Ill-Posed Operaror Equations

and an Inverse Problem in PDE

M. Thamban Nair

Department of Mathematics
Indian Institute of Technology Madras
Chennai, India
Email: mtnair@iitm.ac.in

(Ph.D. IIT Bombay, 1986)

Diamond Jubily Symposium, IIT Bombay

4-6 January 2019

M. Thamban Nair Inverse Problem in PDE

Ill-Posed Operator Equations

Let T : X → Y be a linear operator between normed linear spaces.

The problem of solving the operator equation

Tx = y (∗)

for y ∈ Y is said to be ill-posed if it does not have a unique

solution which depends continuosly on the data y .

M. Thamban Nair Inverse Problem in PDE

Theorem
If T is not bounded below, then (∗) is ill-posed.
In fact, if y ∈ R(T ) and Tx = y , then for every ε > 0, there exists
yε ∈ Y and xε ∈ X such that Txε = yε and

ky − yε k ≤ ε

but
1
kx − xε k ≥ .
ε

Given any sequence (αn ) of poisitve real numbers such that

αn → 0 and n → ∞, there there exists yn ∈ Y and xn ∈ X
such that Txn = yn and
1
ky − yn k ≤ αn but kx − xε k ≥
αn
for all n ∈ N.
M. Thamban Nair Inverse Problem in PDE
Proof of the theorem.
Since T is not bounded below, for every ε > 0, there exists uε ∈ X
with kuε k = 1 such that

kTuε k < ε2 kuε 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

M. Thamban Nair Inverse Problem in PDE

Corollary
If X is infinite dimensional and T is a compact operator, then (∗)
is ill-posed.

Proof.
A compact operator on an infinite dimensiional normed linear
space is not bounded below.

Recall that if k(·, ·) ∈ L2 (Ω × Ω), where Ω is a bounded open

set in Rk , the operator T defined by
Z
(Tx)(s) := k(s, ζ)x(ζ)dζ, x ∈ L2 (Ω),
Ω

is a compact operator from the infinite dimensional space

L2 (Ω) into itself.

M. Thamban Nair Inverse Problem in PDE

An illustration
Let X and Y be Hilbert spaces and T : X → Y be a compact
operator of infinite rank. Let
∞
X
Tx := σn hx, un ivn , x ∈ X,
n=1

be a singular value decompsotion1 of T .

Let y ∈ Y .
For y to be in R(T ), it is necessary that
∞
X |hy , vn i|2
<∞
σn2
n=1

and in that case, Tx = y , where

∞
X hy , vn i
x= un ∈ N(T )⊥ .
σn
n=1
1
See M.T. Nair, Functional Analysis: A First Course, PHI-Learning.
M. Thamban Nair Inverse Problem in PDE
For k ∈ N, let xk = x + √1 uk . Then
σk

1 √
kTx − Txk k = √ kTuk k = σ k ,
σk

but
1
kx − xk k = √ .
σk

As k → ∞,

kTx − Txk k → 0 but kx − xk k → ∞.

M. Thamban Nair Inverse Problem in PDE

Backward heat conduction problem

Find u(·, t) ∈ L2 [0, `] such that

∂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 .

M. Thamban Nair Inverse Problem in PDE

Recall:

For f ∈ L2 [0, `],

∞
2
X
u(x, t) = e −λn t hf , ϕn iϕn (x), (∗)
n=1

with r
nπc 2  nπ 
λn := , ϕn (x) := sin x
` ` `
satisfies the equations (1) and (2) and the initial condition

u(·, 0) = f .

Knowing f := u(·, 0), we otbain u(·, t) in a stable manner.

M. Thamban Nair Inverse Problem in PDE

Inverse problem

From the knowledge of g := u(·, τ ), find f := u(·, t) for

0 ≤ t < τ.
From (∗),
∞
2
X
u(x, τ ) = e −λn τ hf , ϕn iϕn (x). (∗∗)
n=1

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

M. Thamban Nair Inverse Problem in PDE

This is same as solving the operator equation:

Af = g ,

where
∞
2
X
Aϕ := e −λn (τ −t) hϕ, ϕn iϕn .
n=1

Note that A is a compact, positive, self adjoint operator on L2 [0, `]

with eigenvalues
2
µn := e −λn (τ −t) .

µn → 0 exponentially!
The inverse problem is severly ill-posed!

M. Thamban Nair Inverse Problem in PDE

General case
Ω: bounded domain in Rk ;
∂u
= c 2 ∆u, (x, t) ∈ Ω × [0, τ ],
∂t
u(x, t) = 0, x ∈ ∂Ω, t ∈ [0, τ ).
In this case, u(·, 0) = f implies
∞
2
X
u(x, t) = e −λn t hf , ϕn iϕn (x), (∗)
n=1

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

Since a compact operator equation is ill-posed,

Small error in the data can cause large deviation in the
solution.

