Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation...

45
Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion Quadratic C 1 -spline collocation for reaction-diffusion problems Torsten Linss 1 Goran Radojev 2 Helena Zarin 2 1 Fakultät für Mathematik und Informatik, FernUniversität in Hagen, Germany 2 Department of Mathematics and Informatics, University of Novi Sad, Serbia "Numerical analysis for Singularly Perturbed Problems" Workshop dedicated to the 60th birthday of Prof. Martin Stynes TU Dresden, November 16-18, 2011

Transcript of Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation...

Page 1: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Quadratic C1-spline collocation forreaction-diffusion problems

Torsten Linss1 Goran Radojev2 Helena Zarin2

1Fakultät für Mathematik und Informatik, FernUniversität in Hagen, Germany2Department of Mathematics and Informatics, University of Novi Sad, Serbia

"Numerical analysis for Singularly Perturbed Problems"Workshop dedicated to the 60th birthday of Prof. Martin Stynes

TU Dresden, November 16-18, 2011

Page 2: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Outline

Introduction (problem, idea)Layer-adapted meshInterpolation errorCollocation method

StabilityMaximum-norm a priori error boundMaximum-norm a posteriori error boundAn adaptive algorithm

Numerical experimentsConclusion

Page 3: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Reaction-diffusion problem

Reaction-diffusion problemLu := −ε2u′′ + ru = f in (0,1),

u(0) = γ0, u(1) = γ1

ε ∈ (0,1], r ≥ %2 > 0 on [0,1]

(1)

0.2 0.4 0.6 0.8 1.0 1.2

0.5

1.0

1.5

¶ = 10-1

¶ = 10-2

Page 4: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Reaction-diffusion problem

Reaction-diffusion problemLu := −ε2u′′ + ru = f in (0,1),

u(0) = γ0, u(1) = γ1

ε ∈ (0,1], r ≥ %2 > 0 on [0,1]

(1)

0.2 0.4 0.6 0.8 1.0 1.2

0.5

1.0

1.5

¶ = 10-1

¶ = 10-2

Page 5: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Idea

C. de Boor, B. Swartz (SIAM J. Numer. Anal, 1973):general theory for spline-collocation methods applied toclassical (not SP) BVPs

transition to SPBVPs

bounds with “constants” that tend to infinity when ε→ 0

different approach

quadratic C1-splines on a special modified Shishkin mesh

Surla, Uzelac (ZAMM, 1997):quadratic C1-spline collocation, layer-adapted mesh, nodalbasis, mesh points as dof’s⇒ O(N−2 ln2 N)

LRZ (submitted to NA, 2011):B-spline basis⇒ O(N−2 ln2 N)

Page 6: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Idea

C. de Boor, B. Swartz (SIAM J. Numer. Anal, 1973):general theory for spline-collocation methods applied toclassical (not SP) BVPs

transition to SPBVPs

bounds with “constants” that tend to infinity when ε→ 0

different approach

quadratic C1-splines on a special modified Shishkin mesh

Surla, Uzelac (ZAMM, 1997):quadratic C1-spline collocation, layer-adapted mesh, nodalbasis, mesh points as dof’s⇒ O(N−2 ln2 N)

LRZ (submitted to NA, 2011):B-spline basis⇒ O(N−2 ln2 N)

Page 7: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Idea

C. de Boor, B. Swartz (SIAM J. Numer. Anal, 1973):general theory for spline-collocation methods applied toclassical (not SP) BVPs

transition to SPBVPs

bounds with “constants” that tend to infinity when ε→ 0

different approach

quadratic C1-splines on a special modified Shishkin mesh

Surla, Uzelac (ZAMM, 1997):quadratic C1-spline collocation, layer-adapted mesh, nodalbasis, mesh points as dof’s⇒ O(N−2 ln2 N)

LRZ (submitted to NA, 2011):B-spline basis⇒ O(N−2 ln2 N)

Page 8: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Idea

C. de Boor, B. Swartz (SIAM J. Numer. Anal, 1973):general theory for spline-collocation methods applied toclassical (not SP) BVPs

transition to SPBVPs

bounds with “constants” that tend to infinity when ε→ 0

different approach

quadratic C1-splines on a special modified Shishkin mesh

Surla, Uzelac (ZAMM, 1997):quadratic C1-spline collocation, layer-adapted mesh, nodalbasis, mesh points as dof’s⇒ O(N−2 ln2 N)

