Lectures on the parametrix method II Sensitivities and approximations
Alex Kulik
Wroc law University of Science and Technology
NOMP II, 23.03.2021
Differentiability of the solution to the parametrix integral equation
Recall that we have started from the Cauchy problem
(∂s+ Ls,x)ps,t(x, y) = 0, s ∈ (−∞, t),
ps,t(x, y) → δy(x), s % t, (1)
with the 2nd order PDO
Ltf (x) = a(t, x) · ∇f (x) + 1
2b(t, x) · ∇2f (x), which we have replaced by the integral equation
ps,t(x, y) = p0s,t(x, y) + Z t
s
Z
Rd
ps,r(x, z)Υr,t(z, y) dzdr, (2) Υs,t(x, y) = (∂s+ Ls,x)p0s,t(x, y).
The solution to (2) is given by
ps,t(x, y) = p0s,t(x, y) + Z t
s
Z
Rd
p0s,r(x, z)Ξr,t(z, y) dzdr, (3)
Ξ(x, y) =X
Υ~k(x, y).
The further aim is to show that the solution to the integral equation (2) is the unique solution to (1)
Naive explanation: p0s,t(x, y) is C1in s and C2 in x. Differentiating in (3) formally, we get
∇xps,t(x, y) = ∇xp0s,t(x, y) + Zt
s
Z
Rd
∇xp0s,r(x, z)Ξr,t(z, y) dzdr (4)
∇2xxps,t(x, y) = ∇2xxp0s,t(x, y) + Z t
s
Z
Rd
∇2xxp0s,r(x, z)Ξr,t(z, y) dzdr (5)
∂sps,t(x, y) = ∂sp0s,t(x, y) − Ξs,t(x, y) + Z t
s
Z
Rd
∂sp0s,r(x, z)Ξr,t(z, y) dzdr (6) Combining ∇x, ∇2xxwith coefficients we get then
(∂sps,t(x, y) + Ls,xps,t(x, y))
= Υs,t(x, y) − Ξs,t(x, y) + Zt
s
Z
Rd
Υs,r(x, z)Ξr,t(z, y) dzdr
= Υ −
∞
XΥ~k+
∞
XΥ ~ Υ~k
!
(x, y) = 0.
Warm up: parametrix-based expansions for sensitivities w.r.t. external parameters
Let coefficients of the diffusion depend on external parameter θ = (α, β):
a = a(α; t, x), b = b(β; t, x).
Then
ps,t(θ; x, y) = ps,t(θ; x, y) +
∞
X
k=1
(p0~ Υ~k)s,t(θ; x, y) and, provided that the corresponding series converge,
∇θps,t(θ; x, y) = ∇θps,t(θ; x, y) +
∞
X
k=1
∇θ(p0~ Υ~k)s,t(θ; x, y)
with
∇θ
p0~ Υ~k
= ∇θp0 ~ Υ~k+
k−1
X
j=0
p0~ Υ~j~ (∇θΥ) ~ Υ~(k−j−1).
UsingFact 3we get that
∂αp0s,t(θ; x, y) = (t − s)X
i
∂αai(α; t, y)Φit−s(a(θ, t, y), b(θ, t, y); x, y)
∂βp0s,t(θ; x, y) = t − s 2
X
i,j
∂βbij(β; t, y)Φi,jt−s(a(θ, t, y), b(θ, t, y); x, y)
have orders (t − s)1/2and (t − s)0, respectively. Similar but more cumbersome calculations give that ∂αΥs,t(θ; x, y), ∂βΥs,t(θ; x, y) have orders (t − s)−1/2+δand (t − s)−1+δ, and the parametrix series for ∇θps,t(θ; x, y) converges.
∂αps,t(θ; x, y) ≈O(t−s)1/2+δ (t − s)X
i
∂αai(α; s, x)Φit−s(a(s, x), b(s, x); x, y)
∂βps,t(θ; x, y) ≈O(t−s)δ
t − s 2
X
i,j
∂βbij(β; s, x)Φi,jt−s(a(θ, s, x), b(θ, s, x); x, y) The higher order approximations are also available.