One has to use some regularization method for obtaining

stable approximation methods.
We shall consider one such method, the so called, spectral cut-off
method.

M. Thamban Nair Inverse Problem in PDE

Spectral cut-off method:

Recall: If T : X → Y is a compact operator of infinite rank, and if

∞
X
T = σn h·, un ivn
n=1

is an SVD of V , and if y ∈ R(T ), then

∞
X hy , vn i
x := un
σn
n=1

belongs to N(T )⊥ and Tx = y .

A natural way of approximating x would be to take a k-th cut-off
of x:
k
X hy , vn i
xk := un .
σn
n=1

M. Thamban Nair Inverse Problem in PDE

Clearly,
∞
X |hy , vn i|2
kx − xk k2 = →0
σn2
n=k+1

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

M. Thamban Nair Inverse Problem in PDE

 kx − x̃k k2 = kx − xk k2 + kxk − x̃k k2
k
X |hy − ỹ , vn i|2
2
= kx − xk k +
σn2
n=1
|hy − ỹ , vk i|2
≥ kx − xk k2 + .
σk2

Thus,
|hy − ỹ , vk i|2
kx − x̃k k2 ≥ kx − xk k2 +
σk2
and
δ2
kx − x̃k k2 ≤ kx − xk k2 + .
σk2

M. Thamban Nair Inverse Problem in PDE

In particular, if ỹ := y + δvk , then ky − ỹ k = δ and

δ2
kx − x̃k k2 = kx − xk k2 + . (∗)
σk2

For fixed δ, kx − x̃k k can be large (for large k).

Rate of convergence cannot be assertained without having
source conditions on x:

M. Thamban Nair Inverse Problem in PDE

Estimate under source condition(s)
Recall
∞
X
2
kx − xk k = |hx, un i|2 .
n=k+1
Suppose:
x ∈ R(T ∗ ) so that x = T ∗ u. (∗∗)
Then,
∞
X ∞
X
2 ∗ 2
kx − xk k = |hT u, un i| = |hu, Tun i|2
n=k+1 n=k+1
X∞
= σn2 |hu, vn i|2 ≤ kuk2 σk+1
2
.
n=k+1

(∗) and (∗∗) imply:

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 ).

M. Thamban Nair Inverse Problem in PDE

IVP for a Parabolic Problem

H: a Hilbert space;
A : D(A) ⊆ H: a densely defined positive self adjoint
unbounded operator.

Given ϕ0 ∈ H and f ∈ L1 ([0, ∞), H), consider the initial value

problem (IVP):

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 {Eλ : λ ≥ 0} is the resolution of identity of A.

Z ∞ Z ∞
−tA 2 −2tλ 2
ke ϕk := e dkEλ ϕk ≤ dkEλ ϕk2 = kϕk2 .
0 0

Recall spectral theorem:

There exists a resolution of identity {Eλ : λ ≥ 0} such
that Z ∞
Aϕ = λdEλ ϕ, ϕ ∈ D(A),
0
R∞
and in that case D(A) := {ϕ ∈ H : 0 λ2 dkEλ ϕk2 }.

M. Thamban Nair Inverse Problem in PDE

For any continuous function g : [0, ∞) → R,
Z ∞
g (A)ϕ := g (λ)dEλ ϕ, ϕ ∈ D(g (A))
0

is a self adjoint operator with

Z ∞
D(g (A)) := {ϕ ∈ H : |g (λ)|2 dkEλ ϕk2 }.
0

For ϕ ∈ D(g (A),

Z ∞
2
kg (A)ϕk := |g (λ)|2 dkEλ ϕk2 .
0

M. Thamban Nair Inverse Problem in PDE

In particular, we have the following:
For t ≥ 0
Z ∞
tA
e ϕ := e tλ dEλ ϕ, ϕ ∈ D(e tA ),
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

e tA is one-one, onto (since self adjoint), and has bounded

inverse.

R(e −tA ) = D(e tA ) ∀ t ≥ 0.

M. Thamban Nair Inverse Problem in PDE

Let
S(t) := e −tA , t ≥ 0.
Then
{S(t) : t ≥ 0} is a strongly continuous semigroup on H with

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

M. Thamban Nair Inverse Problem in PDE

Final value problem (FVP)

Let τ > 0, ϕτ ∈ H and f ∈ L1 ([0, τ ], H).

