛSd−1 (dθ) ϕ(x + rθ) − ϕ(x) δD(v, x, θ, P1 (v), P2 (v))
0 r Sd−1
Z +∞ Z
dr  
−p.v. 1+α
ΛSd−1 (dθ) ϕ(x′ + rθ) − ϕ(x′ ) δD(v, x′ , θ, P1 (v), P2 (v))
0 r Sd−1
Z +∞ Z
dr  
≤ p.v. 1+α
ΛSd−1 (dθ) ϕ(x + rθ) − ϕ(x) [δD(v, x, θ, P1 (v), P2 (v)) − δD(v, x′ , θ, P1 (v), P2 (v))]
0 r Sd−1
Z +∞ Z
dr  
+ p.v. 1+α
ΛSd−1 (dθ) ϕ(x′ + rθ) − ϕ(x′ ) − (ϕ(x + rθ) − ϕ(x)) δD(v, x′ , θ, P1 (v), P2 (v))
0 r Sd−1
′
=: (D1 + D2 )(v, x, x , P1 , P2 ).
(A.37)
Under (AC) and (DH ), using the mean-value theorem and the fact that g has a bounded a Lipschitz gradient,
we deduce that the continuous map D(v, x, θ, .) defined by (A.34) admits a bounded and continuous functional
derivative such that (Rd )2 ∋ (x, y) 7→ [δD(v, x, θ, m)/δm](y) is 2η-Hölder uniformly w.r.t. the variables v, θ and m,
so that, similarly to (2.12), it is easily seen that there exists C s.t. for all v ∈ [s, t], θ ∈ Sd−1 and for any λ ∈ [0, 1]
|δD(v, x, θ, P1 (v), P2 (v)) − δD(v, x′ , θ, P1 (v), P2 (v))| ≤ C|x − x′ |λ2η d(1−λ)2η,s,t (P1 , P2 ),
which readily gives:
(A.38)

|D1 (v, x, x′ , P1 , P2 )| ≤ C|x−x′ |λ2η d(1−λ)2η,s,t (P1 , P2 ) (kDϕk∞ +kϕk∞ )1{α<1} +kDϕkĊ ϑ 1{α>1} +kϕkC 1+ϑ 1{α=1} .
Let us turn now to D2 . We will establish:
(A.39) |D2 (v, x, x′ , P1 , P2 )| ≤ Cd2η,s,t (P1 , P2 )kDϕkĊ ϑ |x − x′ |1+ϑ−α .
Write indeed assuming first that α < 1:
|D2 (v, x, x′ , P1 , P2 )|
Z 1 Z Z  
dr ′
≤ dλ 1+α
Λ Sd−1 (dθ) Dϕ(x + λrθ) − Dϕ(x + λrθ) · rθδD(v, x′ , θ, P1 (v), P2 (v))
0 r≤|x−x′ | r Sd−1
Z 1 Z Z
dr
+ dλ 1+α
ΛSd−1 (dθ)
0 r≥|x−x′ | r Sd−1
 
× Dϕ(x′ + rθ + λ(x − x′ )) − Dϕ(x′ + λ(x − x′ )) · (x − x′ )δD(v, x′ , θ, P1 (v), P2 (v))
Z dr
Z
dr 
′ ϑ ϑ ′
≤ Cd2η,s,t (P1 , P2 ) 1+α
kDϕk Ċ ϑ |x − x | r + 1+α
kDϕk Ċ ϑ r |x − x |
r∈(0,|x−x′ |] r r≥|x−x′ | r

≤ Cϑ,α d2η,s,t (P1 , P2 )kDϕkĊ ϑ |x − x′ |1+ϑ−α ,

using (A.36) for the last but one inequality and α > ϑ for the last one (see equation (3.5)).
The only modifications needed for α ≥ 1 concern the small jumps. Indeed, we can introduce the compensator
only up to the threshold |x − x′ |. We are simply led to analyze:
Z 1 Z Z  
dr ′ ′
dλ 1+α
Λ Sd−1 (dθ) [Dϕ(x + λrθ) − Dϕ(x)] − [Dϕ(x + λrθ) − Dϕ(x )] · rθ
0 r∈(0,|x−x′|] r Sd−1

×δD(v, x′ , θ, P1 (v), P2 (v))