LRZ (submitted to NA, 2011):B-spline basis⇒ O(N−2 ln2 N)

Page 9: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Idea

C. de Boor, B. Swartz (SIAM J. Numer. Anal, 1973):general theory for spline-collocation methods applied toclassical (not SP) BVPs

transition to SPBVPs

bounds with “constants” that tend to infinity when ε→ 0

different approach

quadratic C1-splines on a special modified Shishkin mesh

Surla, Uzelac (ZAMM, 1997):quadratic C1-spline collocation, layer-adapted mesh, nodalbasis, mesh points as dof’s⇒ O(N−2 ln2 N)

LRZ (submitted to NA, 2011):B-spline basis⇒ O(N−2 ln2 N)

Page 10: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Properties of the exact solution

The Green’s function

‖rG(x , ·)‖1 ≤ 1, ‖Gξ(x , ·)‖1 ≤ (%ε)−1 , ‖Gξξ(x , ·)‖1 ≤ 2ε−2

Lemma (Derivative bounds)

Let r , f ∈ C4[0,1]. Then∣∣u(k)(x)∣∣ ≤ C

{1 + ε−k e−%x/ε + ε−k e−%(1−x)/ε

},

for x ∈ (0,1), k = 0, . . . ,4. Furthermore, the solution can bedecomposed as u = v + w0 + w1. For k = 0, . . . ,4, the regularsolution component v satisfies

∥∥v (k)∥∥∞ ≤ C, while for the layer

parts w0 and w1 we have∣∣w (k)0 (x)

∣∣ ≤ Cε−k e−%x/ε,∣∣w (k)

1 (x)∣∣ ≤ Cε−k e−%(1−x)/ε, x ∈ [0,1].

Page 11: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Properties of the exact solution

The Green’s function

‖rG(x , ·)‖1 ≤ 1, ‖Gξ(x , ·)‖1 ≤ (%ε)−1 , ‖Gξξ(x , ·)‖1 ≤ 2ε−2

Lemma (Derivative bounds)

Let r , f ∈ C4[0,1]. Then∣∣u(k)(x)∣∣ ≤ C

{1 + ε−k e−%x/ε + ε−k e−%(1−x)/ε

},

for x ∈ (0,1), k = 0, . . . ,4. Furthermore, the solution can bedecomposed as u = v + w0 + w1. For k = 0, . . . ,4, the regularsolution component v satisfies

∥∥v (k)∥∥∞ ≤ C, while for the layer

parts w0 and w1 we have∣∣w (k)0 (x)

∣∣ ≤ Cε−k e−%x/ε,∣∣w (k)

1 (x)∣∣ ≤ Cε−k e−%(1−x)/ε, x ∈ [0,1].

Page 12: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Smoothed Shishkin mesh

Shishkin-mesh transition point

λ := min{σε

%ln N,q

}, q ∈ (0,1/2), σ > 0

The mesh ∆ : x0 < x1 < · · · < xN is generated by xi = ϕ(i/N)with the mesh generating function

ϕ(t) :=

λq t t ∈ [0,q],

κ(t) := p(t − q)3 + λq t t ∈ [q,1/2],

1− ϕ(1− t) t ∈ [1/2,1],

where p is chosen such that ϕ(1/2) = 1/2.Note, that ϕ ∈ C1[0,1] with ‖ϕ′‖∞, ‖ϕ′′‖∞ ≤ C.

Page 13: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Smoothed Shishkin mesh

Shishkin-mesh transition point

λ := min{σε

%ln N,q

}, q ∈ (0,1/2), σ > 0

The mesh ∆ : x0 < x1 < · · · < xN is generated by xi = ϕ(i/N)with the mesh generating function

ϕ(t) :=

λq t t ∈ [0,q],

κ(t) := p(t − q)3 + λq t t ∈ [q,1/2],

1− ϕ(1− t) t ∈ [1/2,1],

where p is chosen such that ϕ(1/2) = 1/2.Note, that ϕ ∈ C1[0,1] with ‖ϕ′‖∞, ‖ϕ′′‖∞ ≤ C.

