Professional Documents
Culture Documents
M. Thamban Nair
Department of Mathematics
Indian Institute of Technology Madras
Chennai, India
Email: mtnair@iitm.ac.in
Tx = y (∗)
ky − yε k ≤ ε
but
1
kx − xε k ≥ .
ε
Let
1
xε = x + uε , yε = Txε .
ε
Then Txε = Tx + 1ε Tuε so that
1
kTxε − Txk = kTuε k < ε.
ε
Note that
1 1
kxε − xk = kuε k = .
ε ε
Proof.
A compact operator on an infinite dimensiional normed linear
space is not bounded below.
1 √
kTx − Txk k = √ kTuk k = σ k ,
σk
but
1
kx − xk k = √ .
σk
As k → ∞,
∂u ∂2u
= c2 2 , 0 < x < `, 0 < t < τ, (1)
∂t ∂x
u(0, t) = 0 = u(`, t), (2)
from the knowledge of the final value
u(·, τ ) := g .
with r
nπc 2 nπ
λn := , ϕn (x) := sin x
` ` `
satisfies the equations (1) and (2) and the initial condition
u(·, 0) = f .
and
2
hf , ϕn i = e λn t hu(·, t), ϕn i.
Hence, from (∗∗),
∞
2
X
u(x, τ ) = e −λn (τ −t) hu(·, t), ϕn iϕn (x).
n=1
Af = g ,
where
∞
2
X
Aϕ := e −λn (τ −t) hϕ, ϕn iϕn .
n=1
µn → 0 exponentially!
The inverse problem is severly ill-posed!
with
−∆ϕn = λn ϕn , n ∈ N;
and {ϕn : n ∈ N} is an orthonormal basis of L2 (Ω).
BHCP: From the knowledge of g := u(·, τ ) determine
f := u(·, t).
M. Thamban Nair Inverse Problem in PDE
Regularization - Stable approximation method
as k → ∞.
Suppose the data y in noisy, i.e., we have ỹ in place of y . Then
one may take the approximation as
k
X hỹ , vn i
x̃k := un .
σn
n=1
Then
k
X hy − ỹ , vn i
xk − x̃k = un .
σn
n=1
Thus,
|hy − ỹ , vk i|2
kx − x̃k k2 ≥ kx − xk k2 +
σk2
and
δ2
kx − x̃k k2 ≤ kx − xk k2 + .
σk2
δ2
kx − x̃k k2 = kx − xk k2 + . (∗)
σk2
δ2
kx − x̃k k2 ≤ kuk2 σk+1
2
+ 2
σk+1
M. Thamban Nair Inverse Problem in PDE
Note that
∞
X |hx, un i|2
x ∈ R(T ∗ ) ⇐⇒ < ∞.
σn2
n=1
Theorem
Suppose x ∈ R(T ∗ ) with x = T ∗ u and k is such that δ ≤ σk+1
2 ,
then
kx − x̃k k2 ≤ (1 + kuk2 ) σk+1
2
.
More generally:
Theorem
2ν+1
If x ∈ R((T ∗ T )ν ) and k is such that δ ≤ σk+1 , then
2ν+1
kx − x̃k k2 = O(σk+1 ).
H: a Hilbert space;
A : D(A) ⊆ H: a densely defined positive self adjoint
unbounded operator.
d
u(t) + Au(t) = f (t), u(0) = ϕ0 . (1)
dt
Known2 : Z t
u(t) = e −tA ϕ0 + e −(t−s)A f (s)ds.
0
2
See: A. Pazzy, Semigroups of Linear Operators and Applications to PDE,
Springer-Verlag, 1983
M. Thamban Nair Inverse Problem in PDE
Here, the operator e −tA is defined by
Z ∞
−tA
e ϕ := e −tλ dEλ ϕ.
0
where Z ∞
D(e tA ) := {ϕ ∈ H : e 2tλ dkEλ ϕk2 }
0
and Z ∞
tA 2
ke ϕk = e 2tλ dkEλ ϕk2 } ≥ kϕk2 .
