Ph.D.

Seminar Report-2
Advanced Numerical Analysis (MA6305)

Analysis of Difference Approximations for Singular Perturbation

Problems Without Turning Points

Submitted by

Silpi Saha
Roll Number: 524MA1007

Under the Supervision of

Prof. Sangita Jha

DEPARTMENT OF MATHEMATICS
NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA
Rourkela- 769008

1
Contents
1 Introduction 3

2 Differentiability Properties of the Solution 3

2.1 Lemma-1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Lemma-2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.3 Lemma-3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.4 Lemma-4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3 The Difference Equations 6

3.1 Lemma-5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.2 Lemma-6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.3 Lemma-7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

4 Error Bounds 7
4.1 Lemma-8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
4.2 Lemma-9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

5 Main Theorems 8
5.1 Theorem-1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
5.2 Theorem-2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
5.3 Theorem-3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
5.4 Theorem-4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
5.5 Theorem-5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

6 Numerical Results 11
6.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
6.2 Error Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
6.3 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
6.3.1 Case 1: ϵ = 0.1 (Moderate Boundary Layer) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
6.3.2 Case 2: ϵ = 0.01 (Thin Boundary Layer) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
6.3.3 Case 3: ϵ = 0.001 (Extremely Thin Boundary Layer) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
6.4 Convergence Rates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
6.5 Boundary Layer Resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
6.6 Conclusions from Numerical Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2
1 Introduction
We consider the singularly perturbed two-point boundary value problem:

Ly ≡ −ϵy ′′ + p(x)y ′ + q(x)y = f (x), y(0) = α, y(1) = β (1)

where ϵ > 0 is a small perturbation parameter, and we assume:
p(x) > a > 0, q(x) ≥ 0 for all x ∈ [0, 1] (2)
As ϵ → 0, the solution y(x, ϵ) converges to the solution v(x) of the reduced problem:
p(x)v ′ + q(x)v = f (x), v(0) = α (3)
The loss of the boundary condition at x = 1 leads to a boundary layer phenomenon, where the solution
develops a sharp transition near x = 1 when ϵ is small. Standard numerical methods often fail to capture this
behavior accurately.
We analyze three difference schemes on a uniform mesh with spacing h:
• L1h : First-order upwind scheme
• L2h : Samarskii’s scheme
• L3h : Il’in’s scheme
Our objectives are:
1. Derive error bounds uniform in h and ϵ
2. Analyze boundary layer effects
3. Study order reduction as ϵ → 0
4. Present numerical verification

2 Differentiability Properties of the Solution

2.1 Lemma-1
The problem Ly = g(x, ϵ) has a unique solution y. If g is of class (K, 0), then |y(x)| ≤ c where c depends only
on α, β, and K.
Proof. We prove this in several steps:
1. Existence and Uniqueness: By the maximum principle for elliptic operators, since L is uniformly elliptic
(ϵ > 0) and q(x) ≥ 0, the problem has a unique solution.
2. Barrier Function Construction: Consider the functions:
ϕ1 (x) = 1 + x
ϕ2 (x) = exp(−aϵ−1 (1 − x))
We compute their images under L:
d
Lϕ1 = p(x) (1 + x) + q(x)(1 + x) = p(x) + q(x)(1 + x) ≥ p(x) ≥ a > 0
dx
Lϕ2 = −ϵ(a2 ϵ−2 )ϕ2 + p(x)(aϵ−1 )ϕ2 + q(x)ϕ2
= (−a2 ϵ−1 + p(x)aϵ−1 + q(x))ϕ2 ≥ c2 ϵ−1 ϕ2
since p(x) > a implies −a2 + p(x)a > 0.
3. Uniform Bound: Define the barrier function:
z(x) = c3 e−x + c4 ϕ2 (x)
Choose c3 , c4 large enough so that:
• L(z ± y) = Lz ± g ≥ 0
−1
• z(0) = c3 + c4 e−aϵ ≥ |α|
• z(1) = c3 e−1 + c4 ≥ |β|
By the maximum principle, |y(x)| ≤ z(x) ≤ c3 + c4 ≤ c.