Page 14: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

The interpolation error for piecewise quadratic splines

Notation:

There midpoints of the mesh intervals Ji := [xi−1, xi ] aredenoted with

xi−1/2 := (xi−1 + xi) /2 = xi−1 + hi/2, i = 1, . . . ,N.

For, m, ` ∈ N, m < `, let

Sm` (∆) :=

{s ∈ Cm[0,1] : s|Ji ∈ Π`, for i = 1, . . . ,N

}.

Page 15: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

S02 -interpolation

Given an arbitrary function g ∈ C0[0,1], consider theinterpolation problem of finding I0

2g ∈ S02 (∆) with(

I02g)

i = gi , i = 0, . . . ,N, and(I02g)

i−1/2 = gi−1/2, i = 1, . . . ,N.

Theorem 1

Assume r , f ∈ C4[0,1]. Then the interpolation error u − I02u for

the solution of (1) on a smoothed Shishkin mesh with σ ≥ 3satisfies ∥∥u − I0

2u∥∥∞ ≤ CN−3 ln3 N,

ε2 maxi=1,...,N

∣∣∣(u − I02u)′′

i−1/2

∣∣∣ ≤ CN−2 ln2 N.

Page 16: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

S02 -interpolation

Given an arbitrary function g ∈ C0[0,1], consider theinterpolation problem of finding I0

2g ∈ S02 (∆) with(

I02g)

i = gi , i = 0, . . . ,N, and(I02g)

i−1/2 = gi−1/2, i = 1, . . . ,N.

Theorem 1

Assume r , f ∈ C4[0,1]. Then the interpolation error u − I02u for

the solution of (1) on a smoothed Shishkin mesh with σ ≥ 3satisfies ∥∥u − I0

2u∥∥∞ ≤ CN−3 ln3 N,

ε2 maxi=1,...,N

∣∣∣(u − I02u)′′

i−1/2

∣∣∣ ≤ CN−2 ln2 N.

Page 17: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

S12 -interpolation

Given an arbitrary function g ∈ C0[0,1], consider theinterpolation problem of finding I1

2g ∈ S12 (∆) with(

I12g)

0 = g0,(I12g)

i−1/2 = gi−1/2, i = 1, . . . ,N,(I12g)

N = gN .

Theorem 2

Assume r , f ∈ C4[0,1]. Then the interpolation error u − I12u for

the solution u of (1) on a smoothed Shishkin mesh with σ ≥ 4satisfies

maxi=0,...,N

∣∣∣(u − I12u)

i

∣∣∣ ≤ CN−4 ln4 N,∥∥u − I12u∥∥∞ ≤ CN−3 ln3 N,

ε2 maxi=1,...,N

∣∣∣(u − I12u)′′

i−1/2

∣∣∣ ≤ CN−2 ln2 N.

Page 18: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

S12 -interpolation

Given an arbitrary function g ∈ C0[0,1], consider theinterpolation problem of finding I1

2g ∈ S12 (∆) with(

I12g)

0 = g0,(I12g)

i−1/2 = gi−1/2, i = 1, . . . ,N,(I12g)

N = gN .

Theorem 2

Assume r , f ∈ C4[0,1]. Then the interpolation error u − I12u for

the solution u of (1) on a smoothed Shishkin mesh with σ ≥ 4satisfies

maxi=0,...,N

∣∣∣(u − I12u)

i

∣∣∣ ≤ CN−4 ln4 N,∥∥u − I12u∥∥∞ ≤ CN−3 ln3 N,

ε2 maxi=1,...,N

∣∣∣(u − I12u)′′

i−1/2

∣∣∣ ≤ CN−2 ln2 N.

Page 19: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Let ∆ be an arbitrary partition of [0,1]. Our discretisation is:Find u∆ ∈ S1

2 (∆) such that

u∆,0 = γ0,(Lu∆

)i−1/2 = fi−1/2, i = 1, . . . ,N, u∆,N = γ1.

Let {ϕi}N+1i=0 be the B-spline basis in S1

2 (∆). Then we mayrepresent u∆ as

u∆ :=N+1∑k=0

αkϕk ,

where the αk are determined by collocation.

