Date post: | 07-Dec-2014 |
Category: |
Documents |
Upload: | franco-nelson |
View: | 24 times |
Download: | 1 times |
Numerical AnalysisM1 SMA International
Ecole Centrale de Nantes
Anthony NOUY
Oce : F231
Origin of problems in numerical analysis References
Part I
Introduction
1 Origin of problems in numerical analysis
2 References
Origin of problems in numerical analysis References
Part I
Introduction
1 Origin of problems in numerical analysis
2 References
Origin of problems in numerical analysis References
Origin of problems in numerical analysis I
How to interpret the reality with a computer language: from a continuousworld to a discrete world.
Numerical solution of a dierential equation
Find u : x ∈ Ω 7→ u(x) such that
A(u) = b
Example 1 (1D diusion equation, beam in traction, ...)
− d
dx(α
du
dx) = b(x) for x ∈ Ω = (0, 1), u(0) = u(1) = 0
Origin of problems in numerical analysis References
Origin of problems in numerical analysis II
Approximation (from a continuous to a discrete representation)
Represent a function u on a (nite-dimensional) approximation space :
u(x) =n∑i=1
uihi (x)
The solution is then represented by u = (u1, . . . , un) ∈ Rn.
For the denition of the expansion, dierent alternatives such as methodsbased on a weak formulation of the problem.
Example 2 (Galerkin approximation)
Find u ∈ V = v : (0, 1)→ R; v(0) = v(1) = 0 such that∫Ω
dv
dxαdu
dxdx =
∫Ω
v b dx ∀v ∈ V
and replace function space V by approximation spaceVn = v(x) =
∑n
i=1vihi (x) ⊂ V
Origin of problems in numerical analysis References
Origin of problems in numerical analysis III
If A is a linear operator, the initial continuous equation is then transformed into
Linear systems of equations
Find u ∈ Rn such thatAu = b
where A ∈ Rn×n is a matrix and b ∈ Rn a vector.
In order to construct the system of equation (matrix A and right-hand-side b):
Numerical Integration ∫Ω
f (x) dx ≈K∑k=1
ωk f (xk)
If A is a nonlinear operator:
Origin of problems in numerical analysis References
Origin of problems in numerical analysis IV
Nonlinear system of equations
Find u ∈ Rn such thatA(u) = b
where A : u ∈ Rn 7→ A(u) ∈ Rn.
Remedy: iterative solution techniques which transform the solution of anonlinear equation into solution of linear equations.
Example 3
− d
dx(α(x , u)
du
dx) = b(x , u) for x ∈ Ω = (0, 1), u(0) = u(1) = 0
Eigenproblems
Find (u, λ) ∈ Cn × C such that
Au = λu or Au = λBu
where A,B ∈ Cn×n are matrices.
Origin of problems in numerical analysis References
Origin of problems in numerical analysis V
Example 4 (Eigenmodes of a beam)
Wave equation: solution u(x , t) such that
− ∂
∂x(α∂u
∂x) + ρ
∂2u
∂t2= 0 for x ∈ Ω = (0, 1), u(0, t) = u(1, t) = 0
for which we search solutions of the form u(x , t) = w(x) cos(ωt):
w ∈ Vn,
∫Ω
∂v
∂xα∂w
∂xdx = ω2
∫Ω
ρv w dx ∀v ∈ Vn
Ordinary dierential equations in time
d
dtu(t) + A(u(t); t) = b(t)
Origin of problems in numerical analysis References
Part I
Introduction
1 Origin of problems in numerical analysis
2 References
Origin of problems in numerical analysis References
References for the course
G. Allaire and S. M. Kaber.
Numerical linear algebra.Springer, 2007. → materials for chapters 1 (Linear Algebra), 2 (Linearsystems), 3(Eigenvalues)