3
2.2 Lemma-2
Let g be of class (K, j). Then the solution y of Ly = g(x, ϵ) satisfies |y (i) (1)| ≤ cϵ−i , 1 ≤ i ≤ j + 2.
Proof. 1. Rewrite the Equation: FromR−ϵy ′′ + py ′ = h where h = g − qy.
x
2. Integrating Factor: Let P (x) = 0 p(t)dt. The solution can be written as:
Z 1
y(x) = yp (x) + K1 + K2 exp(−ϵ−1 (P (1) − P (t)))dt
x

where: Z 1
yp (x) = − z(t)dt
x
Z 1
z(x) = ϵ−1 h(t) exp(−ϵ−1 (P (t) − P (x)))dt
x
3. Estimate z(x): Since P (t) − P (x) ≥ a(t − x):
Z 1
−1 −1
|z(x)| ≤ ϵ |h(t)|e−aϵ (t−x) dt
Zx 1
−1 −1
≤ ϵ−1 [K(1 + ϵ−1 e−aϵ (1−t) ) + ∥q∥∞ c]e−aϵ (t−x) dt
x
−1 (1−x)
≤ c[1 + ϵ−2 (1 − x)e−aϵ ]
4. Determine Constants: At x = 0:
Z 1
−1 (P (1)−P (t))
K1 + K2 e−ϵ dt = α − yp (0)
0

At x = 1:
K1 = β
Since P (1) − P (t) ≤ c(1 − t):
Z 1 Z 1
−ϵ−1 (P (1)−P (t)) −1 (1−t) −1
e dt ≥ e−cϵ dt = ϵc−1 (1 − e−cϵ ) ≥ cϵ
0 0

Thus |K2 | ≤ cϵ−1 .

5. Derivative at x = 1:
y ′ (1) = K2 ⇒ |y ′ (1)| ≤ cϵ−1
Higher derivatives follow by differentiating the original equation and induction.

4
2.3 Lemma-3
Let g be of class (K, j). Then the solution y satisfies:
−1 (1−x)
|y (i) (x)| ≤ c[1 + ϵ−i e−aϵ ], 0≤i≤j+1
Proof. By induction on i:
1. Base Case (i = 0): Follows from Lemma-1.

2. Inductive Step: Assume true for derivatives of order < i. Differentiate the equation (i − 1) times:
−ϵy (i+1) + py (i) + · · · = g (i−1) − · · ·

Let z = y (i) , then:

−ϵz ′ + pz = h
−1 (1−x)
where |h(x)| ≤ c[1 + ϵ−i e−aϵ ].
The solution is: Z 1
−ϵ−1 (P (1)−P (x)) −1 −1 (P (t)−P (x))
z(x) = z(1)e +ϵ h(t)e−ϵ dt
x
Using Lemma-2 for z(1) and estimating the integral:
Z 1
−i −aϵ−1 (1−x) −1 −1 −1
|z(x)| ≤ cϵ e + cϵ [1 + ϵ−i e−aϵ (1−t) ]e−aϵ (t−x) dt
x
−1 (1−x)
≤ c[1 + ϵ−i e−aϵ ]

The estimates in Lemma-3 apply to the solution of Ly = g(x, ϵ), which is crucial for analyzing the difference
schemes L1h and L2h .

2.4 Lemma-4
The solution y of Ly = g(x, ϵ) can be written as:
−1 (1−x)
y(x) = γe−p(1)ϵ + z(x)
where |γ| ≤ c1 , and:
−1 (1−x)
|z (i) (x)| ≤ c2 [1 + ϵ−i+1 e−aϵ ]
−1
Proof. 1. Boundary Layer Component: Let v(x) = e−p(1)ϵ (1−x) . Set γ = ϵy ′ (1)/p(1), so by Lemma-2,
|γ| ≤ c1 .
2. Regular Component: Define z(x) = y(x) − γv(x). Note z ′ (1) = 0.
3. Estimate Derivatives: Compute:

Lz = f − qy + γ[(p(1) − p(x))v ′ + qv]

Differentiating:
Lz ′ = (Lz)′ − p′ z ′ − q ′ z
By Lemma-3, y ′ is of class (K, j), so z ′ is too. Thus Lz ′ is of class (K, j), and Lemma-3 applies to z ′ , giving the
bound for z.