Derivatives in x
Lemma
The identity (4) for ∇xps,t(x, y) holds true and
|∇xps,t(x, y)| ≤ C(t − s)−1/2ϕc(t−s)(x, y), ϕt(x, y) = ϕt(y − x).
The key estimate:
∂xip0s,t(x, y) = Φit−s(a(s, x), b(s, x); x, y) has order (t − s)−1/2, hence the integral
Z t s
Z
Rd
|∇xp0s,r(x, z)||Ξr,t(z, y)| dzdr is finite.
Lemma
The identity (5) for ∇2xxps,t(x, y) holds true and
|∇2xxps,t(x, y)| ≤ C(t − s)−1ϕc(t−s)(x, y) Here the situation is more subtle because
∂x2ixjp0s,t(x, y) = Φi,jt−s(a(t, y), b(t, y); x, y) has order (t − s)−1, and the integral
Z t s
Z
Rd
|∂2xixjp0s,r(x, z)||Ξr,t(z, y)| dzdr
has no means to be finite. Before proving the lemma, let us first establish certain H¨older continuity properties of the kernels involved into the parametrix construction.
Proposition For any δ < γ,
|p0t(x1, y) − p0t(x2, y)| ≤ C|x1− x2|2δ tδ
ϕct(x1, y) + ϕct(x2, y)
,
|Υ(x1, y) − Υt(x2, y)| ≤ C|x1− x2|2δt−1+γ−δ
ϕct(x1, y) + ϕct(x2, y)
|Ξ(x1, y) − Ξt(x2, y)| ≤ C|x1− x2|2δt−1+γ−δ
ϕct(x1, y) + ϕct(x2, y)
|pt(x1, y) − pt(x2, y)| ≤ C|x1− x2|2δ tδ
ϕct(x1, y) + ϕct(x2, y) . Similar statement holds true for the increments w.r.t. y.
The proof follows the same ‘parametrix’ strategy we have already discussed.
Proof of the Lemma
We have
∂x2ixjp0s,t(x, y) = Φi,jt−s(a(t, y), b(t, y); x, y) and the following two facts:
Φi,jt−s(a(t, y), b(t, y); x, y) − Φi,jt−s(a(t, x), b(t, x); x, y) has order (t − s)−1+γ;
Z
Rd
Φi,jt−s(a(t, x), b(t, x); x, z) dz = 0.
These facts combined give Z
Rd
Φi,jt−s(a(t, z), b(t, z); x, z) dz
≤ C(t − s)−1+γ.
Then Z
Rd
∇2xxp0s,r(x, z)Ξr,t(z, y) dz
≤ Z
Rd
∇2xxp0s,r(x, z)Ξr,t(x, y) dz +
Z
Rd
∇2xxp0s,r(x, z)(Ξr,t(x, y) − Ξr,t(z, y)) dz
≤ C(r − s)−1+δ(t − r)−1+γ−δϕc(t−r)(x, y)
≤ ˜C(r − s)−1+δ(t − r)−1+γ−δϕc(t−s)/2(x, y) for r < t+s2 , and
Z
Rd
∇2xxp0s,r(x, z)Ξr,t(z, y) dz
≤ C(r − s)−1(t − r)−1+γϕc(t−s)(x, y) for r > t+s2 . These estimates provide
Z t s
Z
Rd
∇2xxp0s,r(x, z)Ξr,t(z, y) dz
dr ≤ C(t − s)−1+γϕc(t−s)(x, y).
PMP and the properties of the evolutionary family of operators
We have proved the following Proposition
For any f ∈ C∞(Rd) and t ∈ R, function u(s, x) = Ps,tf (x) satisfies
(∂s+ Ls,x)u(s, x) = 0, s ∈ (−∞, t), u(s, x) → f (x) s % t.
In other words, u(s, x) is harmonic for (∂s+ Ls) on (−∞, t) × Rdand u(t, x) = f (x).
Each operator Lssatisfies the Positive Maximum Principle (PMP) on C∞2 : f (x∗) ≥ f (x), x ∈ Rd, f (x∗) ≥ 0 =⇒ Lsf (x∗) ≥ 0.