Page 20: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Collocation equation is equivalent to

α0 = γ0, [Lα]i−1/2 = fi−1/2, i = 1, . . . ,N, αN+1 = γ1

with α := (α0, . . . , αN+1)T ∈ RN+2 and

[Lα]i−1/2 := − ε2[

2(αi+1 − αi)

hi(hi + hi+1)− 2(αi − αi−1)

hi(hi−1 + hi)

]+ ri−1/2

[q+

i αi+1 +(1− q+

i − q−i)αi + q−i αi−1

],

q+i :=

hi

4(hi + hi+1)and q−i :=

hi

4(hi + hi−1),

for i = 1, . . . ,N and h0 = hN+1 = 0.

Page 21: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Stability

The operator L is not inverse monotone. Nonetheless, we canestablish its maximum-norm stability.

Theorem 3Assume, there exists a constant κ > 0 such that h1 ≥ κh2,hN ≥ κhN−1 and

max{

hi+1,hi−1}≥ κhi , i = 2, . . . ,N − 1.

Then the operator L is maximum-norm stable with

‖γ‖∞ := maxi=1,...,N

|γi | ≤2(1 + κ)

κmax

i=1,...,N

∣∣∣∣∣ [Lγ]i−1/2

ri−1/2

∣∣∣∣∣≤ 2(1 + κ)

κ%2 ‖Lγ‖∞ ,

for all γ ∈ RN+20 :=

{v ∈ RN+2 : v0 = vN+1 = 0

}.

Page 22: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Maximum-norm a priori error bound

Theorem 4Let u be the solution of (1) and u∆ its approximation by thecollocation method on a smoothed Shishkin mesh with σ ≥ 4. Ifassumptions of Theorems 2 and 3 hold true, then

‖u − u∆‖∞ ≤ CN−2 ln2 N.

Proof.

‖u − u∆‖∞ ≤ ‖u − I12u‖∞ + ‖I1

2u − u∆‖∞

≤ CN−3 ln3 N + ‖α− β‖∞ , I12u =

N+1∑k=0

βkϕk

‖α− β‖∞ ≤ C maxi=1,...,N

∣∣∣[L (α− β)]i−1/2

∣∣∣= C max

i=1,...,N

∣∣∣ε2(u − I12u)′′

i−1/2

∣∣∣ ≤ CN−2 ln2 N �

Page 23: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Maximum-norm a priori error bound

Theorem 4Let u be the solution of (1) and u∆ its approximation by thecollocation method on a smoothed Shishkin mesh with σ ≥ 4. Ifassumptions of Theorems 2 and 3 hold true, then

‖u − u∆‖∞ ≤ CN−2 ln2 N.

Proof.

‖u − u∆‖∞ ≤ ‖u − I12u‖∞ + ‖I1

2u − u∆‖∞

≤ CN−3 ln3 N + ‖α− β‖∞ , I12u =

N+1∑k=0

βkϕk

‖α− β‖∞ ≤ C maxi=1,...,N

∣∣∣[L (α− β)]i−1/2

∣∣∣= C max

i=1,...,N

∣∣∣ε2(u − I12u)′′

i−1/2

∣∣∣ ≤ CN−2 ln2 N �

Page 24: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Maximum-norm a priori error bound

Theorem 4Let u be the solution of (1) and u∆ its approximation by thecollocation method on a smoothed Shishkin mesh with σ ≥ 4. Ifassumptions of Theorems 2 and 3 hold true, then

‖u − u∆‖∞ ≤ CN−2 ln2 N.

Proof.

‖u − u∆‖∞ ≤ ‖u − I12u‖∞ + ‖I1

2u − u∆‖∞

≤ CN−3 ln3 N + ‖α− β‖∞ , I12u =

N+1∑k=0

βkϕk

‖α− β‖∞ ≤ C maxi=1,...,N

∣∣∣[L (α− β)]i−1/2

∣∣∣= C max

i=1,...,N

∣∣∣ε2(u − I12u)′′

i−1/2

∣∣∣ ≤ CN−2 ln2 N �

Page 25: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Maximum-norm a posteriori error bounds

Theorem 5Let u be the solution of (1) and u∆ its approximation by thecollocation method on an arbitrary mesh ∆. Then

‖u − u∆‖∞ ≤ η(ru∆ − f ,∆)

with η(q,∆) := ηI(q,∆) + η3(q,∆) + η4(q,∆) and

ηI(q,∆) := ‖(I02q − q)/r‖∞,

η3(q,∆) :=2%2 max

i=1,...,N

[max

{|qi − qi−1/2|, |qi−1/2 − qi−1|

}·min {1,hi%/(4ε)}] ,

η4(q,∆) := maxi=1,...,N

|qi−1 − 2qi−1/2 + qi |4%2 .

