Equivalence Theorems and Their Applicationstanbui/Equivalence...(Uniform boundedness)...
Transcript of Equivalence Theorems and Their Applicationstanbui/Equivalence...(Uniform boundedness)...
Equivalence Theorems and Their Applications
Tan Bui-Thanh,
Center for Computational Geosciences and OptimizationInstitute for Computational Engineering and Sciences (ICES)
The University of Texas at Austin, USA
September 13, 2010
Time Domain Electromagnetic WavesSpherical Cavity
electromagnetic scattering
Outline
Consistency, Stability and ConvergenceAn Equivalence Theorem for y = Tx: T linearLax Equivalence Theorem for u′(t) = Au(t)Nonlinear maps y = TxEquivalence Theorem for Ay = x
A Discontinuous Spectral Element Method for HyperbolicEquations?? (Next presentation)
Stability of a Discontinuous Spectral Element Method for WavePropagation Problems (Next presentation)
Consistency, Stability and ConvergenceAn Equivalence Theorem for y = Tx: T linearLax Equivalence Theorem for u′(t) = Au(t)Nonlinear maps y = TxEquivalence Theorem for Ay = x
A Discontinuous Spectral Element Method for HyperbolicEquations?? (Next presentation)
Stability of a Discontinuous Spectral Element Method for WavePropagation Problems (Next presentation)
Motivation
QuestionWhat is the first thing you need to do when you derive/invent anew numerical method?
I consistency
I stability
I convergence
Motivation
QuestionWhat is the first thing you need to do when you derive/invent anew numerical method?
I consistency
I stability
I convergence
Consistency, Stability and Convergence
What is consistency?
I Consistency is a measure of how close a discretization is tothe “continuous” problem = how good you approximateoperators and functions
What is stability?
I Stability means that the propagated error is controlled by theerror in the data = continuity of solution w.r.t the data,uniform boundedness of the discrete operator
What is convergence?
I Convergence means that the discrete solution converges tothe exact solution = error between the exact and discretesolutions converges to zero
Consistency, Stability and Convergence
What is consistency?
I Consistency is a measure of how close a discretization is tothe “continuous” problem = how good you approximateoperators and functions
What is stability?
I Stability means that the propagated error is controlled by theerror in the data = continuity of solution w.r.t the data,uniform boundedness of the discrete operator
What is convergence?
I Convergence means that the discrete solution converges tothe exact solution = error between the exact and discretesolutions converges to zero
Consistency, Stability and Convergence
What is consistency?
I Consistency is a measure of how close a discretization is tothe “continuous” problem = how good you approximateoperators and functions
What is stability?
I Stability means that the propagated error is controlled by theerror in the data = continuity of solution w.r.t the data,uniform boundedness of the discrete operator
What is convergence?
I Convergence means that the discrete solution converges tothe exact solution = error between the exact and discretesolutions converges to zero
Consistency, Stability and Convergence
What is consistency?
I Consistency is a measure of how close a discretization is tothe “continuous” problem = how good you approximateoperators and functions
What is stability?
I Stability means that the propagated error is controlled by theerror in the data = continuity of solution w.r.t the data,uniform boundedness of the discrete operator
What is convergence?
I Convergence means that the discrete solution converges tothe exact solution = error between the exact and discretesolutions converges to zero
Consistency, Stability and Convergence
What is consistency?
I Consistency is a measure of how close a discretization is tothe “continuous” problem = how good you approximateoperators and functions
What is stability?
I Stability means that the propagated error is controlled by theerror in the data = continuity of solution w.r.t the data,uniform boundedness of the discrete operator
What is convergence?
I Convergence means that the discrete solution converges tothe exact solution = error between the exact and discretesolutions converges to zero
Consistency, Stability and Convergence
What is consistency?
I Consistency is a measure of how close a discretization is tothe “continuous” problem = how good you approximateoperators and functions
What is stability?
I Stability means that the propagated error is controlled by theerror in the data = continuity of solution w.r.t the data,uniform boundedness of the discrete operator
What is convergence?
I Convergence means that the discrete solution converges tothe exact solution = error between the exact and discretesolutions converges to zero
The importance of equivalence theorems
Which of the three (consistency, stability, convergence) is themost difficult? Why?
I Convergence: needs knowledge about the exact solution
Alternate route for convergence: Equivalence theorems
consistency + stability → convergence
Peter Lax 1953“Well-posedness of the original differential equationproblem and consistency imply the equivalence betweenstability and convergence of difference methods”
stability ⇔ convergence
The importance of equivalence theorems
Which of the three (consistency, stability, convergence) is themost difficult? Why?
I Convergence: needs knowledge about the exact solution
Alternate route for convergence: Equivalence theorems
consistency + stability → convergence
Peter Lax 1953“Well-posedness of the original differential equationproblem and consistency imply the equivalence betweenstability and convergence of difference methods”
stability ⇔ convergence
The importance of equivalence theorems
Which of the three (consistency, stability, convergence) is themost difficult? Why?