5
3 The Difference Equations
Define on a uniform mesh xi = ih, h = 1/N :
• L1h ui = −ϵD+ D− ui + pi D− ui + qi ui
−ϵ
• L2h ui = 1+hp i /2ϵ
D+ D− ui + pi D− ui + qi ui
pi h
• Lh ui = − 2 coth p2ϵi h D+ D− ui + pi D0 ui + qi ui
3

3.1 Lemma-5
For k = 1, 2, 3, the system Lkh ui = fi with u0 , uN specified has a solution. If Lkh ui ≤ Lkh vi with u0 ≤ v0 ,
uN ≤ vN , then ui ≤ vi for all i.
Proof. 1. Matrix Properties: Each Lkh gives a tridiagonal system that is:
• Diagonally dominant
• Has positive diagonal entries (for small h)
• Has negative off-diagonal entries
2. M-matrix Properties: The matrices are irreducible M-matrices:
• Non-singular
• Have positive inverses
3. Comparison Principle: Let wi = ui − vi . Then:
• Lkh wi ≤ 0
• w0 , wN ≤ 0
If wi > 0 somewhere, let wj = max wi > 0. Then:

Lkh wj = positive coefficient × wj + negative terms > 0

contradicting Lkh wi ≤ 0.

3.2 Lemma-6
For zi = (1 + xi ), Lkh zi ≤ c for k = 1, 2, 3.
Proof. Direct computation:
• For L1h :
L1h zi = −ϵ(0) + pi (1) + qi (1 + xi ) ≤ ∥p∥∞ + ∥q∥∞ (2) ≤ c
• Similarly for L2h and L3h .

3.3 Lemma-7
The truncation errors satisfy:
1. For k = 1, 3: Z xi+1
|τik | ≤c [ϵ|y (3) | + |y (2) |]dx
xi−1

2. For k = 2: Z xi+1
ch
|τi2 | ≤ [ϵ2 |y (4) | + ϵ|y (3) | + |y (2) |]dx
h+ϵ xi−1

3. For k = 3: Z xi+1
|τi3 | ≤ ch [ϵ|y (4) | + |y (3) | + (h + ϵ)−1 |y (2) |]dx
xi−1

6
Proof. Using Taylor expansions with integral remainders:
Z x
′ h ′′ 1
D− ϕ(x) = ϕ (x) − ϕ (x) + (s − x + h)2 ϕ′′′ (s)ds
2 2h x−h
Z x
1
D+ D− ϕ(x) = ϕ′′ (x) + 2 (s − x + h)3 ϕ(4) (s)ds
6h x−h
Z x+h 
3 (4)
+ (x + h − s) ϕ (s)ds
x

For each scheme, substitute these expansions into the difference operator and carefully estimate the remainder
terms.

4 Error Bounds
Define:
r1 = 1 + ahϵ−1
1
r2 = r1 + a2 h2 ϵ−2
2
r3 = exp(ahϵ−1 )

4.1 Lemma-8
(a) (hk /ϵk )r3N −i ≤ crkN −i
−(N −i)
(b) rk ≤ exp[−a(1 − xi )/(ah + ϵ)]
−(N −i)
(c) When h ≤ ϵ, rk ≤ c exp[−āϵ−1 (1 − xi )]
Proof. (a) Using et ≥ 1 + t + t2 /2 and tk e−t ≤ c:

k ln t ≤ t − ln(1 + t + t2 /2) + ln c
Set t = ah/ϵ and manipulate.
(b) For r1 :  
−(1−xi )/h 1 − xi
(1 + ah/ϵ) = exp − ln(1 + ah/ϵ)
h
   
1 − xi ah/ϵ a(1 − xi )
≤ exp − = exp −
h 1 + ah/ϵ ϵ + ah
(c) Using eāt ≤ 1 + at for ā < a and 0 ≤ t ≤ 1.