Page 26: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Maximum-norm a posteriori error bounds

Proof. q := ru∆ − f

(u − u∆) (x) =

∫ 1

0G(x , ξ)

(L(u − u∆)

)(ξ) dξ

=N∑

i=1

∫Ji

G(x , ξ)[qi−1/2 − q(ξ)

]dξ

=

∫ 1

0

(I02q − q

)(ξ)G(x , ξ) dξ

+N∑

i=1

∫Ji

G(x , ξ)[qi−1/2 −

(I02q)(ξ)]

Page 27: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Maximum-norm a posteriori error bounds

Proof. q := ru∆ − f

(u − u∆) (x) =

∫ 1

0G(x , ξ)

(L(u − u∆)

)(ξ) dξ

=N∑

i=1

∫Ji

G(x , ξ)[qi−1/2 − q(ξ)

]dξ

=

∫ 1

0

(I02q − q

)(ξ)G(x , ξ) dξ

+N∑

i=1

∫Ji

G(x , ξ)[qi−1/2 −

(I02q)(ξ)]

Page 28: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Maximum-norm a posteriori error bounds

Proof. q := ru∆ − f

(u − u∆) (x) =

∫ 1

0G(x , ξ)

(L(u − u∆)

)(ξ) dξ

=N∑

i=1

∫Ji

G(x , ξ)[qi−1/2 − q(ξ)

]dξ

=

∫ 1

0

(I02q − q

)(ξ)G(x , ξ) dξ

+N∑

i=1

∫Ji

G(x , ξ)[qi−1/2 −

(I02q)(ξ)]

Page 29: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Maximum-norm a posteriori error bounds

1) ∣∣∣∣∣∫ 1

0

(I02q − q

)(ξ)G(x , ξ) dξ

∣∣∣∣∣ ≤ ηI(q,∆)

2)N∑

i=1

∫Ji

G(x , ξ)[qi−1/2 −

(I02q)(ξ)]

dξ =N∑

i=1

Ii

whereIi =

∫Ji

G(x , ξ)(ξ − xi−1/2)Ri(ξ) dξ

and

Ri(ξ) =qi − qi−1

hi+ 2(ξ − xi−1/2

)qi−1 − 2qi−1/2 + qi

h2i

Page 30: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Maximum-norm a posteriori error bounds

1) ∣∣∣∣∣∫ 1

0

(I02q − q

)(ξ)G(x , ξ) dξ

∣∣∣∣∣ ≤ ηI(q,∆)

2)N∑

i=1

∫Ji

G(x , ξ)[qi−1/2 −

(I02q)(ξ)]

dξ =N∑

i=1

Ii

whereIi =

∫Ji

G(x , ξ)(ξ − xi−1/2)Ri(ξ) dξ

and

Ri(ξ) =qi − qi−1

hi+ 2(ξ − xi−1/2

)qi−1 − 2qi−1/2 + qi

h2i

Page 31: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Maximum-norm a posteriori error bounds

2a)

|Ii | ≤hi

2‖Ri‖∞,Ji

∫Ji

G(x , ξ) dξ

2b)

|Ii | ≤hi

8

{‖Ri‖∞,Ji

∫Ji

|Gξ(x , ξ)| dξ +∥∥R′i

∥∥∞,Ji

∫Ji

G(x , ξ) dξ}

2a,b) ∣∣∣∣∣N∑

i=1

Ii

∣∣∣∣∣ ≤ 2%2 max

i=1,...,N

hi

2‖Ri‖∞,Ji

min{

1,hi%

}

+1%2 max

i=1,...,N

h2i

8∥∥R′i

∥∥∞,Ji

= η3(q,∆) + η4(q,∆) �

Page 32: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Maximum-norm a posteriori error bounds