Z
dr
≤ Cd2η,s,t (P1 , P2 ) kDϕkĊ ϑ r1+ϑ ≤ Cα,ϑ d2η,s,t (P1 , P2 )kDϕkĊ ϑ |x − x′ |1+ϑ−α .
r∈(0,|x−x′|] r1+α
This indeed gives (A.39). Plugging (A.38) and (A.39) into (A.37) we derive the statement (A.31). 
28 NOUFEL FRIKHA, VALENTIN KONAKOV, AND STÉPHANE MENOZZI

The bound in (3.6) (resp. (3.7)) then readily follows from (A.28), (A.29) (resp. (A.30), (A.31)) and the a priori
bounds (3.4), (3.5) which have been established in Lemma 3.1 proven above.
We conclude this section by discussing a bit why in the current McKean-Vlasov setting the absolute continuity
condition on the spectral measure (Assumption (AC)) seems essential.
Remark A.2 (About the absolute continuity of the spectral measure). The key point is that, when considering
the difference of the non-local operators associated with two different measure arguments it seems difficult to make
explicitly the distance between the considered measures appear. Let us illustrate this fact considering the elementary
Pd
following case. Consider ν(dz) = 12 i=1 (δei + δ−ei ) r1+α
dr
, where the (ei )i∈[[1,d]] correspond to the canonical vectors,
corresponding, up to a suitable normalizing constant to the example of the cylindrical Laplacian. In this case we
cannot do the change of variable of (A.32) and directly write for the difference:

(LvP1 ,α − LvP2 ,α )ϕ(x)

Z
dr  
d
X 1 X +∞  
≤ p.v. ϕ(x + (−1)j σxv,P1 rei ) − ϕ(x) − ϕ(x + (−1)j σxv,P2 rei ) − ϕ(x) .
i=1
2 0 r1+α
j∈{0,1}

(A.40)
As usual, the strategy consists in separating the small and large jumps. Let us first consider the case α < 1. In
that case we write, for a threshold T (possibly depending on (σxv,Pi )i∈{1,2} ):
Z
dr  
d
X 1 X T  
p.v. ϕ(x + (−1)j σxv,P1 rei ) − ϕ(x) − ϕ(x + (−1)j σxv,P2 rei ) − ϕ(x)
i=1
2 0 r1+α
j∈{0,1}
Z Z
d
X 1 X T
dr 1 
= p.v. dλDϕ(x + (−1)j ((1 − λ)σxv,P1 + λσxv,P2 )ei r) · (−1)j (σxv,P1 − σxv,P2 )rei
i=1
2 0 r1+α 0
j∈{0,1}

≤ Cα,T kDϕk∞ d2η (P1 (v), P2 (v)).

Hence, since the required term d2η (P1 (v), P2 (v)) appears this way, it seems natural to consider a threshold T at
a macro scale, and in particular independent of σxv,P2 − σxv,P1 . This choice anyhow yields problems for the large
jumps, i.e. those above the threshold T . Namely,
Z +∞
dr  
Xd
1 X j v,P1
  j v,P2
p.v. ϕ(x + (−1) σx re i ) − ϕ(x) − ϕ(x + (−1) σx re i ) − ϕ(x)
i=1
2 T r1+α
j∈{0,1}
d
X Z +∞
1 X dr
≤ 1+α
kϕkĊ β |σxv,P2 − σxv,P1 |β rβ ,
i=1
2 T r
j∈{0,1}

for any β ∈ (0, 1) s.t. β < α in order to preserve some integrability and where kϕkĊ β again stands for the β-
Hölder modulus of ϕ (homogeneous Hölder norm) which can be estimated from kϕk∞ , kDϕk∞ through elementary
interpolation. Hence, we cannot recover that way the required control with exactly the distance, we only end up
with:
Z +∞
dr  
Xd
1 X j v,P1
  j v,P2
p.v. ϕ(x + (−1) σx re i ) − ϕ(x) − ϕ(x + (−1) σx re i ) − ϕ(x)
i=1
2 T r1+α
j∈{0,1}

≤ Cα,β,T kϕkĊ β dβη (P1 (v), P2 (v)).