K. Atkinson and W. Han.
Theoretical Numerical Analysis: A Functional Analysis Framework.Springer, 2009.→ materials for chapters 4 (Nonlinear equations), 5(Approximation/Interpolation)→ a quite abstract introduction to numerical analysis (very instructive),with an introduction to functional analysis
E. Suli and D. Mayers.
An Introduction to Numerical Analysis.Cambridge University Press, 2003.→ a clear and simple presentation of all the ingredients of the course
G. Allaire.
Numerical Analysis and Optimization.Cambridge University Press, 2007.→ additional material for numerical solution of PDE and optimizationproblems→ a natural continuation of the course
Matrices Reduction of matrices Vector and matrix norms
Part II
Linear algebra
3 Matrices
4 Reduction of matrices
5 Vector and matrix norms
Matrices Reduction of matrices Vector and matrix norms
Part II
Linear algebra
3 Matrices
4 Reduction of matrices
5 Vector and matrix norms
Matrices Reduction of matrices Vector and matrix norms
Vector space
Let V be a vector space with nite dimension n, on the eld K (R or C).Let E = e1, . . . , en be a basis of V . A vector v ∈ V admits a uniquedecomposition
v =n∑i=1
viei
where the (vi )ni=1 are the components of v on the basis E . When a basis is
chosen and when there is no ambiguity, we can identify V to Kn (Rn or Cn)and let v = (vi )
ni=1, represented by the column vector
v =
v1...vn
We denote respectively by vT and vH the transpose and conjugate transpose ofv , which are the following row vectors
vT =(v1 . . . vn
)vH =
(v1 . . . vn
)where a denotes the complex conjugate of a.
Matrices Reduction of matrices Vector and matrix norms
Canonical inner product
We denote by (·, ·) : V × V → K the canonical inner product dened for allu, v ∈ V by
(u, v) = uT v = vTu =n∑i=1
uivi if K = R
(u, v) = uHv = vHu =n∑i=1
uivi if K = C
It is called euclidian inner product if K = R and hermitian inner product ifK = C.
Matrices Reduction of matrices Vector and matrix norms
Orthogonality
Orthogonality on a vector space V must be thought with respect to an innerproduct (·, ·). If not mentioned, we classically consider the canonical innerproduct.Two vectors u, v ∈ V are said orthogonal with respect to inner product (·, ·) ifand only if (u, v) = 0.A vector v is said orthogonal to a linear subspace U ⊂ V , which is denotedv ⊥ U, if and only if (v , u) = 0 for all u ∈ U. Two linear subspaces U ⊂ V andU ′ ⊂ V are said orthogonal, and it is denoted U ⊥ U ′, if
(u, u′) = 0 ∀u ∈ U, ∀u′ ∈ U ′
For a given subspace U ⊂ V , we denote by U⊥ its orthogonal complement,which is the largest subspace orthogonal to U. The orthogonal complement ofa vector v ∈ V is denoted by v⊥.
Matrices Reduction of matrices Vector and matrix norms
Matrices
Let V and W be two vector spaces with dimension n and m respectively, withbases E = (ei )
ni=1 and F = (fi )
mi=1. A linear map A : V →W , relatively to
those bases, is represented by a matrix A with m rows and n columns
A =
a11 a12 . . . a1na21 a22 . . . a2n...
......
am1 am2 . . . amn
where the coecients aij are such that
Aej =m∑i=1
aij fi , 1 6 j 6 n
We denote (A)ij = aij . The j-th column of A represents the vector Aej in thebasis F .
Denition 5
The set of matrices with m rows and n columns with entries in the eld K is avector space denotedMm,n(K) or Km×n.
