The document discusses a trilinear embedding theorem for divergence-form operators with complex coefficients. It proves that if matrices A, B, C are appropriately p,q,r-elliptic, then there is a bound on the integral of the product of the gradients of the semigroups associated with the operators. The proof uses a Bellman function technique and shows the relationship to the concept of p-ellipticity. It generalizes previous work on bilinear embeddings to the trilinear case.
GLYCOSIDES Classification Of GLYCOSIDES Chemical Tests Glycosides
Trilinear embedding for divergence-form operators with complex coefficients
1. Trilinear embedding for divergence-form
operators
Vjekoslav Kovač (U. of Zagreb, PMF–MO)
Joint work with
Andrea Carbonaro (U. of Genova),
Oliver Dragičević (U. of Ljubljana),
Kristina Ana Škreb (U. of Zagreb, Civil Eng.)
Supported by HRZZ UIP-2017-05-4129 (MUNHANAP)
Appl Math 20, Brijuni, Sep 15, 2020
1/18
2. Ellipticity and p-ellipticity
Ω ⊆ Rd an open set
A: Ω → Cd×d with L∞ coefficients (Note: non-smooth and complex)
A is elliptic if ∃λ, Λ ∈ (0, ∞) s.t.
Re 〈A(x)ξ, ξ〉Cd > λ|ξ|2
for ξ ∈ Cd
, for a.e. x ∈ Ω
⃒
⃒ 〈A(x)ξ, η〉Cd
⃒
⃒ 6 Λ |ξ| |η| for ξ, η ∈ Cd
, for a.e. x ∈ Ω
Take p ∈ (0, ∞], but think of p ∈ (1, ∞)
A is p-elliptic if additionally ∃c ∈ (0, ∞) s.t.
Re
⟨︀
A(x)ξ, ξ + |1 − 2/p|ξ̄
⟩︀
Cd > c|ξ|2
for ξ ∈ Cd
, for a.e. x ∈ Ω
(Carbonaro–Dragičević, 2016)
2/18
3. Properties of p-elliptic matrices
Δp(A) := ess inf
x∈Ω
min
ξ∈Cd
|ξ|=1
Re
⟨︀
A(x)ξ, ξ + |1 − 2/p|ξ̄
⟩︀
Cd
A is p-elliptic ⇐⇒ Δp(A) > 0
Δp0 (A) = Δp(A) where 1/p0 + 1/p = 1
Ap(Ω) = the class of p-elliptic matrix functions on Ω
Ap(Ω) decreases in p ∈ [2, ∞)
{elliptic on Ω} = A2(Ω)
{real elliptic on Ω} =
⋂︁
p∈[2,∞)
Ap(Ω)
3/18
4. Divergence-form operators
Boundary conditions reflect the choice of U :
• Dirichlet: U = H1
0
(Ω) = C∞
c
(Ω) in H1(Ω)
• Neumann: U = H1(Ω) = W1,2(Ω)
• mixed: U =
{︀
u|Ω : u ∈ C∞
c
(Rd)
}︀
in H1(Ω), ⊆ ∂Ω closed
Divergence-form operator informally: “LA,U u = −div(A∇u)”
Rigorously:
〈LA,U u, v〉L2(Ω) =
∫︁
Ω
〈A∇u, ∇v〉Cd for u ∈ D (LA), v ∈ U ,
where D (LA,U ) :=
{︀
u ∈ U : RHS extends boundedly to L2(Ω)
}︀
(TA,U
t
)t>0 is the operator semigroup on L2(Ω) generated by −LA,U
4/18
5. Bilinear embeddings
A bilinear embedding is any estimate of the form
∫︁ ∞
0
∫︁
X
|∇Ttf(x)| |∇Ttg(x)| dμ(x) dt 6 C kfkLp(X) kgkLq(X) ,
where (X, μ) is a measure space, (Tt)t>0 is an operator semigroup,
and p, q ∈ (1, ∞) are s.t. 1/p + 1/q = 1
Some history of bilinear embeddings:
• Estimates for the Ahlfors–Beurling operator and iterated Riesz
transf. (Petermichl–Volberg, 2002; Nazarov–Volberg, 2003):
∫︁
R2
(R2
1
f)(x) g(x) dx = −2
∫︁ ∞
0
∫︁
R2
(︀
∂x1 Ttf(x)
)︀(︀
∂x1 Ttg(x)
)︀
dx dt,
where (Ttf)t>0 is the heat extension of f
5/18
6. More history of bilinear embeddings
• Dimension-free Littlewood–Paley estimates
(Dragičević–Volberg, 2006)
• Dimension-free estimates for Schrödinger operators
(Dragičević–Volberg, 2011, 2012)
• Dimension-free estimates for Riesz transforms associated with a
Riemannian manifold (Carbonaro–Dragičević, 2013)
• Functional calculus for generators of symmetric contraction
semigroups (Carbonaro–Dragičević, 2017)
• Bilinear embedding for divergence-form operators with
complex coefficients (Carbonaro–Dragičević, 2020)
– concept of p-ellipticity
– connection with Lp
-dissipativity of sesquilinear forms
6/18
7. Trilinear embedding
Take p, q, r ∈ (1, ∞) s.t. 1/p + 1/q + 1/r = 1
Theorem (Carbonaro–Dragičević–K.–Škreb, 2020)
Suppose that A, B, C : Ω → Cd×d are mx{p, q, r}-elliptic. Then for
f ∈ (Lp ∩ L2)(Ω), g ∈ (Lq ∩ L2)(Ω) and h ∈ (Lr ∩ L2)(Ω) we have
∫︁ ∞
0
∫︁
Ω
⃒
⃒TA,U
t
f
⃒
⃒
⃒
⃒∇TB,V
t
g
⃒
⃒
⃒
⃒∇TC,W
t
h
⃒
⃒dx dt 6 C kfkLp(Ω) kgkLq(Ω) khkLr(Ω)
When Ω = Rd, the same conclusion holds if only: A is p-elliptic, B is
q-elliptic and (1 + q/r)-elliptic, C is r-elliptic and (1 + r/q)-elliptic.