One could object that we do not have exploited the full regularity of ϕ ∈ C 1+ϑ (Rd , R). But for the small jumps
this will not a priori improve the dependence in the distance (see also the small jump part for α > 1 below).
If now α > 1 the problems are reversed. Indeed, the large jumps can be readily controlled with the expected
bound. Indeed,
Z +∞
dr  
X d
1 X j v,P1
  j v,P2
p.v. ϕ(x + (−1) σ x re i ) − ϕ(x) − ϕ(x + (−1) σx re i ) − ϕ(x)
i=1
2 T r1+α
j∈{0,1}
d
X Z +∞
1 X dr
= p.v. 1+α
kDϕk∞ |σxv,P2 − σxv,P1 |r ≤ Cα,T kDϕk∞ d2η (P1 (v), P2 (v)).
i=1
2 T r
j∈{0,1}
 WELL-POSEDNESS OF SOME NON-LINEAR STABLE DRIVEN SDES 29

On the other hand, for the small jumps,h we are led to use explicitly the ϑ-Hölder regularity of the gradient Dϕ
1 η

i
(recall from (3.5) that ϑ = ϑ(α, η, ε) = α 1 − ε − α − α ). Write precisely,
Z
dr  
d
X 1 X T  
p.v. ϕ(x + (−1)j σxv,P1 rei ) − ϕ(x) − ϕ(x + (−1)j σxv,P2 rei ) − ϕ(x)
i=1
2 0 r1+α
j∈{0,1}
Z Z
dr 
d
X 1 X j
T 1 
= (−1) p.v. dλ Dϕ(x + λ(−1)j σxv,P1 rei ) · σxv,P1 +Dϕ(x + λ(−1)j · σxv,P1 rei )σxv,P2
i=1
2 0 r1+α 0
j∈{0,1}

−Dϕ(x + λ(−1)j σxv,P2 rei ) · σxv,P2 rei
Z Z
dr 
d
X 1 X j
T 1 
= (−1) p.v. dλ (Dϕ(x + λ(−1)j σxv,P1 rei ) − Dϕ(x)) · (σxv,P1 − σxv,P2 )
i=1
2 0 r1+α 0
j∈{0,1}


+ Dϕ(x + λ(−1)j σxv,P1 rei ) − Dϕ(x + λ(−1)j σxv,P2 rei ) · σxv,P2 rei
Z T  
dr ϑ ϑ
≤ C kDϕk Ċ ϑ r d η (P1 (v), P2 (v)) + d η (P1 (v), P2 (v))
0 rα
 
≤ Cα,ϑ kDϕkĊ ϑ dη (P1 (v), P2 (v)) + dϑη (P1 (v), P2 (v)) ,
where we used a cancellation argument for the first inequality, and exploit as well that α − ϑ = 1 − η + εα < 1 for
ε small enough, for the last one. In any case we see that this is not enough to conclude whereas assumption (AC)
allows to transfer all the sensitivity analysis to the spectral measure (see again the proof of the previous lemma).
A.4. Proof of Lemma 3.3. The stated controls follow reproducing the transfer arguments of the diffusion coeffi-
cient to the density of the Lévy measure in equation (A.32) for the integro-differential part. The gradient term is
controlled directly. Namely, from (3.11),

 
α,(τ,ξ)
|RP1 (v, y)| ≤ | b(v, y, P1 (v)) − b(v, θv,τ (ξ), P1 (v)) · Dw(v, y)| + | LsP1 ,α − L̃sP1 ,α,(τ,ξ) w(v, y)|
Z
2η
 
≤ kb(v, ·, P1 (v))kĊ 2η |y − θv,τ (ξ)| |Dw(v, y)| + p.v. ϕ(y + σ(v, y, P1 (v))z) − ϕ(y) ν(dz)
Rd
Z
 
−p.v. ϕ(y + σ(v, θv,τ (ξ), P1 (v))z) − ϕ(y) ν(dz)
Rd
Z +∞ Z
2η dr  
≤ kb(v, ·, P1 (v))kĊ 2η |y − θv,τ (ξ)| |Dw(v, y)| + p.v. ΛSd−1 (dθ) ϕ(y + rθ) − ϕ(y)
0 r1+α Sd−1
 (σ(v,y,P1 (v)))−1 θ
g( |(σ(v,y,P1 (v)))
−1 θ| )
×

−1
|(σ(v, y, P1 (v))) θ| d+α det σ(v, y, P1 (v))
(σ(v,θ
v,τ (ξ),P (v)))−1 θ
1
g( |(σ(v,θv,τ (ξ),P1 (v)))−1 θ| )
−
.
|(σ(v, θv,τ (ξ), P1 (v)))−1 θ|d+α det σ(v, θv,τ (ξ), P1 (v))
Again the Lipschitz regularity of g, the non-degeneracy and 2η-Hölder spatial continuity of σ and the symmetry of
the Lévy measure yield:
 
α,(τ,ξ)
RP1 (v, y) ≤ C kb(v, ·, P1 (v))kC˙2η |Dy w(v, y)| + kσ(v, ·, P1 (v))kC 2η (kwk∞ + kDx w(v, ·)k∞ )1{α<1}

(A.41) +kDx w(v, ·)kC ϑ 1{α>1} + kw(v, ·)kC 1+ϑ 1{α=1} |y − θv,τ (ξ)|2η ,
which the is required statement.