Matrices Reduction of matrices Vector and matrix norms
Transpose
We denote AH the adjoint (or conjugate transpose) matrix of a complex matrixA = (aij ) ∈ Cm×n, dened by
(AH)ij = aji
We denote AT the transpose of a real matrix A = (aij ) ∈ Rn×m, dened by
(AT )ij = aji
We have the following characterization of AH and AT :
(Au, v) = (u,AHv) ∀u ∈ Cn, v ∈ Cm
(Au, v) = (u,AT v) ∀u ∈ Rn, v ∈ Rm
Matrices Reduction of matrices Vector and matrix norms
Product
To the composition of two linear maps corresponds the multiplication of theassociated matrices. If A = (aik) ∈ Km×q and B = (bkj ) ∈ Kq×n, the productAB ∈ Km×n is dened by
(AB)ij =
q∑k=1
aikbkj
We have(AB)T = BTAT , (AB)H = BHAH
The set of square matricesMn,n(K) is simply denotedMn(K) = Kn×n. In thefollowing, unless it is mentioned, we only consider square matrices.
Matrices Reduction of matrices Vector and matrix norms
Inverse
We denote by In the identity matrix on Kn×n, associated with the identity mapfrom V to V . If there is no ambiguity, we simply denote In = I and
(I )ij = δij
where δij is the Knonecker delta.A matrix is invertible if there exists a matrix denoted A−1 (unique if it exists)and called the inverse matrix of A, such that AA−1 = A−1A = I . A matrixwhich is not invertible is said singular. If A and B are invertible, we have
(AB)−1 = B−1A−1, (AT )−1 = (A−1)T ≡ A−T , (AH)−1 = (A−1)H ≡ A−H
Matrices Reduction of matrices Vector and matrix norms
Particular matrices
Denition 6
A matrix A ∈ Cn×n is said
Hermitian if A = AH
Normal if AAH = AHA
Unitary if AAH = AHA = I
Denition 7
A matrix A ∈ Rn×n is said
Symmetric if A = AT
Orthogonal if AAT = ATA = I
Matrices Reduction of matrices Vector and matrix norms
Particular matrices
A matrix A ∈ Kn×n is said diagonal if aij = 0 for i 6= j and we denote
A = diag(aii ) = diag(a11, . . . , ann) =
a11 0 . . . 0
0. . .
. . ....
.... . .
. . . 00 . . . 0 ann
A matrix A is said upper triangular if aij = 0 for i > j :
A =
a11 a12 . . . a1n0 a22 . . . a2n...
. . .. . .
...0 . . . 0 ann
A matrix A is said lower triangular if aij = 0 for j > i :
A =
a11 0 . . . 0
a21 a22. . .
......
.... . . 0
an1 an2 . . . ann
Matrices Reduction of matrices Vector and matrix norms
Properties of triangular matrices
Let Ln ⊂ Kn×n be the set of lower triangular matrices, and Un ⊂ Kn×n be theset of upper triangular matrices.
Theorem 8
If A,B ∈ Ln, then AB ∈ Ln
If A,B ∈ Un, then AB ∈ Un
A ∈ Ln (or Un) is invertible if and only if all its diagonal terms are nonzero.
If A ∈ Ln, A−1 ∈ Ln (if it exists)
If A ∈ Un, A−1 ∈ Un (if it exists)
Matrices Reduction of matrices Vector and matrix norms
Trace
Denition 9
The trace of a matrix A ∈ Kn×n is dened as
tr(A) =n∑i=1
aii
Property 10
tr(A + B) = tr(A) + tr(B), tr(AB) = tr(BA)
Matrices Reduction of matrices Vector and matrix norms
Determinant
Let Sn denote the set of permutations of 1, . . . , n. For σ ∈ Sn, we denote bysign(σ) the signature of the permutation, with sign(σ) = +1 (resp. −1) if σ isan even (resp. odd) permutation of 1, . . . , n.