2a)

|Ii | ≤hi

2‖Ri‖∞,Ji

∫Ji

G(x , ξ) dξ

2b)

|Ii | ≤hi

8

{‖Ri‖∞,Ji

∫Ji

|Gξ(x , ξ)| dξ +∥∥R′i

∥∥∞,Ji

∫Ji

G(x , ξ) dξ}

2a,b) ∣∣∣∣∣N∑

i=1

Ii

∣∣∣∣∣ ≤ 2%2 max

i=1,...,N

hi

2‖Ri‖∞,Ji

min{

1,hi%

}

+1%2 max

i=1,...,N

h2i

8∥∥R′i

∥∥∞,Ji

= η3(q,∆) + η4(q,∆) �

Page 33: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Maximum-norm a posteriori error bounds

2a)

|Ii | ≤hi

2‖Ri‖∞,Ji

∫Ji

G(x , ξ) dξ

2b)

|Ii | ≤hi

8

{‖Ri‖∞,Ji

∫Ji

|Gξ(x , ξ)| dξ +∥∥R′i

∥∥∞,Ji

∫Ji

G(x , ξ) dξ}

2a,b) ∣∣∣∣∣N∑

i=1

Ii

∣∣∣∣∣ ≤ 2%2 max

i=1,...,N

hi

2‖Ri‖∞,Ji

min{

1,hi%

}

+1%2 max

i=1,...,N

h2i

8∥∥R′i

∥∥∞,Ji

= η3(q,∆) + η4(q,∆) �

Page 34: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

An adaptive algorithm

Idea: to adaptively design a mesh for which the localcontributions to the a posteriori error estimator

µi (u∆,∆) :=

∥∥∥∥∥ I02q − q

r

∥∥∥∥∥∞,Ji

+|qi−1 − 2qi−1/2 + qi |

4%2

+2%2

[max

{|qi − qi−1/2|, |qi−1/2 − qi−1|

}min

{1,

hi%

}]are the same on each mesh interval, i.e.,

µi−1 (u∆,∆) = µi (u∆,∆) , i = 1, . . . ,N.

This is equivalent to

Qi (u∆,∆) =1N

N∑j=1

Qj (u∆,∆) , Qi (u∆,∆) := µi (u∆,∆)1/2 .

Page 35: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

An adaptive algorithm

Algorithm modification(Kopteva, Stynes SIAM J. Numer. Anal., 2001):

We stop the algorithm when

hiQ̃i (u∆,∆) ≤ γ

N

N∑j=1

hjQ̃j (u∆,∆) ,

for some user chosen constant γ > 1, where

Q̃i (u∆,∆) :=(

h2i + µi (u∆,∆)

)1/2.

Page 36: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

An adaptive algorithm

Algorithm (de Boor, 1973)

1. Fix N, r and a constant γ > 1. The initial mesh ∆[0] isuniform with mesh size 1/N.

2. For k = 0,1, . . . , given the mesh ∆[k ], compute thediscrete solution u[k ]

∆[k ] on this mesh using the S12 -collocation

method. Set h[k ]i = x [k ]

i − x [k ]i−1 for each i . Compute the

piecewise-constant monitor function M [k ] defined by

M [k ](x) := Q̃[k ]i := Q̃i

(u[k ]

∆[k ] ,∆[k ])

for x ∈(xk

i−1, xki).

The total integral of the monitor function is

I[k ] :=

∫ 1

0M [k ](t) dt =

N∑j=1

h[k ]j Q̃[k ]

j .

Page 37: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

An adaptive algorithm

3. Test mesh: If

h[k ]j Q̃[k ]

j ≤ γI[k ]N−1 for all j = 1, . . . ,N

then go to Step 5. Otherwise, continue to Step 4.4. Generate a new mesh by equidistributing the monitor

function M [k ], i.e., choose the new mesh ∆[k+1] such that∫ x [k+1]i

x [k+1]i−1

M [k ](t) dt =I[k ]

N, i = 1, . . . ,N.

Return to Step 2.

5. Set ∆∗ = ∆[k ] and u∗∆∗ = u[k ]

∆[k ] then stop.