Appendix B. Theorem 2.3 in the supercritical case under the sole condition α > 2η ∨ (1 − η)
In order to establish Theorem 2.3 under (AS ) and the sole condition α > 2η ∨ (1 − η) in the supercritical case
α < 1, we proceed as follows. Restarting from (3.21) with h ∈ C γ1 , with α + γ1 > 1 and γ1 ∈ (0, 2η] to be specified
later on and employing (3.6) instead of (3.23), similarly to (3.27), we derive:
 Z t 
1 γ 2η
1−( α − α1 )
(B.1) |w(s, x)| ≤ C d2η,s,t (P1 , P2 )(t − s) + (t − s)kwk∞ + dv(v − s) α kDw(v, ·)k∞
s
30 NOUFEL FRIKHA, VALENTIN KONAKOV, AND STÉPHANE MENOZZI

Observe that, for the first term of the above r.h.s., we have assumed that ( α1 − γα1 ) < 1 ⇐⇒ α + γ1 > 1. This
δ
seems to be somehow a necessary condition to derive a smoothing effect for the contribution HP 1 ,P2
.
Let us now proceed from the cancellation argument as in (3.28). We get from (3.7) that in the current case:
γ1


1
|HP1 ,P2 uδ2 (v, y) − HP1 ,P2 uδ2 (v, θv,τ (ξ))| ≤ Cd(1−λ)2η (P1 (v), P2 (v))|y − θv,τ (ξ)|2λη (t − v)− α − α
(B.2) +Cd2η (P1 (v), P2 (v))|y − θv,τ (ξ)|ϑ (t − v)−1+ε
h
i
recalling from (3.5) that 0 < ϑ := α 1 − ε − 1
α − γα1 < α − 1 + γ1 .
Now, similarly to (3.33), using (B.2) on the time interval [s, t+s t+s
2 ] and (3.6) on the time interval [ 2 , t], taking
also λ large enough so that 2λη + α > 1, we get:
Z t+s
2 1 2λη 1−γ1
|Dw(s, x)| ≤ C (v − s)− α + α d(1−λ)2η (P1 (v), P2 (v))(t − v)− α dv
s
Z t+s
2 1 ϑ
+ (v − s)− α + α d2η (P1 (v), P2 (v))(t − v)−1+ε dv
s
Z t
1 1−γ1
+ (v − s)− α d2η (P1 (v), P2 (v))(t − v)− α dv
t+s
2
Z t
1 2η

+ (v − s)− α + α kw(v, .)k∞ + kDw(v, .)k∞ dv
s 2 2λη γ1 2 γ1
≤ C d(1−λ)2η,s,t (P1 , P2 )(t − s)1−( α − α − α ) + d2η,s,t (P1 , P2 )(t − s)1−( α − α )
Z t
1 2η

(B.3) + (v − s)− α + α kwk∞ + kDw(v, ·)k∞ dv ,
s

provided that ϑ is sufficiently large so that

(B.4) ϑ + α > 1.


With the notation of (3.34) for Φ(v) v∈[s,t] , taking in our current supercritical case Ξ = α1 − γα1 + 2ε < 1, we derive
from (B.1) and (B.3):
 Z t 
1 γ1 2η
|w(s, x)| ≤ C d2η,s,t (P1 , P2 )(t − s)1−( α − α ) + (t − s)kwk∞ + sup Φ(v) (v − s) α (t − v)−Ξ dv
v∈[s,t] s
 
γ
1
1−( α − α1 ) 1+ 2η −Ξ
(B.5) ≤ C d2η,s,t (P1 , P2 )(t − s) + (t − s)kwk∞ + sup Φ(v)(t − s) α ,
v∈[s,t]

and
 2 2λη γ1 2 γ1
|Dw(s, x)| ≤ C d(1−λ)2η,s,t (P1 , P2 )(t − s)1−( α − α − α ) + d2η,s,t (P1 , P2 )(t − s)1−( α − α )
Z t 
1 2η
+ sup Φ(v) (v − s)− α + α (t − v)−Ξ dv
v∈[s,t] s
 2 2λη γ1 2 γ1 1 2η