Denition 11
The determinant of a matrix A ∈ Kn×n is dened as
det(A) =∑σ∈Sn
sign(σ)aσ(1)1 . . . aσ(n)n
Property 12
det(AB) = det(BA) = det(A)det(B)
Matrices Reduction of matrices Vector and matrix norms
Image, Kernel I
Denition 13
The image of A ∈ Km×n is a linear subspace of Km dened by
Im(A) = Av ; v ∈ Kn
The rank of a matrix A, denoted rank(A), is the dimension of Im(A):
rank(A) = dim(Im(A)) 6 min(m, n)
Denition 14
The kernel of A ∈ Km×n is a linear subspace of Kn dened by
Ker(A) = v ∈ Kn;Av = 0
The dimension of Ker(A) is called the nullity of A.
Property 15
dim(Im(A)) + dim(Ker(A)) = n
Matrices Reduction of matrices Vector and matrix norms
Image, Kernel II
Property 16
For A ∈ Rm×n,
Ker(AT ) + Im(A) = Rm, Ker(AT ) = Im(A)⊥
Ker(A) + Im(AT ) = Rn, Ker(A) = Im(AT )⊥
Proof.
Let us prove that Ker(AT ) = Im(A)⊥, which implies Ker(AT ) + Im(A) = Rm.First, u ∈ Ker(AT )⇒ ATu = 0 ⇒ vTATu = 0 ∀v ⇒ uT y = 0 ∀y ∈ Im(A) ⇒Ker(AT ) ⊂ Im(A)⊥.Secondly, u ∈ Im(A)⊥ ⇒ uTAv = 0 ∀v ⇒ vT (ATu) = 0 ∀v ⇒ ATu = 0 ⇒Im(A)⊥ ⊂ Ker(AT ).
Exercice.
Finish the proof.
Matrices Reduction of matrices Vector and matrix norms
Eigenvalues and eigenvectors I
Denition 17
Eigenvalues λi = λi (A), 1 6 i 6 n, of a matrix A ∈ Kn×n are the n roots of itscharacteristic polynomial
pA : λ ∈ C 7→ pA(λ) = det(A− λI )
The eigenvalues may be real or complex. An eigenvalue is said of multiplicity kif it is a root of pA with multiplicity k. The spectrum of matrix A is thefollowing subset of the complex plane
sp(A) = λi (A)ni=1
We have
tr(A) =n∑i=1
λi (A), det(A) =n∏i=1
λi (A)
Matrices Reduction of matrices Vector and matrix norms
Eigenvalues and eigenvectors II
Denition 18
The spectral radius ρ(A) of a matrix A is dened by
ρ(A) = max16i6n
|λi (A)|
Property 19
λ ∈ sp(A) if and only if the following equation has (at least) a nontrivialsolution v ∈ Cn\0:
Av = λv
Denition 20
For λ ∈ sp(A), a vector v satisfying Av = λv is called an eigenvector of Aassociated with λ. The linear subspace v ∈ Kn;Av = λv (with dimension atleast one) is called the eigenspace associated with λ.
Matrices Reduction of matrices Vector and matrix norms
Part II
Linear algebra
3 Matrices
4 Reduction of matrices
5 Vector and matrix norms
Matrices Reduction of matrices Vector and matrix norms
Reduction of matrices
Let V be a vector space with dimension n and A : V → V a linear map on V .Let A be the matrix associated with A, relatively to the basis E = (ei )
ni=1 of V .
Relatively to another basis F = (fi )ni=1 of V , the application A is associated
with another matrix B such that
B = P−1AP
where P is an invertible matrix whose j-th column is composed by thecomponents of fj on the basis E .
Denition 21
Matrices A and B are said similar when they represent the same linear map intwo dierent basis, i.e. when there exists an invertible matrix P such thatB = P−1AP.
Matrices Reduction of matrices Vector and matrix norms
Theorem 22 (Triangularization)
For A ∈ Cn×n, there exists a unitary matrix U such that U−1AU is a triangularmatrix, called the Schur form of A (if upper triangular).
Remark.
The previous theorem says that there exists a nested sequence of A-invariantsubspaces 0 = V0 ⊂ V1 ⊂ . . . ⊂ Vn = Cn and there exists an orthonormalbasis of Cn such that Vi is the span of the rst i basis vectors.