Consider the final value problem (FVP):

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

where ϕ0 := u(0). In particular,

Z τ
−τ A
ϕτ = u(τ ) = e ϕ0 + e −(τ −s)A f (s)ds.
0

M. Thamban Nair Inverse Problem in PDE

Theorem
Let ϕτ ∈ H and f ∈ L1 ([0, τ ], H). If the FVP (2) has a solution
u(·) with u(τ ) = ϕτ , then
Z τ
ψ(t) := ϕτ − e −(τ −s)A f (s)ds
t

belongs to D(e (τ −t)A ) and

u(t) = e (τ −t)A ψ(t).

In the above, if f = 0, then

u(t) = e (τ −t)A ϕτ .

Since A is an unbounded operator,

Small error in ϕτ can lead to larege error in the
solution.
M. Thamban Nair Inverse Problem in PDE
Proof of theorem.
Let ϕ0 := u(0). Then
Z t
−tA
u(t) = e ϕ0 + e −(t−s)A f (s)ds. (3)
0

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

M. Thamban Nair Inverse Problem in PDE

Continues.
Therefore, (3) implies
 Z τ  Z t
−tA τ A −(τ −s)A
u(t) = e eϕτ − e f (s)ds + e −(t−s)A f (s)ds
0 0
 Z τ  Z t
= e (τ −t)A ϕτ − e −(τ −s)A f (s)ds + e −(t−s)A f (s)ds.
0 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

M. Thamban Nair Inverse Problem in PDE

Continues.
Hence,
Z τ
−(τ −t)A −(τ −t)A
e u(t) = ϕτ − e e −(t−s)A f (s)ds.
t

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

M. Thamban Nair Inverse Problem in PDE

In view of the above theorem, we introdice the following definition.

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

belongs to D(e (τ −t)A ), then u(·) defined by

u(t) = e (τ −t)A ψ(t)

is called the mild solution of the FVP (2).

M. Thamban Nair Inverse Problem in PDE

FVP as operator equation
The problem of finding a mild solution u(·) of the FVP with
u(τ ) = ϕτ can be posed as a problem of solving the operator
equation
At u(t) = ψ(t), (4)
where
At ϕ := e −(τ −t)A ϕ, ϕ ∈ H,
Z τ
ψ(t) := ϕτ − e −(τ −s)A f (s)ds.
t
Note that:
At is an injective bounded self adjoint operator.
R(At ) = D(e (τ −t)A ) is dense in H.
At−1 = e (τ −t)A : R(At ) → H is not continuous.

Hence, (4) is ill-posed.

M. Thamban Nair Inverse Problem in PDE

 A mild solution is not necessary to be a solution of the FVP:

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

if and only if ϕτ ∈ D(Ae τ A ).

M. Thamban Nair Inverse Problem in PDE

Proof.
For t ≥ 0 and h > 0,

u(t + h) − u(t) e (τ −t−h)A ϕτ − e (τ −t)A ϕτ e −hA u(t) − u(t)

= = .
h h h
Since −A is the infinitesimal generator of the semigroup
{e −hA : h ≥ 0},

u(t + h) − u(t)
lim exists ⇐⇒ u(t) ∈ D(−A)
h→0 h
⇐⇒ ϕτ ∈ D(Ae τ A ).

Thus, u 0 (t) exists for every t ≥ 0 iff e τ A ϕτ ∈ D(−A) iff

ϕτ ∈ D(Ae τ A ), and in that case u 0 (t) = −Au(t), u(τ ) = ϕτ .

M. Thamban Nair Inverse Problem in PDE

Truncated spectral regularization (TRS)

Let ϕτ and f ∈ L1 ([0, τ ], H).

Recall that, the mild solution of the FVP is
Z ∞
(τ −t)A
u(t) = e ψ(t) = e (τ −t)λ dEλ (ψ(t)) (5)
0
Rτ
whenever ψ(t) := ϕτ − t e −(τ −s)A f (s)ds belongs to D(e τ A ).

Since small error in the data (ϕτ , f ) can lead to large

error in the solution u(·), we have to look for a
regularized solution which depends continuously on the
data (ϕτ , f ).

M. Thamban Nair Inverse Problem in PDE

Looking at the expression in (5) for the mild solution, we define
such a regularized solution as
Z β
uβ (t) = e (τ −t)λ dEλ (ψ(t)) (6)
0

for each β > 0.

The following theorem shows that uβ (·) is an approximation of u(·)
for large β.

M. Thamban Nair Inverse Problem in PDE

Theorem
Under the assumption ψ(t) ∈ D(e τ A ),

ku(t) − uβ (t)k → 0 as β → ∞.