≤ C d(1−λ)2η,s,t (P1 , P2 )(t − s)1−( α − α − α )
+ d2η,s,t (P1 , P2 )(t − s)1−( α − α ) + sup Φ(v)(t − s)1−( α − α )−Ξ
v∈[s,t]

which in turn gives

 2 2λη γ1 2 γ1
(t − s)Ξ |Dw(s, x)| ≤ C d(1−λ)2η,s,t (P1 , P2 )(t − s)1−( α − α − α )+Ξ + d2η,s,t (P1 , P2 )(t − s)1−( α − α )+Ξ
1 2η

+ sup Φ(v)(t − s)1−( α − α ) .
v∈[s,t]

Taking into account that 0 < t − s ≤ T is sufficiently small, from the preceding inequality, we obtain
 2 2λη γ1 2 γ1

sup Φ(v) ≤ C d(1−λ)2η,s,t (P1 , P2 )(t − s)1−( α − α − α )+Ξ + d2η,s,t (P1 , P2 )(t − s)1−( α − α )+Ξ
v∈[s,t]
 WELL-POSEDNESS OF SOME NON-LINEAR STABLE DRIVEN SDES 31

and plugging the previous estimate into (B.5) yields


1 γ1 2 2η γ1 2 2λη 2η γ1
|w(s, x)| ≤ C d2η,s,t (P1 , P2 )[(t − s)1−( α − α ) + (t − s)2−( α − α − α ) ] + d(1−λ)2η,s,t (P1 , P2 )(t − s)2−( α − α − α − α )

+ (t − s)kwk∞
 
γ γ
1
1−( α − α1 ) 2
2−( α − 2λη − 2η − α1 )
≤ C d2η,s,t (P1 , P2 )(t − s) + d(1−λ)2η,s,t (P1 , P2 )(t − s) α α + (t − s)kwk∞

where for the last inequality we used the fact that 2η + α > 1.
We now select γ1 = 2η. We importantly recall the condition on the regularization parameter ϑ: 1 − α < ϑ <
α − 1 + γ = α − 1 + 2η so that a necessary condition is α + η > 1. Taking into account this condition, we next
optimize the previous inequality on w w.r.t h ∈ C γ1 (Rd ). We thus deduce

 
1 2η 2 2λη 4η
(B.6) d2η,s,t (T(P1 ), T(P2 )) ≤ C d2η,s,t (P1 , P2 )(t − s)1−( α − α ) + d(1−λ)2η,s,t (P1 , P2 )(t − s)2−( α − α − α )

for any λ ∈ [0, 1] s.t. 2λη + α > 1 under the assumption that α + η > 1.
Our final step consists in iterating the previous analysis. We aim at establishing an estimate similar to (B.6) but
for the map T2 and a distance d(1−λ)2η,s,t for a well-chosen λ ∈ [0, 1]. Keeping in mind the inequality (B.6), we
proceed in a similar manner but replacing Pi by T(Pi ). For h ∈ C γ2 , we consider uδi (s, x), i = 1, 2 satisfying

( T(P )

∂s + As i uδi (s, x) = 0, (s, x) ∈ [0, t) × Rd ,
(B.7)
uδi (t, x) = hδ (x), x ∈ Rd ,

and uδi ∈ L∞ ([0, t], C α+γ2 (Rd , R)).

To compare both semigroups, we again write the PDE satisfied by w(s, x) := wδ (s, x) = (uδ1 − uδ2 )(s, x). Namely,
(B.8)  
 

 ∂s + AsT(P1 ) )w(s, x) = − b(s, x, T(P1 )) − b(s, x, T(P2 )) · Dx uδ2 (s, x)

 
T(P1 ),α T(P2 ),α δ
 + (L s − L s )u 2 (s, x) =: −HT(P1 ),T(P2 ) uδ2 (s, x), (s, x) ∈ [0, t) × Rd ,



w(t, x) = 0, x ∈ Rd .

The inequality (3.23) now becomes

1−γ2
(B.9) |HT(P1 ),T(P2 ) uδ2 (v, y)| ≤ Cd2η (T(P1 )(v), T(P2 )(v))(t − v)− α .