Proposition
Let a function u(s, x) be harmonic for (∂s+ Ls) on (−∞, t) × Rdwith Ls satisfying the PMP.
Then u(t, x) ≥ 0 implies u(s, x) ≥ 0, s < t.
Corollary
I. For any f ≥ 0 and s ≤ t, Ps,tf (x) ≥ 0.
II. For any s ≤ r ≤ t, Ps,t= Ps,rPr,t. III. For any s ≤ t and f ∈ C∞2,
Ps,tf (x) = f (x) + Z t
s
Ps,rLrf (x) dr.
To prove, apply Proposition to I. u(s, x) = Ps,tf (x);
II. u(s, x) = ±Ps,tf (x) ∓ Ps,rPr,tf (x), s ≤ r;
III. u(s, x) = ±Ps,tf (x) ∓ f (x) ∓Rt
Ps,rLrf (x) dr.
Summary
Assume that
A.1 Coefficients a(t, x), b(t, x) are bounded.
A.2 Coefficient b(t, x) is uniformly elliptic: for some β > 0, v>b(t, x)v ≥ β|v|2. A.3 Coefficient b(t, x) is H¨older continuous: for some γ ∈ (0,12],
|b(t, x) − b(t0, x)| ≤ C|t − t0|γ, |b(t, x) − b(t, x0)| ≤ C|x − x0|2γ. Then the kernel ps,t(x, y) is the unique Fundamental Solution to the Cauchy problem for
∂s+ Ls in the class of kernels fs,t(x, y) such that for any ε > 0, s < t
|fs,t(x, y)| ≤ Ceε|y−x|2.
The corresponding operator family {Ps,t} is positivity preserving, evolutionary, and satisfies Forward and Backward Kolmogorov’s equations
Back to Processes: SDEs and Martingale Problems
The kernel ps,t(x, y) is the transition density for a Markov process X = {Xt} with the generating family
Ltf (x) = a(t, x) · ∇f (x) + 1
2b(t, x) · ∇2f (x).
The process X is the unique weak solution to the SDE
dXt= a(t, Xt) dt + σ(t, Xt) dWt, σ(t, x)σ(t, x)∗= b(t, x).
Existence of some weak solution to SDE follows by compactness argument By the Itˆo formula, for any weak solution X and u ∈ C1,2,
u(r, Xr) − Z r
s
(∂v+ Lv)u(v, Xv) dv is a martingale.
Taking u(r, x) = Pr,tf (x), we get that
u(r, Xr), r ∈ [s, t]
is a martingale, which yields
In the time homogeneous setting, the above argument is closely related to the notion of the martingale problem.
D. W. Stroock, S. R. S. Varadhan (1979), Multidimensional Diffusion Processes, Springer, Berlin
Ethier, S.N., Kurtz, T.G. (1986) Markov Processes: Characterization and Convergence.
Wiley, New York
Recall that a process X is said to be a solution to the martingale problem (L, C∞2 (Rd)), if for every f ∈ C∞2 (Rd) the process
f (Xt) − Z t
0
Lf (Xs) ds, t ≥ 0
is a martingale w.r.t. the natural filtration of X. The martingale problem is said to be well-posed, if
1 for any probability measure π on Rdthere exists a solution X to the martingale problem with X0∼ π, and
2 any two solutions with same π have the same distribution.
The above argument shows that the martingale problem for Lf (x) = a(x) · ∇f (x) +1
2b(x) · ∇2f (x), f ∈ C∞2 (Rd)
Some other possibilities
We have
|(∇x)kps,t(x, y)| ≤ C(t − s)−k/2ϕc(t−s)(x, y), k = 1, 2, which yields
|∂sps,t(x, y)| ≤ C(t − s)−1ϕc(t−s)(x, y).
I.In the time-homogeneous setting, this can be extended to
|(∂t)mpt(x, y)| ≤ C(t − s)−mϕc(t−s)(x, y), m ≥ 1.