Theorem 23 (Diagonalization)
For a normal matrix A ∈ Cn×n, i.e. such that AHA = AAH , there exists aunitary matrix U such that U−1AU is diagonal.
For a symmetric matrix A ∈ Rn×n, there exists an orthogonal matrix Osuch that O−1AO is diagonal.
Matrices Reduction of matrices Vector and matrix norms
Singular values and vectors
Denition 24
The singular values of A ∈ Km×n are the eigenvalues of√AHA ∈ Kn×n.
Singular values of A are real non-negative numbers.
Denition 25
σ ∈ R+ is a singular value of A if and only if there exists normalized vectorsu ∈ Km and v ∈ Kn such that we have simultaneously
Av = σu and AHu = σv
u and v are respectively called the left and right singular vectors of Aassociated with singular value σ.
Matrices Reduction of matrices Vector and matrix norms
Singular value decomposition (SVD) I
Theorem 26
For A ∈ Km×n, there exist two orthogonal (if K = R) or unitary (if K = C)matrices U ∈ Km×m and V ∈ Kn×n such that
A = USVH
where S = diag(σi ) ∈ Rm×n is a diagonal matrix, with σi the singular values ofA. The columns of U are the left singular vectors of A, and the columns of Vare the right singular vectors of A.
If n = m, S = diag(σi ) =
σ1. . .
σm
. If n 6= m,
S = diag(σi ) ∈ Rm×n must be interpreted as follows (0kl is a k × l matrix withzero entries):
σ1. . . 0m(n−m)
σn
if n > m,
σ1
. . .
σn0(m−n)n
if n < m,
Matrices Reduction of matrices Vector and matrix norms
Truncated Singular Value Decomposition (SVD)
The SVD of A can be written
A = USVH =
min(n,m)∑i=1
σiuivHi
After ordering the singular values by decreasing values (σ1 ≥ σ2 ≥ . . .), matrixA can be approximated by a rank-K matrix AK obtained by a truncation of theSVD:
AK =K∑i=1
σiuivHi
We have the following error estimate:
‖A− AK‖F2‖A‖F
=
min(n,m)∑i=K+1
σ2i
Matrices Reduction of matrices Vector and matrix norms
Illustration: SVD for data compression
Initial image (778× 643) Singular values Rank-10 SVD
Matrices Reduction of matrices Vector and matrix norms
Illustration: SVD for data compression
Initial image (778× 643) Singular values Rank-20 SVD
Matrices Reduction of matrices Vector and matrix norms
Illustration: SVD for data compression
Initial image (778× 643) Singular values Rank-30 SVD
Matrices Reduction of matrices Vector and matrix norms
Illustration: SVD for data compression
Initial image (778× 643) Singular values Rank-40 SVD
Matrices Reduction of matrices Vector and matrix norms
Illustration: SVD for data compression
Initial image (778× 643) Singular values Rank-50 SVD
Matrices Reduction of matrices Vector and matrix norms
Illustration: SVD for data compression
Initial image (778× 643) Singular values Rank-100 SVD
Matrices Reduction of matrices Vector and matrix norms
Part II
Linear algebra
3 Matrices
4 Reduction of matrices
5 Vector and matrix norms
Matrices Reduction of matrices Vector and matrix norms
Vector norms
Denition 27
A norm on vector space V is an application ‖ · ‖ : V → R+ verifying
‖v‖ = 0 if and only if v = 0
‖αv‖ = |α|‖v‖ for all v ∈ V and ∀α ∈ K‖u + v‖ 6 ‖u‖+ ‖v‖ for all u, v ∈ V (triangle inequality)
Example 28 (For V = Kn)
(2-norm) ‖v‖2 =(∑n
i=1|vi |2
)1/2(1-norm) ‖v‖1 =
∑n
i=1|vi |
(∞-norm) ‖v‖∞ = maxi∈1,...,n |vi |
(p-norm) ‖v‖p =(∑n
i=1|vi |p
)1/pfor p > 1.