M. Thamban Nair Inverse Problem in PDE

Theorem
Under the assumption ψ(t) ∈ D(e τ A ),

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

M. Thamban Nair Inverse Problem in PDE

Now, we show that

uβ (·) is stable under perturbations in the data (ϕτ , f ).

Suppose ϕ̃τ ∈ H and f˜ ∈ L1 ([0, τ ], H) are the noisy data, in place
of the actual data ϕτ and f , respectively.
Let Z β
ũβ (t) = e (τ −t)λ dEλ (ψ̃(t)),
0
where Z τ
ψ̃(t) := ϕ̃τ − e −(τ −s)A f˜(s)ds.
t

M. Thamban Nair Inverse Problem in PDE

Theorem
Let ϕτ , ϕ̃τ ∈ H and f , f˜ ∈ L1 ([0, τ ], H). The for each t ∈ [0, τ ]
and β > 0,

kuβ (t) − ũβ (t)k ≤ e (τ −t)β kψ(t) − ψ̃(t)k

≤ e (τ −t)β (kϕτ − ϕ̃τ k + kf − f˜k1 ).

Suppose kϕτ − ϕ̃τ k + kf − f˜k1 ≤ δ for some δ > 0. Then we

obtain
kuβ (t) − ũβ (t)k ≤ e (τ −t)β δ.
Observation:
For a fixed β > 0,

kuβ (t) − ũβ (t)k → 0 as δ → 0.

uβ (t) is stable under perturbations in the data (ϕτ , f ).

M. Thamban Nair Inverse Problem in PDE

Convergence

From the above theorem we have

Theorem
Let ϕτ , ϕ̃τ ∈ H and f , f˜ ∈ L1 ([0, τ ], H) such that

kϕτ − ϕ̃τ k + kf − f˜k1 ≤ δ

for some δ > 0. The for each t ∈ [0, τ ] and β > 0,

ku(t) − ũβ (t)k ≤ ku(t) − uβ (t)k + e (τ −t)β δ.

1 1
β≈ log p , 0 < p < 1,
τ −t δ
⇒
ku(t) − ũβ (t)k = o(1) as δ → 0.

M. Thamban Nair Inverse Problem in PDE

Proof of Theorem.
We observe that
Z τ
ψ(t) − ψ̃(t) = ϕt − ϕ̃τ ) − e −(τ −s)A˜(f (s) − f˜(s))ds
t

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 .

M. Thamban Nair Inverse Problem in PDE

Continues.
Hence,
Z β
kuβ (t) − ũβ (t)k2 = e 2(τ −t)λ dEλ kψ(t) − ψ̃(t)k2
0
≤ e 2(τ −t)β kψ(t) − ψ̃(t)k2

Thus,

kuβ (t) − ũβ (t)k ≤ e 2(τ −t)β kψ(t) − ψ̃(t)k

≤ e 2(τ −t)β (kϕt − ϕ̃τ k + kf − f˜k1 ).

M. Thamban Nair Inverse Problem in PDE

Continues.
Hence,
Z β
kuβ (t) − ũβ (t)k2 = e 2(τ −t)λ dEλ kψ(t) − ψ̃(t)k2
0
≤ e 2(τ −t)β kψ(t) − ψ̃(t)k2

Thus,

kuβ (t) − ũβ (t)k ≤ e 2(τ −t)β kψ(t) − ψ̃(t)k

≤ e 2(τ −t)β (kϕt − ϕ̃τ k + kf − f˜k1 ).

Next we obtain an estimate for the error under an additional

smoothness assumption on u(·).

M. Thamban Nair Inverse Problem in PDE

Theorem
Suppose ϕτ ∈ H and f ∈ L1 ([0, τ ], H) are such that for each
t ∈ [0, τ ), ψ(t) ∈ D(e τ A ) and there exists a monotonically
increasing continuous function ht ()˙ : [0, τ ] → [0, ∞) such that

(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)

M. Thamban Nair Inverse Problem in PDE

Proof.
Recall that
u(t) = e (τ −t)A ψ(t).
Hence,
Z ∞
ku(t) − uβ (t)k 2
= e 2(τ −t)λ dkEλ (ψ(t))k2
β
Z ∞
1
= ht (λ)2 e 2(τ −t)λ dkEλ (ψ(t))k2
β ht (λ)2
Z ∞
1
≤ ht (λ)2 e 2(τ −t)λ dkEλ (ψ(t))k2
ht (β)2 β

M. Thamban Nair Inverse Problem in PDE

Continues.
By the assumption,
Z ∞
ht (λ)e (τ −t)λ dkEλ (ψ(t))k2 = kht (A)e (τ −t)A ψ(t)k2
0
= kht (A)u(t)k2 ≤ ρ2t .