Knopova, V., Kulik, A. (2018) Parametrix construction of the transition probability density of the solution to an SDE driven by α-stable noise. Annales de l’Institut Henri Poincar´e: Probabilites et Statistiques 54
pt(x, y) = p0t(x, y) + Z t/2
0
Z
Rd
p0t−s(x, z)Ξs(z, y) dzds
+ Z t/2
0
Z
Rd
p0s(x, z)Ξt−s(z, y) dz ds.
∂tΥ~(k+1)t (x, y) = Z t/2
0
Z
Rd
(∂tΥ~k)t−s(x, z)Υs(z, y) dzds
+ Zt/2Z
Φ~ks (x, z)(∂tΥ)t−s(z, y) dzds
II.Provided that a, b are Cm in x,
|(∇x+ ∇y)mps,t(x, y)| ≤ Cϕc(t−s)(x, y).
S.D. Eidel’man (1969) Parabolic Systems. North-Holland & Wolters-Noordhoff, Amsterdam.
Approximate fundamental solutions: the way to avoid (5)
Define for η > 0
ps,t,η(x, y) = p(0)s,t+η(x, y) + Z t
s
(p(0)s,r+η∗ Ξr+η,t+η)(x, y) dr.
Denote
Ps,t,ηf (x) = Z
Rd
ps,t,η(x, y)f (y) dy,
∆s,t,ηf (x) = (∂s+ Ls,x)Ps,t,ηf (x).
Lemma
Each Ps,t,ηf (x) is C1 in t and C2 in x, thus ∆s,t,ηf (x) is well defined. For each f ∈ C∞(Rd),
Ps,t,ηf → Ps,tf, η → 0 uniformly on 0 ≤ t − s ≤ T and
∆s,t,ηf → 0, η → 0
(∂sps,t,η(x, y) + Ls,xps,t,η(x, y))
= Υs,t+η(x, y) − (p0s,s+η∗ Ξs+η,t+η)(x, y) +
Z t s
(Υs,r+η∗ Ξr+η,t+η)(x, y) dr.
→ Υ −
∞
X
k=1
Υ~k+
∞
X
k=1
Υ ~ Υ~k
!
s,t
(x, y) = 0, η → 0.
We have proved the following Proposition
For any f ∈ C∞(Rd) and t ∈ R, sequence uη(s, x) = Ps,t,ηf (x) satisfies
(∂s+ Ls,x)uη(s, x) → 0, s ∈ (−∞, t − τ ), uη(s, x) → u(s, x) := Ps,tf (x) .
In such a case, we call function u(s, x) approximate harmonic
Rebuilding the theory using approximate harmonic functions
A. Kulik (2019) Approximation in law of locally α-stable L´evy-type processes by non-linear regressions. Electronic Journal of Probability 24
V. Knopova, A. Kulik, R. Schilling (2020) Construction and heat kernel estimates of general stable-like Markov processes, arXiv:2005.08491
Proposition
Let a function u(s, x) beapproximateharmonic for (∂s+ Ls) on (−∞, t) × Rdwith Ls
satisfying the PMP.
Then u(t, x) ≥ 0 implies u(s, x) ≥ 0, s < t.
Corollary
I. For any f ≥ 0 and s ≤ t, Ps,tf (x) ≥ 0.
II. For any s ≤ r ≤ t, Ps,t= Ps,rPr,t. III. For any s ≤ t and f ∈ C∞2,
Ps,tf (x) = f (x) + Z t
s
Ps,rLrf (x) dr.
Proposition
Let a function u(s, x) beapproximateharmonic for (∂s+ Ls) and Y be a weak solution to the SDE.
Then u(r, Yr), r ∈ [s, t] is a martingale.
Mr,η:= uη(r, Yr) − Z r
s
(∂v+ Lv)uη(v, Yv) dv → u(r, Yr), η → 0.
Taking u(r, x) = Ps,tf (x) we get that X, Y have the same laws.
Summary
I. ps,t(x, y) defines a Markov process X.
II. X is the unique weak solution to SDE.
III. ps,t(x, y) solves the Forward Equation.
IV. ps,t(x, y) solves the Backward Equation in the approximate sense.
These properties are based on the parametrix integral equation only, and allow easy extension to the L´evy-type PDOs