0
S(t)ϕ − ϕ
+ Aϕ → 0 as t → 0.
t
−A is the infinitesimal generator of S, i.e.,
S(t)ϕ − ϕ
−Aϕ := lim , ϕ ∈ D(−A)
t→0 t
where
S(t)ϕ − ϕ
D(−A) := {ϕ ∈ H : lim exists}.
t→0 t
d
u(t) + Au(t) = f (t), u(τ ) = ϕτ . (2)
dt
Suppose u is a solution of (2). Then
Z t
−tA
u(t) = e ϕ0 + e −(t−s)A f (s)ds, u(τ ) = ϕτ ,
0
u(t) = e (τ −t)A ϕτ .
In particular,
Z τ
−τ A
ϕτ = e ϕ0 + e −(τ −s)A f (s)ds.
0
Since e −τ A ϕ0 ∈ D(e τ A ),
Z τ
ϕτ − e −(τ −s)A f (s)ds = e −τ A ϕ0 ∈ D(e τ A ).
0
Hence, Z τ
ϕ0 = e τA
ϕτ − e −(τ −s)A f (s)ds .
0
That is,
Z τ
−(τ −t)A
e u(t) = (ϕτ − e −(τ −s)A f (s)ds
0
Z t
+e −(τ −t)A e −(t−s)A f (s)ds
0
Z τ
−(τ −t)A
= ϕτ − e e −(t−s)A f (s)ds
0
Z t
−(τ −t)A
+e e −(t−s)A f (s)ds.
0
Thus, Z τ
ϕτ − e −(τ −s)A f (s)ds ∈ D(e (τ −t)A )
t
and Z τ
(τ −t)A
u(t) = e ϕτ − e −(τ −s)A f (s)ds .
t
Definition
If ϕτ ∈ H and f ∈ L1 ([0, τ ], H) are such that
Z τ
ψ(t) := ϕτ − e −(τ −s)A f (s)ds ∀ t ∈ [0, τ )
t
Theorem
Let ϕτ ∈ D(e τ A ) and let u : [0, τ ] → H be defined by
u(t) = e (τ −t)A ϕτ , t ≥ 0. Then u is a solution of the FVP
ut + Au(t) = 0, u(τ ) = ϕτ
u(t + h) − u(t)
lim exists ⇐⇒ u(t) ∈ D(−A)
h→0 h
⇐⇒ ϕτ ∈ D(Ae τ A ).
ku(t) − uβ (t)k → 0 as β → ∞.
ku(t) − uβ (t)k → 0 as β → ∞.
Proof.
Since Z ∞
2
ku(t)k = e 2(τ −t)λ dkEλ (ψ(t))k2 < ∞,
0
we obtain
Z ∞
2
ku(t) − uβ (t)k = e 2(τ −t)λ dkEλ (ψ(t))k2 → 0 as β → ∞.
β
1 1
β≈ log p , 0 < p < 1,
τ −t δ
⇒
ku(t) − ũβ (t)k = o(1) as δ → 0.
and Z β
uβ (t) − ũβ (t) = e (τ −t)λ dEλ (ψ(t) − ψ̃(t)).
0
Note that
Z τ
kψ(t) − ψ̃(t)k ≤ kϕt − ϕ̃τ k + ke −(τ −s)A k kf (s) − f˜(s)kds
Zt τ
≤ kϕt − ϕ̃τ k + kf (s) − f˜(s)kds
t
≤ kϕt − ϕ̃τ k + kf − f˜k1 .
Thus,
Thus,
(i) ht (λ) → ∞ as λ → ∞,
(i) u(t) ∈ D(ht (A)),
(ii) kht (A)u(t)k ≤ ρt for some ρt > 0.
Then
ρt
ku(t) − uβ (t)k ≤ .
ht (β)
Hence, we have
Theorem
Let
ξt (λ) := ht (λ)e (τ −t)λ , λ>0
and
β = βt := ξt−1 (ρ/δ).
Then
2ρ
ku(t) − ũβ (t)k ≤ .
h(ξt−1 (ρ/δ))
In particular,
and
δ
kuαL (t) − ũαL (t)k ≤ . (8)
α
Note that the estimate
1
e (τ −t)β = .
α
That is,
1 1
β= ln .
τ −t α
equivalently,
1 1 p
β= ln and α = δ/ρ,
τ −t α
that is, if
1 ρ
β := ln .
2(τ − t) δ