Matrices Reduction of matrices Vector and matrix norms
Useful inequalities
(·, ·) denote the canonical inner product.
Theorem 29 (Cauchy-Schwartz inequality)
|(u, v)| 6 ‖u‖2‖v‖2
Theorem 30 (Hölder's inequality)
Let 1 ≤ p, q ≤ ∞ such that 1
p+ 1
q= 1, then
|(u, v)| 6 ‖u‖p‖v‖q
Theorem 31 (Minkowski inequality)
Let 1 6 p 6∞, then‖u + v‖p 6 ‖u‖p + ‖v‖p
Minkowski inequality is in fact the triangular inequality for the norm ‖ · ‖p.
Matrices Reduction of matrices Vector and matrix norms
Matrix norms I
Denition 32
A norm on Km×n is a map ‖ · ‖ : Km×n → R+ which veries
‖A‖ = 0 is and only if A = 0
‖αA‖ = |α|‖A‖ for all A ∈ Km×n and ∀α ∈ K‖A + B‖ 6 ‖A‖+ ‖B‖ for all A,B ∈ Km×n (triangle inequality)
For square matrices (n = m), a matrix norm is a norm which satises thefollowing additional inequality
‖AB‖ 6 ‖A‖‖B‖ for all A ∈ Kn×n, B ∈ Kn×n
An important class of matrix norms is the class of subordinate matrix norms.
Denition 33 (subordinate matrix norm)
Given norms ‖ · ‖ on Kn and Km, we can dene a natural norm on Km×n,subordinate to the vectors norms, and dened by
‖A‖ = maxv∈Cn :v 6=0
‖Av‖‖v‖ = max
v∈Cn :‖v‖61
‖Av‖ = maxv∈Cn :‖v‖=1
‖Av‖
Matrices Reduction of matrices Vector and matrix norms
Matrix norms II
Example 34
When considering classical vector norms on Kn, we have the followingcharacterization of the subordinate norms of a square matrix A ∈ Kn×n:
‖A‖1 = maxv‖Av‖1‖v‖1 = maxj
∑i |aij |
‖A‖∞ = maxv‖Av‖∞‖v‖∞ = maxi
∑j |aij |
‖A‖2 = maxv‖Av‖2‖v‖2 =
√ρ(AHA) =
√ρ(AAH) = ‖AH‖2.
Note that ‖A‖2 corresponds to the dominant singular value of A.
Property 35
For all unitary matrix U (i.e. UUH = I ), we have
‖A‖2 = ‖AU‖2 = ‖UA‖2 = ‖UHAU‖2
If A is normal (i.e. AAH = AHA), then ‖A‖2 = ρ(A).
Matrices Reduction of matrices Vector and matrix norms
Matrix norms III
Theorem 36
Let A be a square matrix and ‖ · ‖ an arbitrary matrix norm. Then
ρ(A) 6 ‖A‖
For ε > 0, there exists at least one subordinate matrix norm such that
‖A‖ 6 ρ(A) + ε
Conditioning Direct methods Iterative methods
Part III
Systems of linear equations
6 Conditioning
7 Direct methodsTriangular systemsGauss eliminationLU factorizationCholesky factorizationHouseholder method and QR factorizationComputational work
8 Iterative methodsGeneralitiesJacobi, Gauss-Seidel, RelaxationProjection methodsKrylov subspace methods
Conditioning Direct methods Iterative methods
The aim is to introduce dierent strategies for the solution of a system of linearequations
Ax = b
with A ∈ Rn×n, b ∈ Rn.