This new estimate will be combined with (B.6) in the sequel. Similarly to (B.1), from (B.6), we get

 Z t 
1 γ 2η
1−( α − α2 )
|w(s, x)| ≤ C d2η,s,t (T(P1 ), T(P2 ))(t − s) + (t − s)kwk∞ + dv(v − s) kDw(v, ·)k∞
α

s
 Z t 
γ2
2
2−( α − 2η − ) 2η
(B.10) ≤ C d(1−λ)2η,s,t (P1 , P2 )(t − s) α α + (t − s)kwk∞ + dv(v − s) α kDw(v, ·)k∞
s
32 NOUFEL FRIKHA, VALENTIN KONAKOV, AND STÉPHANE MENOZZI

where we used the direct inequality d2η,s,t (P1 , P2 ) ≤ d(1−λ)2η,s,t (P1 , P2 ). Remark also that similarly to (B.3),
using directly (B.9), as soon as η + α > 1 and γ2 + α > 1, one has
Z t 1 1−γ2
|Dw(s, x)| ≤ C (v − s)− α d2η (T(P1 )(v), T(P2 )(v))(t − v)− α dv
s
Z t
1 2η

+ (v − s)− α + α kw(v, .)k∞ + kDw(v, .)k∞ dv
s
Z t 1 1−γ2
≤C (v − s)− α d2η,s,v (T(P1 ), T(P2 ))(t − v)− α dv
s
Z t
1 2η

+ (v − s)− α + α kw(v, .)k∞ + kDw(v, .)k∞ dv
s
Z t
2 2η 1−γ2
≤ Cd(1−λ)2η,s,t (P1 , P2 ) (v − s)1−( α − α ) (t − v)− α dv
s
Z t
1 2η

+C (v − s)− α + α kw(v, .)k∞ + kDw(v, .)k∞ dv
s
 Z t
3
2−( α − 2η
γ
− α2 ) 1 2η

(B.11) ≤ C d(1−λ)2η,s,t (P1 , P2 )(t − s) α + (v − s)− α + α kw(v, .)k∞ + kDw(v, ·)k∞ dv ,
s
for any λ s.t. 2λη + α > 1 and where we used (B.6) for the last but one inequality. We now proceed in a completely
analogous manner as the first iteration. We thus conclude
2 2η γ2
kwk∞ ≤ Cd(1−λ)2η,s,t (P1 , P2 )(t − s)2−( α − α − α ) .
We now select γ2 = (1 − λ)2η with λ large enough so that 2λη + α > 1 and small enough so that α + γ2 > 1.
Note that this choice is licit under our current assumption α + η > 1 and that it also yields 2 − ( α2 − 2η γ2
α − α ) > 0.
γ2 d
We next optimize the previous inequality on w w.r.t h ∈ C (R )
2 2η γ2
d(1−λ)2η,s,t (T(2) (P1 ), T(2) (P2 )) ≤ Cd(1−λ)2η,s,t (P1 , P2 )(t − s)2−( α − α − α ) .
This shows that T(2) is a contraction on the complete metric space As,s+T,µ w.r.t the distance d(1−λ)2η,s,s+T
under the condition α + η > 1 provided T is sufficiently small. According to the Banach fixed point theorem, T
admits a unique fixed point which is the unique solution to the martingale problem on the time interval [s, s + T ].
We eventually conclude following the same lines of reasonings as those employed at the end of the proof of Theorem
2.3. We omit the remaining technical details.
Remark B.1. We point out that in the previous argument only two iterations of the map T were needed to derive
our result. One may naturally ask if one can do better by performing more iteration of T. This is not clear since
the necessary condition α + η > 1 (coming from the regularization parameter ϑ) already appears at the first iterate
and seems to be the natural one if one wants to be able to compare the distance dγ,s,t (T(P1 , P2 )) with the distance
d2η,s,t (P1 , P2 ) in order to prove a contraction property.

Acknowledgment
For the second and the third authors, the study has been funded by the Russian Science Foundation (project No
17-11-01098).