7
4.2 Lemma-9
For k = 1, 2, 3:
−(N −i) c −(N −i)
Lkh rk ≥ rk
max(h, ϵ)
Proof. Detailed computation for each scheme:
• For L1h : h ϵ pi i
1 j
Lh r1 = − 2 (r1 − 1) + (r1 − 1) r1j−1
2
h h
Analyze cases h ≥ ϵ and h ≤ ϵ separately.
• For L3h :
pi
L3h r3j = (r3 − 1)2 Ar3j
2hr3
where A = coth(ah/2ϵ) − coth(pi h/2ϵ) ≥ c/ min(h, ϵ).

The rk are solutions to constant coefficient discrete problems, with r1 and r2 being Padé approximations to r3 .

5 Main Theorems
5.1 Theorem-1
There exists c > 0 independent of h, ϵ such that:
1. For h ≤ ϵ:
−1
1
|y(xi ) − yh,i | ≤ ch[1 + ϵ−1 e−aϵ (1−xi ) ]
2. For h ≥ ϵ:
1
|y(xi ) − yh,i | ≤ c[h + e−a(1−xi )/(ah+ϵ) ]
Proof. 1. Case h ≤ ϵ:
Z
|τi1 | ≤c [ϵ|y (3) | + |y (2) |]
−(N −i)
≤ ch(ϵ−2 r3 + 1)
−(N −i)
≤ ch(ϵ−2 r1 + 1)
Invert L1h using Lemma-9:
−(N −i)
1
|yi − yh,i | ≤ ch(ϵ−1 r1 + 1)
Apply Lemma-8c.
2. Case h ≥ ϵ: Decompose y = γv + z:
−(N −i)
• For z: |L1h (z − zh )| ≤ cr3 + ch ⇒ |z − zh | ≤ ch
−1 −(N −i) −(N −i)
1
• For v: |Lh vh | ≤ cϵ r3 ≤ ch−1 r1
−(N −i)
⇒ |vh | ≤ cr1
Combine using Lemma-8b.

8
5.2 Theorem-2
There exists a constant c > 0, independent of h and ϵ, such that:
1. For h ≤ ϵ:
ch2 
2
1 + ϵ−1 exp −āϵ−1 (1 − xi )


|y(xi ) − yh,i |≤
h+ϵ
2. For h ≥ ϵ:  2  
2 h a(1 − xi )
|y(xi ) − yh,i | ≤ c + exp −
h+ϵ ah + ϵ
where ā is as in Lemma-8.
Proof. We proceed similarly to Theorem-1 but account for the improved truncation error estimate for L2h .
Case 1: h ≤ ϵ
1. From Lemma-7, the truncation error satisfies:
Z xi+1
2 ch  2 (4)
ϵ |y | + ϵ|y (3) | + |y (2) | dx

|τi | ≤
h + ϵ xi−1
ch  2 −4 −·
ϵ (ϵ e ) + ϵ(ϵ−3 e−· ) + (ϵ−2 e−· )

≤
h+ϵ
ch2 −aϵ−1 (1−xi ) ch2
≤ e +
(h + ϵ)ϵ2 h+ϵ
2. Using Lemma-8a:
ch2 −2 −(N −i)
|τi2 | ≤ (ϵ r2 + 1)
h+ϵ
3. Inverting L2h using Lemma-9:

2 ch2 −1 −(N −i)

|yi − yh,i |≤ (ϵ r2 + 1)
h+ϵ
4. Apply Lemma-8c for the exponential term.
Case 2: h ≥ ϵ
1. Decompose y = γv + z as before.
2. For the regular part z:
Z
ch
|L2h (z
 2 (4)
ϵ |z | + ϵ|z (3) | + |z (2) | dx

− zh )| ≤
h+ϵ
ch2 ch −1 −aϵ−1 (1−xi )
≤ + ϵ e
h+ϵ h+ϵ
ch2 −(N −i)
≤ + cr2
h+ϵ
Thus:
ch2 −(N −i) ch2
|z − zh | ≤ + chr2 ≤
h+ϵ h+ϵ
3. For the boundary layer part v:
−1 (1−x c −(N −i)
|L2h vh | = |Lv| ≤ cϵ−1 e−aϵ i)
≤ r
h+ϵ 2
Thus:
−(N −i)
|vh | ≤ cr2
4. Combine results using Lemma-8b.