The embedding constant C only depends on p, q, r and the
∗-ellipticity constants of A, B, C
7/18
8. The idea of proof
Let us illustrate it in the very special case d = 1, Ω = R,
A = B = C = I (one-dimensional heat semigroups)
Note that this special case is classical and accessible by tools from
harmonic analysis (boundedness of maximal and square functions)
The following proof by the so-called Bellman function technique was
given by K.–Škreb, 2018
The approach is direct, avoids the need for any classical tools, and
extends (after a lot of extra work) to the general case
8/18
9. Properties of B
Let us find a C1 and “piecewise” C2 function
B = B (u, v, w, U, V, W) of 6 real variables s.t.
(B 1) Domain:
u, v, w, U, V, W > 0, up
6 U, vq
6 V, wr
6 W
(B 2) Range:
0 6 B (u, v, w, U, V, W) 6 C
(︀ 1
p
U + 1
q
V + 1
r
W
)︀
(B 3) Certain concavity:
− 1
2
( d2
B
⏟ ⏞
quadratic form
) (u, v, w, U, V, W)
⏟ ⏞
at a point
(4u, 4v, 4w, 4U, 4V, 4W)
⏟ ⏞
on a vector
> u|4v||4w|
9/18
10. Construction of B
WLOG assume q > r and use the ansatz:
B (u, v, w, U, V, W) = C
(︀ 1
p
U + 1
q
V + 1
r
W
)︀
− α(u, v, w)
α(u, v, w) = up
γ
(︁ vq
up
⏟ ⏞
t
,
wr
up
⏟ ⏞
s
)︁
γ(t, s) =
⎧
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎩
a1 + b1t + c1s; 1 6 s 6 t
a2 + b2t + c2s
1
p0
; s 6 1 6 t
a3 + b3t
1
p0
+ c3s
1
p0
; s 6 t 6 1
a4 + b4t
2
q s
1
r
− 1
q + c4s
1
p0
; t 6 s 6 1
a5 + b5t
2
q + c5t
2
q s1− 2
q + d5s; t 6 1 6 s
a6 + b6t + c6t
2
q s1− 2
q + d6s; 1 6 t 6 s
Adjust the coefficients so that γ is C1 on (0, ∞)2
10/18
11. Construction of B
γ(t, s) =
⎧
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎩
a + bt + cs; 1 6 s 6 t
a(p−1)−c
p−1
+ bt + cp
p−1
s
1
p0
; s 6 1 6 t
a(p−1)−(b+c)
p−1
+ bp
p−1
t
1
p0
+ cp
p−1
s
1
p0
; s 6 t 6 1
a(p−1)−(b+c)
p−1
+ bq
2
t
2
q s
1
r
− 1
q + 2cpr−bp(q−r)
2r(p−1)
s
1
p0
; t 6 s 6 1
2ar(p−1)−b(q+r)
2r(p−1)
+ bq2
2p(q−2)
t
2
q + bq(q−r)
2r(q−2)
t
2
q s1− 2
q
+ 2cr−b(q−r)
2r
s; t 6 1 6 s
a + bq
p(q−2)
t + bq(q−r)
2r(q−2)
t
2
q s1− 2
q + 2cr−b(q−r)
2r
s; 1 6 t 6 s
Choose a, b, c appropriately
Similar to the function constructed by Nazarov–Treil, 1995, used in
bilinear embeddings
11/18
12. Finalizing the proof
Assume f, g, h > 0 and assume for simplicity that B ∈ C2 (mollify it)
u(x, t) := (Ttf)(x), . . . , U(x, t) := (Ttfp
)(x), . . .
b(x, t) := B
(︀
u(x, t), v(x, t), w(x, t), U(x, t), V(x, t), W(x, t)
)︀
=⇒
(︀
∂t − 1
2
∂2
x
)︀
b(x, t) = (∇B )(u, v, . . .) ·
(︀
∂t − 1
2
∂2
x
)︀
(u, v, . . .)