I Convergence: needs knowledge about the exact solution
Alternate route for convergence: Equivalence theorems
consistency + stability → convergence
Peter Lax 1953“Well-posedness of the original differential equationproblem and consistency imply the equivalence betweenstability and convergence of difference methods”
stability ⇔ convergence
The importance of equivalence theorems
Which of the three (consistency, stability, convergence) is themost difficult? Why?
I Convergence: needs knowledge about the exact solution
Alternate route for convergence: Equivalence theorems
consistency + stability → convergence
Peter Lax 1953“Well-posedness of the original differential equationproblem and consistency imply the equivalence betweenstability and convergence of difference methods”
stability ⇔ convergence
More on Stability
The “easiest” among the three?
I purely the property of the discrete problem: Knowledge aboutthe exact solution/operators is not needed
I An analogy: uniqueness implies existence (matrix theory,Fredolm theory). Think about the proof of Banach fixed pointtheorem. (next talk about inverse problem theory)
Extremely important for computer implementation?
I round-off errors
More on Stability
The “easiest” among the three?
I purely the property of the discrete problem: Knowledge aboutthe exact solution/operators is not needed
I An analogy: uniqueness implies existence (matrix theory,Fredolm theory). Think about the proof of Banach fixed pointtheorem. (next talk about inverse problem theory)
Extremely important for computer implementation?
I round-off errors
More on Stability
The “easiest” among the three?
I purely the property of the discrete problem: Knowledge aboutthe exact solution/operators is not needed
I An analogy: uniqueness implies existence (matrix theory,Fredolm theory). Think about the proof of Banach fixed pointtheorem. (next talk about inverse problem theory)
Extremely important for computer implementation?
I round-off errors
More on Stability
The “easiest” among the three?
I purely the property of the discrete problem: Knowledge aboutthe exact solution/operators is not needed
I An analogy: uniqueness implies existence (matrix theory,Fredolm theory). Think about the proof of Banach fixed pointtheorem. (next talk about inverse problem theory)
Extremely important for computer implementation?
I round-off errors
More on Stability
The “easiest” among the three?
I purely the property of the discrete problem: Knowledge aboutthe exact solution/operators is not needed
I An analogy: uniqueness implies existence (matrix theory,Fredolm theory). Think about the proof of Banach fixed pointtheorem. (next talk about inverse problem theory)
Extremely important for computer implementation?
I round-off errors
General settings
DefinitionsLet V ,W be Banach spaces, and T,Th : V →W
i) Wellposedness: continuity of T,Th
ii) Consistency: Th is said to be consistent with T iflimh→0 ‖(T h − T ) v0‖ = 0, ∀v0 ∈D ⊂ V , D dense in V ,
iii) Stability: Th is called stable if suph ‖Th‖ <∞. (Uniformboundedness)
iv) Convergence: Th is said to converge to T, iflimh→0 ‖(T h − T ) v‖ = 0, ∀v ∈ V
Replace h by n if n is more natural
Equivalence Theorem for Linear Operators
TheoremA consistent family of Th is convergent if and only if it is stable.
Proof.
⇒) Since limh→0 ‖(T h − T ) v‖ = 0, ∀v ∈ V , the family Th ispointwise uniformly bounded continuous linear operators. Theuniform boundedness principle yields suph ‖Th‖ <∞, which isexactly stability.
⇐) By triangle inequality (three-ε argument), v0 ∈D,
‖Tv −Thv‖ ≤ ‖T‖ ‖v − v0‖+ ‖(T−Th) v0‖+ ‖Th‖ ‖v0 − v‖ .
The proof is complete by the following two facts. First, theconsistency implies ∃h0 : ∀h ≤ h0 such that‖(T−Th) v0‖ ≤ ε/3. Second, the density of D allows us topick v0 such that ‖v − v0‖ ≤ ε
3max{‖T‖,suph‖Th‖} .
Equivalence Theorem for Linear Operators
TheoremA consistent family of Th is convergent if and only if it is stable.
Proof.
⇒) Since limh→0 ‖(T h − T ) v‖ = 0, ∀v ∈ V , the family Th ispointwise uniformly bounded continuous linear operators. Theuniform boundedness principle yields suph ‖Th‖ <∞, which isexactly stability.
⇐) By triangle inequality (three-ε argument), v0 ∈D,
‖Tv −Thv‖ ≤ ‖T‖ ‖v − v0‖+ ‖(T−Th) v0‖+ ‖Th‖ ‖v0 − v‖ .