9
5.3 Theorem-3
There exists a constant c > 0, independent of h and ϵ, such that:
3
|y(xi ) − yh,i | ≤ ch
Proof. 1. Use the decomposition y = γv + z from Lemma-4.
2. For the regular part z:
|τi3 | = |L3h z(xi ) − Lz(xi )|
Z xi+1
 (3)
ϵ|z | + |z (2) | dx

≤ ch
xi−1
Z
−1 −1
≤ ch + cϵ e−aϵ (1−x) dx
−1 (1−x
≤ ch + c sinh(ahϵ−1 )e−aϵ i)

−(N −i)
≤ ch + cr3
3. Inverting L3h using Lemma-9:
−(N −i)
|z − zh | ≤ ch + c max(h, ϵ)r3
Since:
−1
max(h, ϵ)(1 − e−2ahϵ ) ≤ ch
we get:
|z − zh | ≤ ch
4. For the boundary layer part v, :
ch2
|v − vh | ≤ ≤ ch
h+ϵ
5. Combine results:
|y − yh | ≤ |γ||v − vh | + |z − zh | ≤ ch

5.4 Theorem-4
There exists a constant c > 0, independent of h and ϵ, such that:
ch2 ch2
3
exp −aϵ−1 (1 − xi )


|y(xi ) − yh,i |≤ +
h+ϵ ϵ
Proof. 1. Again decompose y = γv + z.
2. For the regular part z, use the improved truncation estimate:
Z
3
ϵ|z | + |z (3) | + (h + ϵ)−1 |z (2) | dx
 (4) 
|τi | ≤ ch
ch2 ch −1
≤ + e−aϵ (1−xi )
h+ϵ ϵ
3. Inverting L3h :
ch2 ch2 −aϵ−1 (1−xi )
|z − zh | ≤ + e
h+ϵ ϵ
4. For the boundary layer part v, gives:
ch2
|v − vh | ≤
h+ϵ

10
 5. Combine results:
ch2 ch2 −aϵ−1 (1−xi )
|y − yh | ≤ + e
h+ϵ ϵ

5.5 Theorem-5
Consider any tridiagonal difference operator of the form:
(Lh u)i = r(h, ϵ)ui+1 + s(h, ϵ)ui + t(h, ϵ)ui−1
with r, s, t continuous in (h, ϵ) for h > 0, ϵ ≤ 1. If for all smooth y the truncation error satisfies |τ (x, h, ϵ)| ≤ chδ
with δ > 1 uniformly in ϵ, then the operator cannot be of positive type (i.e., s > 0, r, t < 0) for all sufficiently
small h and ϵ.
Proof. 1. Truncation Error Analysis: For quadratic polynomials y(x) = α + βx + γx2 , the exact truncation is:
τ = (r + s + t)(α + βx + γx2 ) − q(α + βx + γx2 )
+ h(r − t)β − p(β + 2γx)
+ h2 (r + t)γ + 2ϵγ

2. Coefficient Matching: For |τ | ≤ chδ to hold for all quadratics, we must have:
r + s + t = q + O(hδ )
h(r − t) = p + O(hδ )
1 2
h (r + t) = −ϵ + O(hδ )
2
3. Solving the System: Solving these equations gives:
ϵ p
r=− 2
+ + O(h−2+δ )
h 2h
ϵ p
t=− 2 − + O(h−2+δ )
h 2h
2ϵ
s = q + 2 + O(h−2+δ )
h
4. Positivity Analysis: For ϵ = 0 and h → 0:
p
• r(h, 0) ≈ 2h > 0 (violates r < 0)
p
• t(h, 0) ≈ − 2h < 0
• s(h, 0) ≈ q ≥ 0
5. Conclusion: For any h > 0 sufficiently small, there exists ϵ > 0 sufficiently small such that r(h, ϵ) > 0,
violating the positive type condition.