Page 38: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Test problem

−ε2u′′(x) + 4u(x) = cos 12x , x ∈ (0,1), u(0) = u(1) = 0

Discrete maximum-norm

‖uε − uε∆‖∞ ≈ χεN := max

i=1,...,Nm=0,...,M

∣∣∣(uε − uε∆) (xi−1 + mM−1hi)∣∣∣

χN := maxk=0,...,20

χ10−k

N

Rates of convergence

p̃N :=lnχN − lnχ2N

ln 2(Bakhvalov and uniform mesh)

pN :=lnχN − lnχ2N

ln 2 + ln ln N − ln ln 2N(two meshes of Shishkin type)

Page 39: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Test problem

−ε2u′′(x) + 4u(x) = cos 12x , x ∈ (0,1), u(0) = u(1) = 0

Discrete maximum-norm

‖uε − uε∆‖∞ ≈ χεN := max

i=1,...,Nm=0,...,M

∣∣∣(uε − uε∆) (xi−1 + mM−1hi)∣∣∣

χN := maxk=0,...,20

χ10−k

N

Rates of convergence

p̃N :=lnχN − lnχ2N

ln 2(Bakhvalov and uniform mesh)

pN :=lnχN − lnχ2N

ln 2 + ln ln N − ln ln 2N(two meshes of Shishkin type)

Page 40: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Test problem

−ε2u′′(x) + 4u(x) = cos 12x , x ∈ (0,1), u(0) = u(1) = 0

Discrete maximum-norm

‖uε − uε∆‖∞ ≈ χεN := max

i=1,...,Nm=0,...,M

∣∣∣(uε − uε∆) (xi−1 + mM−1hi)∣∣∣

χN := maxk=0,...,20

χ10−k

N

Rates of convergence

p̃N :=lnχN − lnχ2N

ln 2(Bakhvalov and uniform mesh)

pN :=lnχN − lnχ2N

ln 2 + ln ln N − ln ln 2N(two meshes of Shishkin type)

Page 41: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Table: Maximum-norm errors of the collocation method onlayer-adapted and uniform mesh

smoothed standard Bakhvalov uniformShishkin mesh Shishkin mesh mesh mesh

N χN pN χN pN χN p̃N χN p̃N

26 3.198e-03 2.49 2.879e-03 2.29 1.024e-04 2.07 1.574e-01 0.0027 8.375e-04 2.10 8.375e-04 2.10 2.439e-05 2.03 1.575e-01 0.0028 2.588e-04 2.08 2.588e-04 2.08 5.987e-06 2.01 1.575e-01 0.0029 7.800e-05 2.05 7.800e-05 2.05 1.485e-06 2.01 1.574e-01 0.00210 2.335e-05 2.03 2.335e-05 2.03 3.698e-07 2.00 1.574e-01 0.00211 6.940e-06 2.02 6.940e-06 2.02 9.229e-08 2.00 1.574e-01 0.00212 2.046e-06 2.01 2.046e-06 2.01 2.305e-08 2.00 1.574e-01 0.00213 5.971e-07 2.00 5.971e-07 2.00 5.761e-09 2.00 1.574e-01 0.00214 1.726e-07 2.00 1.726e-07 2.00 1.440e-09 2.00 1.575e-01 0.00215 4.947e-08 2.00 4.947e-08 2.00 3.599e-10 2.00 1.575e-01 0.00216 1.406e-08 2.00 1.406e-08 2.00 8.998e-11 2.00 1.574e-01 0.00217 3.966e-09 2.00 3.966e-09 2.00 2.249e-11 2.00 1.574e-01 0.00218 1.111e-09 — 1.111e-09 — 5.623e-12 — 1.574e-01 —

Page 42: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Table: A posteriori-error estimates for smoothed Shishkin meshes;ε = 10−12