References
[Bas03] R. F. Bass. Stochastic differential equations driven by symmetric stable processes, pages 302–313. Springer Berlin Heidel-
berg, Berlin, Heidelberg, 2003.
[BBC04] R. F. Bass, K. Burdzy, and Z.-Q. Chen. Stochastic differential equations driven by stable processes for which pathwise
uniqueness fails. Stochastic Processes and their Applications, 111(1):1 – 15, 2004.
[CD18] R. Carmona and F. Delarue. Probabilistic theory of mean field games with applications. I, volume 83 of Probability Theory
and Stochastic Modelling. Springer, Cham, 2018. Mean field FBSDEs, control, and games.
[CdR19] P.-E. Chaudru de Raynal. Strong well posedness of McKean-Vlasov stochastic differential equations with Hölder drift.
Stochastic Processes and their Applications, 2019.
[CdRF18a] P.-E. Chaudru de Raynal and N. Frikha. From the Backward Kolmogorov PDE on the Wasserstein space to propagation
of chaos for McKean-Vlasov SDEs. arXiv:1907.01410, 2018.
[CdRF18b] P.-E. Chaudru de Raynal and N. Frikha. Well-posedness for some non-linear diffusion processes and related PDE on the
Wasserstein space. ArXiV:1811.06904, 2018.
[CdRHM18] P-E. Chaudru de Raynal, I. Honoré, and S. Menozzi. Sharp Schauder Estimates for some Degenerate Kolmogorov Equations.
2018. working paper or preprint.
 WELL-POSEDNESS OF SOME NON-LINEAR STABLE DRIVEN SDES 33