The proof is complete by the following two facts. First, theconsistency implies ∃h0 : ∀h ≤ h0 such that‖(T−Th) v0‖ ≤ ε/3. Second, the density of D allows us topick v0 such that ‖v − v0‖ ≤ ε
3max{‖T‖,suph‖Th‖} .
Equivalence Theorem for Linear Operators
TheoremA consistent family of Th is convergent if and only if it is stable.
Proof.
⇒) Since limh→0 ‖(T h − T ) v‖ = 0, ∀v ∈ V , the family Th ispointwise uniformly bounded continuous linear operators. Theuniform boundedness principle yields suph ‖Th‖ <∞, which isexactly stability.
⇐) By triangle inequality (three-ε argument), v0 ∈D,
‖Tv −Thv‖ ≤ ‖T‖ ‖v − v0‖+ ‖(T−Th) v0‖+ ‖Th‖ ‖v0 − v‖ .
The proof is complete by the following two facts. First, theconsistency implies ∃h0 : ∀h ≤ h0 such that‖(T−Th) v0‖ ≤ ε/3. Second, the density of D allows us topick v0 such that ‖v − v0‖ ≤ ε
3max{‖T‖,suph‖Th‖} .
Numerical Integrations
A suitable settingI V = (C [a, b] , ‖·‖∞)
I Tf =∫ ba f dx, f ∈ V , then T is linear and bounded,
|Tf | ≤ (b− a) ‖f‖∞I Tnf =
∑ni=1wifi, f ∈ V , then Tn is linear and bounded,
|Tnf | ≤ (∑n
i=1wi) ‖f‖∞I The family Tn is consistent if ∀f ∈ P, P space of
polynomials (P ⊂ V , dense ?), limn→∞ |Tnf −Tf | = 0.
I The family Tn is stable if supn ‖Tn‖ <∞.
I The family Tn is convergent if ∀f ∈ Vlimn→∞ |Tnf −Tf | = 0
Examples
Stable : Trapezoidal, Simpson, Gauss quadrature, and etc
Unstable : Newton-Cotes
Numerical Derivatives
A suitable settingI V =
(Ck [a, b] , ‖·‖Ck
), W = (C [a, b] , ‖·‖∞)
I D(k)f, f ∈ V is linear and bounded,∥∥D(k)f
∥∥∞ ≤ ‖f‖Ck
I Denote D(k)h f, f ∈ V the numerical derivative
I The family D(k)h is consistent if
limh→0
∥∥∥(D(k)h −D
(k))f∥∥∥∞
= 0 ∀f ∈D ⊂ V , D is dense
I The family D(k)h is stable if suph
∥∥∥D(k)h
∥∥∥∞<∞.
I The family D(k)h is convergent if ∀f ∈ V
limh→0
∥∥∥D(k)h f −D(k)f
∥∥∥∞
= 0
Examples: Stability of forward differentiation∥∥∥D(1)h
∥∥∥ = sup‖f‖C1=1 supx∣∣∣f(x+h)−f(x)
h
∣∣∣ =sup‖f‖C1=1 supx |f ′(x+ θh)| ≤ 1
Numerical Derivatives
Convergence is independent of Stability
limh→0
∥∥∥D(1)h f −D(1)f
∥∥∥∞
= limh→0
supx
∣∣∣∣f(x+ h)− f(x)h
− f ′(x)∣∣∣∣ =
limh→0
supx
∣∣f ′(x+ θh)− f ′(x)∣∣ = 0
Linear time dependent problems
One step method
Consider the problem u′(t) = Au(t), u(0) = u0, and an abstractone step method
v(h) = Bhu0,
v(nh) = Bnhu0.
Well-posedness of the continuous problem
Let S : V → V , u(t) = S(t)u0, we require u(t) is continuous w.r.tt and supt ‖S(t)‖ <∞.
Well-posedness of the discrete problem
For each 0 ≤ h ≤ h0, we require suph ‖Bh‖ <∞
Linear time dependent problems
Consistency
∀v0 ∈D ⊂ V , D is dense,
limh→0‖Bhu(t)− S(t+ h)v0‖ = 0
Stability
‖Bnh‖ <∞, ∀h, n : nh ≤ T
Convergence
limk→∞
∥∥∥Bnkhkv0 − S(t)v0
∥∥∥ = 0, where limk→∞ nkhk = t
Stability ⇔ Convergence
Stability ⇒ Convergence
Using triangle inequality and three-ε trick we have
‖v(nh)− u(t)‖ = ‖Bnhu0 − S(t)u0‖ ≤
‖Bnh‖ ‖v − u0‖︸ ︷︷ ︸
stability + density
+ ‖Bnhv − S(t)v‖︸ ︷︷ ︸consistency
+ ‖S(t)‖ ‖v − u0‖︸ ︷︷ ︸wellposedness + density
→ 0,
Stability ⇔ Convergence
Convergence ⇒ Stability
If n is finite then by the discrete wellposedness we have
suph‖Bn
h‖ ≤ suph‖Bh‖n <∞.