N χN η ηI η3 η4 χN/η

26 3.198e-03 6.468e-02 1.662e-03 5.302e-02 1.000e-02 4.945e-0227 8.375e-04 2.437e-02 2.190e-04 1.991e-02 4.238e-03 3.437e-0228 2.588e-04 8.510e-03 2.730e-05 6.905e-03 1.578e-03 3.042e-0229 7.800e-05 2.807e-03 3.401e-06 2.265e-03 5.386e-04 2.779e-02210 2.335e-05 8.878e-04 4.246e-07 7.138e-04 1.736e-04 2.630e-02211 6.940e-06 2.724e-04 5.302e-08 2.185e-04 5.381e-05 2.548e-02212 2.046e-06 8.168e-05 6.624e-09 6.545e-05 1.623e-05 2.504e-02213 5.971e-07 2.407e-05 8.278e-10 1.927e-05 4.797e-06 2.481e-02214 1.726e-07 6.996e-06 1.035e-10 5.600e-06 1.397e-06 2.468e-02215 4.947e-08 2.011e-06 1.293e-11 1.609e-06 4.017e-07 2.461e-02216 1.406e-08 5.723e-07 1.616e-12 4.579e-07 1.144e-07 2.457e-02217 3.966e-09 1.616e-07 2.021e-13 1.293e-07 3.231e-08 2.455e-02218 1.111e-09 4.529e-08 2.526e-14 3.624e-08 9.058e-09 2.454e-02

Page 43: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Table: A posteriori-error estimates for smoothed Shishkin meshes,robustness of the estimator; N = 214

ε χN η ηI η3 η4 χN/η

1 1.715e-09 8.335e-08 7.159e-13 6.846e-08 1.489e-08 2.058e-0210−2 1.720e-07 6.971e-06 3.587e-12 5.580e-06 1.392e-06 2.468e-0210−3 1.726e-07 6.996e-06 8.304e-11 5.599e-06 1.397e-06 2.468e-0210−4 1.726e-07 6.996e-06 1.013e-10 5.600e-06 1.397e-06 2.468e-0210−6 1.726e-07 6.996e-06 1.034e-10 5.600e-06 1.397e-06 2.468e-0210−8 1.726e-07 6.996e-06 1.035e-10 5.600e-06 1.397e-06 2.468e-0210−12 1.726e-07 6.996e-06 1.035e-10 5.600e-06 1.397e-06 2.468e-0210−16 1.726e-07 6.996e-06 1.035e-10 5.600e-06 1.397e-06 2.468e-0210−20 1.726e-07 6.996e-06 1.035e-10 5.600e-06 1.397e-06 2.468e-02

Page 44: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

Table: The adaptive algorithm, ε = 10−12

N χN p̃N η ηI η3 η4 χN/η #iter26 2.24e-04 0.63 2.07e-03 1.07e-04 1.67e-03 2.97e-04 1.08e-01 827 1.45e-04 2.09 1.12e-03 1.43e-05 9.88e-04 1.13e-04 1.30e-01 1428 3.39e-05 2.17 1.26e-04 1.65e-06 1.09e-04 1.46e-05 2.70e-01 629 7.53e-06 2.37 3.96e-05 2.15e-07 3.56e-05 3.77e-06 1.90e-01 5210 1.46e-06 0.80 8.96e-06 2.60e-08 8.07e-06 8.67e-07 1.63e-01 5211 8.37e-07 3.86 2.86e-06 3.50e-09 2.60e-06 2.58e-07 2.92e-01 4212 5.75e-08 0.27 4.94e-07 4.11e-10 4.41e-07 5.27e-08 1.16e-01 4213 4.76e-08 3.65 1.63e-07 5.01e-11 1.48e-07 1.50e-08 2.92e-01 4214 3.79e-09 -0.11 6.38e-08 1.23e-11 5.10e-08 1.28e-08 5.93e-02 3215 4.10e-09 3.35 1.41e-08 8.96e-13 1.27e-08 1.35e-09 2.91e-01 3216 4.03e-10 1.17 2.33e-09 1.03e-13 2.11e-09 2.21e-10 1.73e-01 3217 1.80e-10 1.72 6.17e-10 1.26e-14 5.59e-10 5.83e-11 2.91e-01 3218 5.44e-11 — 1.83e-10 1.56e-15 1.69e-10 1.48e-11 2.97e-01 3a.r. 1.83 1.95 3.00 1.94 2.02

Page 45: Quadratic C1-spline collocation for reaction-diffusion ... filegeneral theory for spline-collocation methods applied to classical (not SP) BVPs transition to SPBVPs bounds with “constants”

Introduction Layer-adapted mesh The interpolation error The collocation method Numerical experiments Conclusion

robust convergence of almost second order

further investigation:higher-order splines (a posteriori estimator)convection-diffusion problems (stability issue)