[CdRMP19] P.-E. Chaudru de Raynal, S. Menozzi, and E. Priola. Schauder estimates for drifted fractional operators in the supercritical
case. February 2019. working paper or preprint.
[CZZ17] Z.Q. Chen, X. Zhang, and G. Zhao. Well-posedness of supercritical SDE driven by Lévy processes with irregular drifts.
https://arxiv.org/pdf/1709.04632.pdf, 2017.
[Fun84] T. Funaki. A certain class of diffusion processes associated with nonlinear parabolic equations. Zeitschrift für Wahrschein-
lichkeitstheorie und Verwandte Gebiete, 67(3):331–348, Oct 1984.
[Gär88] J. Gärtner. On the Mckean-Vlasov Limit for Interacting Diffusions. Mathematische Nachrichten, 137(1):197–248, 1988.
[Gra92] C. Graham. Nonlinear diffusion with jumps. Ann. Inst. H. Poincaré Probab. Statist., 28(3):393–402, 1992.
[HM16] L. Huang and S. Menozzi. A Parametrix Approach for some Degenerate Stable Driven SDEs. Annales Instit. H. Poincaré,
52(4):1925–1975, 2016.
[HMP19] L. Huang, S. Menozzi, and E. Priola. Lp Estimates For Degenerate Non-Local Kolmogorov Operators. Journal de
Mathématiques Pures et Appliquées, 121:162–215, 2019.
[HWZ19] Z. Hao, M. Wu, and X. Zhang. Schauder’s estimate for nonlocal kinetic equations and its applications. arXiv:1903.09967,
2019.
[Jac05] N. Jacob. Pseudo differential operators and Markov processes, volume 1. Imperial College Press, 2005.
[JMW05a] B. Jourdain, S. Méléard, and W. A. Woyczynski. A probabilistic approach for nonlinear equations involving the fractional
Laplacian and a singular operator. Potential Anal., 23(1):55–81, 2005.
[JMW05b] B. Jourdain, S. Méléard, and W. A. Woyczynski. Probabilistic approximation and inviscid limits for one-dimensional
fractional conservation laws. Bernoulli, 11(4):689–714, 2005.
[JMW08] B. Jourdain, S. Méléard, and W. A. Woyczynski. Nonlinear SDEs driven by Lévy processes and related PDEs. ALEA Lat.
Am. J. Probab. Math. Stat., 4:1–29, 2008.
[Jou97] B. Jourdain. Diffusions with a nonlinear irregular drift coefficient and probabilistic interpretation of generalized burgers’
equations. ESAIM: PS, 1:339–355, 1997.
[Kac56] M. Kac. Foundations of kinetic theory. In Proceedings of the Third Berkeley Symposium on Mathematical Statistics and
Probability, Volume 3: Contributions to Astronomy and Physics, pages 171–197, Berkeley, Calif., 1956. University of
California Press.
[Kol00] V. Kolokoltsov. Symmetric stable laws and stable-like jump diffusions. Proc. London Math. Soc., 80:725–768, 2000.
[Kol10] V. N. Kolokoltsov. Nonlinear Markov processes and kinetic equations, volume 182 of Cambridge Tracts in Mathematics.
Cambridge University Press, Cambridge, 2010.
[KP10] N. V. Krylov and E. Priola. Elliptic and parabolic second-order PDEs with growing coefficients. Comm. Partial Differential
Equations, 35(1):1–22, 2010.
[Kry96] N. V. Krylov. Lectures on elliptic and parabolic equations in Hölder spaces. Graduate Studies in Mathematics 12. AMS,
1996.
[Lac18] D. Lacker. On a strong form of propagation of chaos for McKean-Vlasov equations. Electron. Commun. Probab., 23:11 pp.,
2018.
[LM16] J. Li and H. Min. Weak solutions of mean-field stochastic differential equations and application to zero-sum stochastic
differential games. SIAM Journal on Control and Optimization, 54(3):1826–1858, 2016.
[McK66] H. P. McKean. A class of markov processes associated with nonlinear parabolic equations. Proceedings of the National
Academy of Sciences of the United States of America, 56(6):1907–1911, 12 1966.
[McK67] H. P. McKean. Propagation of chaos for a class of non-linear parabolic equations. Stochastic Differential Equations (Lecture
Series in Differential Equations, Session 7, Catholic Univ., 1967), pages 41–57, 1967.
[MP14] R. Mikulevicius and H. Pragarauskas. On the Cauchy problem for integro-differential operators in Hölder classes and the
uniqueness of the martingale problem. Potential Anal., 40(4):539–563, 2014.
[MV18] Y. S. Mishura and A. Y. Veretennikov. Existence and uniqueness theorems for solutions of McKean–Vlasov stochastic
equations. Preprint Arxiv arXiv:1603.02212, 2018.
[Oel84] K. Oelschlager. A martingale approach to the law of large numbers for weakly interacting stochastic processes. Ann.
Probab., 12(2):458–479, 05 1984.
[Pri12] E. Priola. Pathwise uniqueness for singular SDEs driven by stable processes. Osaka J. Math., 49(2):421–447, 2012.
[Pri18] E. Priola. Davie’s type uniqueness for a class of SDEs with jumps. Ann. Inst. Henri Poincaré Probab. Stat., 54(2):694–725,
2018.
[RZ18] M. Röckner and X. Zhang. Well-posedness of distribution dependent SDEs with singular drifts. arXiv e-prints, page
arXiv:1809.02216, Sep 2018.
[Sch87] M. Scheutzow. Uniqueness and non-uniqueness of solutions of Vlasov-McKean equations. Journal of the Australian Math-
ematical Society. Series A. Pure Mathematics and Statistics, 43(2):246–256, 1987.
[ST85] T. Shiga and H. Tanaka. Central limit theorem for a system of Markovian particles with mean field interactions. Zeitschrift
für Wahrscheinlichkeitstheorie und Verwandte Gebiete, 69(3):439–459, Sep 1985.
[Szn91] A.-S. Sznitman. Topics in propagation of chaos. In Paul-Louis Hennequin, editor, Ecole d’Eté de Probabilités de Saint-Flour
XIX — 1989, pages 165–251, Berlin, Heidelberg, 1991. Springer Berlin Heidelberg.
[Szt10] P. Sztonyk. Estimates of tempered stable densities. J. Theoret. Probab., 23(1):127–147, 2010.
[TTW74] H. Tanaka, M. Tsuchiya, and S. Watanabe. Perturbation of drift-type for Lévy processes. J. Math. Kyoto Univ., 14(1):73–
92, 1974.
[Vil09] C. Villani. Optimal transport, volume 338 of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles
of Mathematical Sciences]. Springer-Verlag, Berlin, 2009. Old and new.
[Wat07] T. Watanabe. Asymptotic estimates of multi-dimensional stable densities and their applications. Transactions of the
American Mathematical Society, 359(6):2851–2879, 2007.
[ZZ18] X. Zhang and G. Zhao. Dirichlet problem for supercritical non-local operators. arXiv:1809.05712, 2018.
34 NOUFEL FRIKHA, VALENTIN KONAKOV, AND STÉPHANE MENOZZI

Noufel Frikha, Université de Paris, Laboratoire de Probabilités, Statistiques et Modélisation, F-75013 Paris, France.
E-mail address: [email protected]

Valentin Konakov, Laboratory of Stochastic Analysis, Higher School of Economics, Pokrovsky boulevard 11, Moscow,
Russian federation.
E-mail address: [email protected]

Stéphane Menozzi, Laboratoire de Modélisation Mathématique d’Evry (LaMME), UMR CNRS 8070, Université d’Evry
Val d’Essonne, Université Paris-Saclay, 23 Boulevard de France 91037 Evry and Laboratory of Stochastic Analysis,
Higher School of Economics, 31 Shabolovka, Moscow, Russian federation.
E-mail address: [email protected]