⏟ ⏞
=0
− 1
2
(d2
B )(u, v, . . .)(∂xu, ∂xv, . . .)
Using (B 3) we get
±u(x, t) ∂xv(x, t) ∂xw(x, t) 6
(︀
∂t − 1
2
∂2
x
)︀
b(x, t)
12/18
13. Finalizing the proof
Integrating by parts and using B > 0 we get for δ, T > 0:
±
∫︁
R×(δ,T−δ)
k(x, T − t) u(x, t) ∂xv(x, t) ∂xw(x, t) dx dt
6
∫︁
R
k(x, δ) b(x, T − δ) dx
Letting δ → 0 and using (B 2):
±
∫︁
R×(0,T)
k(x, T − t) u(x, t) ∂xv(x, t) ∂xw(x, t) dx dt 6 b(0, T)
6
C
p
2πT
(︁ 1
p
∫︁
R
f(y)p
e− y2
2T dy +
1
q
∫︁
R
g(y)q
e− y2
2T dy +
1
r
∫︁
R
h(y)r
e− y2
2T dy
)︁
Observe lim
T→∞
p
2πT k(x, T − t) = 1 uniformly over
(x, t) ∈ [−R, R] × (0, T1] and let T → ∞, R → ∞, T1 → ∞
13/18
14. Relation to p-ellipticity
For α: C3 → R, A, B, C ∈ Cd×d, (u, v, w) ∈ C3, and
(ζ, η, ξ) ∈ (Cd)3 we define the generalized Hessian form of α with
respect to (A, B, C),
HA,B,C
α [(u, v, w); (ζ, η, ξ)],
as the standard inner product of
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
ReA − ImA
ImA ReA
ReB − ImB
ImB ReB
ReC − ImC
ImC ReC
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
Re ζ
Im ζ
Re η
Im η
Re ξ
Im ξ
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
∈ (Rd
)6
14/18
15. Relation to p-ellipticity
and
(︀
Hess(α; (u, v, w)) ⊗ IRd
)︀
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
Re ζ
Im ζ
Re η
Im η
Re ξ
Im ξ
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
∈ (Rd
)6
Here one has to interpret Hess(α; (u, v, w)) as the 6 × 6 real
Hessian matrix of the function
R6
→ R, (ur, ui, vr, vi, wr, wi) 7→ α(ur + iui, vr + ivi, wr + iwi)
15/18
16. Relation to p-ellipticity
Lemma
If α is given by the formula α(u, v, w) := |u|a|v|b|w|c for some
a, b, c ∈ [0, ∞〉, then
HA,B,C
α [(u, v, w); (ζ, η, ξ)]
>
1
2
|u|a
|v|b
|w|c
(︁
a2
Δa(A)
⃒
⃒
⃒
ζ
u
⃒
⃒
⃒
2
+ b2
Δb(B)
⃒
⃒
⃒
η
v
⃒
⃒
⃒
2
+ c2
Δc(C)
⃒
⃒
⃒
ξ
w
⃒
⃒
⃒
2
− 2(Λ(A) + Λ(B))
⃒
⃒
⃒
ζ
u
⃒
⃒
⃒
⃒
⃒
⃒
η
v
⃒
⃒
⃒ − 2(Λ(A) + Λ(C))
⃒
⃒
⃒
ζ
u
⃒
⃒
⃒
⃒
⃒
⃒
ξ
w
⃒
⃒
⃒
− 2(Λ(B) + Λ(C))
⃒
⃒
⃒
η
v
⃒
⃒
⃒
⃒
⃒
⃒
ξ
w
⃒
⃒
⃒
)︁
16/18
17. A consequence
“Vertical” square function, defined as
(G A
f)(x) :=
(︂∫︁ ∞
0
⃒
⃒
⃒∇(TA
t
f)(x)
⃒
⃒
⃒
2
dt
)︂1/2
,
is only bounded in a very restrictive range even for real elliptic A
(Auscher, 2007)
“Conical” square function, defined as
(C A
f)(x) :=
(︂∫︁ ∫︁
{|x−y|<
p
t}
⃒
⃒
⃒∇(TA
t
f)(y)
⃒
⃒
⃒
2 dy dt
td/2
)︂1/2
,
is bounded in the full range for real elliptic A
(Auscher–Hofmann–Martell, 2012)
The main theorem here reproves and refines their result
17/18