Hence, we have

ku(t) − uβ (t)k ≤ ρt /ht (β).

M. Thamban Nair Inverse Problem in PDE

Combining the last two theorems, we obtain the following.

M. Thamban Nair Inverse Problem in PDE

Combining the last two theorems, we obtain the following.
Theorem
Suppose ϕ̃τ and f˜ are noisy data such that

kϕτ − ϕ̃τ k + kf − f˜k1 ≤ δ

for some noise level δ > 0. Then

kuβ (t) − ũβ (t)k ≤ e (τ −t)β δ.

If ρt > 0 and ht (·) are as in last theorem, then we have

ρt
ku(t) − ũβ (t)k ≤ + e (τ −t)β δ.
ht (β)

M. Thamban Nair Inverse Problem in PDE

Parameter choice strategy

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,

ku(t) − ũβ (t)k → 0 as δ → 0.

M. Thamban Nair Inverse Problem in PDE

Proof.
Note that
ρt ρt
= e (τ −t)β δ ⇐⇒ ξt (β) := ht (β)e (τ −t)β =
ht (β) δ
⇐⇒ β = ξt−1 (ρt /δ).

Thus, for the choice of β = ξt−1 (ρt /δ),

ρt
ku(t) − ũβ (t)k ≤ + e (τ −t)β δ
ht (β)
2ρ
≤ −1
h(ξt (ρ/δ))

Since h(ξt−1 (ρt /)) → ∞ as → 0.

ku(t) − ũβ (t)k → 0 as → 0.

M. Thamban Nair Inverse Problem in PDE

Remarks on optimality
Recall that the operator At : H → H defined by

At ϕ := e −(τ −t)A ϕ, ϕ∈H

is injective, continuous, self adjoint, with R(At ) dense in H.

Therefore,
u(·) is a generalized solution
Z τ
At u(t) = ϕτ − e −(τ −s)A f (s)ds (7)
t

if and only if it is a solution.

Let u(t) be the solution of (7) and let uαL (·) be the Lavrentive
regularized solution, i.e.,
Z τ
(At + αI )uαL (t) = ψ(t) := ϕτ − e −(τ −s)A f (s)ds.
t

M. Thamban Nair Inverse Problem in PDE

Then, from the standard theory, we know that

ku(t) − uαL (t)k → 0 as α → 0

and
δ
kuαL (t) − ũαL (t)k ≤ . (8)
α
Note that the estimate

kuβ (t) − ũβ (t)k ≤ δe (τ −t)β

obtained eariler is same as (8) if we take β such that

1
e (τ −t)β = .
α
That is,
1 1
β= ln .
τ −t α

M. Thamban Nair Inverse Problem in PDE

Next, suppose

u(t) = At v (t) with kv (t)k ≤ ρt , (9)

equivalently,

u(t) ∈ D(e (τ −t)A ) with ke (τ −t)A u(t)k ≤ ρt . (10)

Then we have the estimate

ku(t) − uαL (t)k ≤ ρt α.

 
1
Under the choice β := τ −t ln α1 , the above estimate takes the
form
ku(t) − uαL (t)k ≤ ρt e −(τ −t)β . (11)
This is same as the estimate obtained earlier for ku(t) − uβ (t)k
under the (10) .

M. Thamban Nair Inverse Problem in PDE

Thus, we can conclude:

If ht (λ) := e (τ −t)λ , then the the estimate obtained

under TSR is same as the order optimal rate possible
for the Lavrentive regularization for the source
condition (9), if

1 1 p
β= ln and α = δ/ρ,
τ −t α
that is, if
1 ρ
β := ln .
2(τ − t) δ

Similar conclusion can be made for a general ht (·) as well.

M. Thamban Nair Inverse Problem in PDE

References

1 A. Jana, Regularization of Ill-Posed Nonhomogeneous

Parabolic Problems, Ph.D. Thesis, IIT Madras, August, 2018.
2 M.T. Nair, Functional Analysis: A First Course,
Prentice-Hall of India, New Delhi, 2002 (Fourth Print: 2014).
3 M.T. Nair, Linear Operator Equations: Approximation and
Regularization World Scientific, Singapore, New York, 2009.
4 A. Pazy, Semigroups of Linear Operators and Applications to
Partial Differential Equations, Springer - Verlag, New York,
1983.

M. Thamban Nair Inverse Problem in PDE

Thank you for your attention

M. Thamban Nair Inverse Problem in PDE