Now k →∞ : limk nkhk = t, by convergence we have for eachu0 ∈ V , the family Bnk
hkis uniformly bounded. Then by the
uniform boundedness principle, we have
supk
∥∥∥Bnkhk
∥∥∥ <∞.In both cases, the stability condition is proved.
A nonlinear setting for y = Tx
T is nonlinearT : V →W , and V ,W are Banach
I Wellposedness: T is continuous
I Consistency: convergence on a dense subspace D
∀v ∈D ⊂ V : limn→∞
‖Tnv −Tv‖ = 0
I Stability ∀ε, ‖v − v0‖ ≤ δ, such that ‖Tnv −Tnv0‖ ≤ ε(replace uniform boundedness by equi-continuity)
I Convergence
∀v ∈ V ,∃n0 : n ≥ n0 limn→∞
‖Tnv −Tv‖ = 0
Equivalence Theorem for nonlinear Operators
TheoremA consistent family of Th is convergent if and only if it is stable.
Proof.
⇒) Again the three-ε trick, ‖v − v0‖ < δ
‖Tnv −Tnv0‖ ≤‖Tnv −Tv‖︸ ︷︷ ︸
convergence
+ ‖Tv −Tv0‖︸ ︷︷ ︸Wellposedness
+ ‖Tv0 −Tnv0‖︸ ︷︷ ︸convergence
≤ ε
⇐) By the three-ε trick, ∀v ∈ V , by density ∃v0 ∈D,‖v − v0‖ < δ
‖Tv −Tnv‖ ≤‖Tv −Tv0‖︸ ︷︷ ︸wellposedness
+ ‖Tv0 −Tnv0‖︸ ︷︷ ︸Consistency
+ ‖Tnv0 −Tnv‖︸ ︷︷ ︸Stability
≤ ε
Equivalence Theorem for nonlinear Operators
TheoremA consistent family of Th is convergent if and only if it is stable.
Proof.
⇒) Again the three-ε trick, ‖v − v0‖ < δ
‖Tnv −Tnv0‖ ≤‖Tnv −Tv‖︸ ︷︷ ︸
convergence
+ ‖Tv −Tv0‖︸ ︷︷ ︸Wellposedness
+ ‖Tv0 −Tnv0‖︸ ︷︷ ︸convergence
≤ ε
⇐) By the three-ε trick, ∀v ∈ V , by density ∃v0 ∈D,‖v − v0‖ < δ
‖Tv −Tnv‖ ≤‖Tv −Tv0‖︸ ︷︷ ︸wellposedness
+ ‖Tv0 −Tnv0‖︸ ︷︷ ︸Consistency
+ ‖Tnv0 −Tnv‖︸ ︷︷ ︸Stability
≤ ε
Equivalence Theorem for nonlinear Operators
TheoremA consistent family of Th is convergent if and only if it is stable.
Proof.
⇒) Again the three-ε trick, ‖v − v0‖ < δ
‖Tnv −Tnv0‖ ≤‖Tnv −Tv‖︸ ︷︷ ︸
convergence
+ ‖Tv −Tv0‖︸ ︷︷ ︸Wellposedness
+ ‖Tv0 −Tnv0‖︸ ︷︷ ︸convergence
≤ ε
⇐) By the three-ε trick, ∀v ∈ V , by density ∃v0 ∈D,‖v − v0‖ < δ
‖Tv −Tnv‖ ≤‖Tv −Tv0‖︸ ︷︷ ︸wellposedness
+ ‖Tv0 −Tnv0‖︸ ︷︷ ︸Consistency
+ ‖Tnv0 −Tnv‖︸ ︷︷ ︸Stability
≤ ε
Linear equation Ay = x
Wellposedness
Let A : W → V , and V ,W are Banach. The problem iswellposed if A is continuous and bijective. We know that theinverse exists, i.e. T = A−1, and T is continuous (Why??).
The problem can be converted to the form y = Tx which we havealready discussed.
Consistency, Stability and ConvergenceAn Equivalence Theorem for y = Tx: T linearLax Equivalence Theorem for u′(t) = Au(t)Nonlinear maps y = TxEquivalence Theorem for Ay = x
A Discontinuous Spectral Element Method for HyperbolicEquations?? (Next presentation)
Stability of a Discontinuous Spectral Element Method for WavePropagation Problems (Next presentation)
Consistency, Stability and ConvergenceAn Equivalence Theorem for y = Tx: T linearLax Equivalence Theorem for u′(t) = Au(t)Nonlinear maps y = TxEquivalence Theorem for Ay = x
A Discontinuous Spectral Element Method for HyperbolicEquations?? (Next presentation)
Stability of a Discontinuous Spectral Element Method for WavePropagation Problems (Next presentation)