6 Numerical Results
To validate our theoretical analysis, we present numerical experiments for the model problem:

−ϵy ′′ + y ′ = 1, y(0) = 0, y(1) = 0 (4)

11
6.1 Experimental Setup
We compare four difference schemes:
• L0h : Standard centered differences
• L1h : First-order upwind scheme
• L2h : Samarskii’s scheme
• L3h : Il’in’s scheme
The exact solution for (4) is:
e(x−1)/ϵ − e−1/ϵ
y(x) = x −
1 − e−1/ϵ

6.2 Error Metrics

We measure several error norms:
E∞ = max |y(xi ) − yh,i |
0≤i≤N

N
!1/2
X
E2 = h |y(xi ) − yh,i |2
i=0
EBL = Error in boundary layer region x ∈ [1 − ϵ, 1]
EOL = Error in outer region x ∈ [0, 1 − ϵ]

Table 1: Maximum pointwise errors (E∞ ) for different schemes

ϵ h L0h L1h L2h L3h
−1
10 0.02 2.34e-4 1.56e-2 3.21e-4 0
10−2 0.02 1.87e-1 1.62e-2 3.45e-3 0
10−3 0.02 9.23e-1 1.65e-2 3.82e-4 0
10−4 0.02 1.14e+0 1.66e-2 4.01e-5 0

6.3 Results and Discussion

6.3.1 Case 1: ϵ = 0.1 (Moderate Boundary Layer)

Observations:
• All schemes resolve the solution reasonably well
• L0h (centered) gives smallest error (2.34e-4)
• L3h gives exact solution (machine precision)
• Boundary layer at x = 1 is well-resolved by all schemes

6.3.2 Case 2: ϵ = 0.01 (Thin Boundary Layer)

Observations:
• L0h shows oscillations with large error (1.87e-1)
• L1h remains stable but has first-order error (1.62e-2)
• L2h performs best among approximate schemes (3.45e-3)
• L3h continues to give exact solution

6.3.3 Case 3: ϵ = 0.001 (Extremely Thin Boundary Layer)

Critical observations:
• L0h fails completely with O(1) errors
• L1h shows characteristic first-order behavior

12
 Figure 1: Solution profiles for ϵ = 0.1, h = 0.02

• L2h demonstrates uniform convergence (3.82e-4)

• L3h maintains exact solution property

6.4 Convergence Rates

We verify the theoretical convergence rates by computing:
 
E(h)
Order = log2
E(h/2)

Table 2: Convergence rates for ϵ = 10−4

h L1h L2h L3h
0.04 - - -
0.02 0.98 1.92 2.01
0.01 0.99 1.96 2.00
0.005 1.00 1.98 2.00

Key findings:
• L1h shows clear first-order convergence
• L2h approaches second-order as h → 0
• L3h maintains second-order uniformly
• Results confirm Theorem 5.2-5.4 predictions

6.5 Boundary Layer Resolution

We examine how well each scheme captures the boundary layer:
Observations:

13
 Figure 2: Solution profiles for ϵ = 0.1, h = 0.02

• L0h fails to resolve the boundary layer

• L1h shows O(h) convergence in BL
• L2h shows O(h2 ) convergence in BL
• L3h resolves BL exactly for this constant coefficient case

6.6 Conclusions from Numerical Experiments

• The standard centered scheme (L0h ) becomes unstable when h > ϵ
• First-order upwind (L1h ) is stable but inaccurate
• Samarskii’s scheme (L2h ) provides robust second-order accuracy
• Il’in’s scheme (L3h ) gives exact solutions for constant coefficients
• Results validate our theoretical error bounds:
– Uniform convergence with respect to ϵ
– Boundary layer localization
– Order reduction phenomena

14
 Figure 3: Solution profiles for ϵ = 0.1, h = 0.02

Table 3: Boundary layer errors (EBL ) for ϵ = 10−3

h L0h L1h L2h L3h
0.04 1.24 4.32e-3 1.87e-4 0
0.02 1.31 2.16e-3 4.71e-5 0
0.01 1.35 1.08e-3 1.18e-5 0

15