Conditioning Direct methods Iterative methods
Part III
Systems of linear equations
6 Conditioning
7 Direct methodsTriangular systemsGauss eliminationLU factorizationCholesky factorizationHouseholder method and QR factorizationComputational work
8 Iterative methodsGeneralitiesJacobi, Gauss-Seidel, RelaxationProjection methodsKrylov subspace methods
Conditioning Direct methods Iterative methods
Condition number
Let consider the following two systems of equations10 7 8 77 5 6 58 6 10 97 5 9 10
x =
32233331
⇒ x =
1111
10 7 8 77 5 6 58 6 10 97 5 9 10
x =
32.122.933.130.9
⇒ x =
9.2−12.64.5−1.1
We observe that a little modication of the right-hand side leads a largemodication in the solution.If an error is made on the input data (here the right-hand side), the error onthe solution may be drastically amplied.This phenomenon is due to a bad conditioning of the matrix A. It reveals thatfor badly conditioned matrices, the solution of systems of equations obtainedwith nite precision computers has to be considered carefully or even notconsidered as a good solution.
Conditioning Direct methods Iterative methods
Denition 37
Let A ∈ Kn×n be an invertible matrix and let ‖ · ‖ be a matrix normsubordinate to the vector norm ‖ · ‖. The condition number of A is dened as
cond(A) = ‖A‖‖A−1‖
Let b ∈ Kn be the right-hand side of a system and let δA ∈ Kn×n and δb ∈ Kn
be perturbations of matrix A and vector b.
Property 38
If x and xε are solutions of the following systems
Ax = b, Aεxε = bε,
with ‖A− Aε‖ = O(ε) and ‖b − bε‖ = O(ε), then
‖x − xε‖‖x‖ 6 cond(A)
(‖A− Aε‖‖A‖ +
‖b − bε‖‖b‖
)+ O(ε2)
Conditioning Direct methods Iterative methods
Property 39
For every matrix A and every matrix norm, cond(A) > 1,cond(A) = cond(A−1), cond(αA) = cond(A), ∀α 6= 0.
For every matrix A, the condition number cond2(A) = ‖A‖2‖A−1‖2associated with the 2-norm veries
cond2(A) =maxi σi (A)
mini σi (A)
where the σi (A) are the singular values of A.
For a normal matrix A,
cond2(A) =maxi |λi (A)|mini |λi (A)|
where the λi (A) are the eigenvalues of A.
For unitary or orthogonal matrix A, the condition number cond2(A) = 1.
The condition number cond2(A) is invariant trough unitarytransformation: cond2(A) = cond2(AU) = cond2(UA) = cond2(UHAU)for every unitary matrix U.
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Part III
Systems of linear equations
6 Conditioning
7 Direct methodsTriangular systemsGauss eliminationLU factorizationCholesky factorizationHouseholder method and QR factorizationComputational work
8 Iterative methodsGeneralitiesJacobi, Gauss-Seidel, RelaxationProjection methodsKrylov subspace methods
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Principle of direct methods I
For solving
Ax = b,
direct methods consist in determining an invertible matrix M such that
MAx = Mb
is an upper triangular system. This is called the elimination step. Then, asimple backward substitution can be performed to solve this triangular system.
Do not compute the inverse !!!
In practice, the solution x of Ax = b is not obtained by rst computing theinverse A−1 and then computing the matrix-vector product A−1b. Indeed, itwould be equivalent to solving n systems of linear equations.
For simplicity, we use sometimes the notation M−1x but the inverse is nevercomputed in practise. This operation corresponds to the solution of a system ofequations (generally easy due to properties of M: diagonal, triangular).
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Part III
Systems of linear equations
6 Conditioning
7 Direct methodsTriangular systemsGauss eliminationLU factorizationCholesky factorizationHouseholder method and QR factorizationComputational work
8 Iterative methodsGeneralitiesJacobi, Gauss-Seidel, RelaxationProjection methodsKrylov subspace methods
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Triangular systems of equations I
If A is lower triangular, the systema11 0 . . . 0
a21 a22. . .
......
.... . . 0
an1 an2 . . . ann
x1
...xn
=
b1...bn
is solved by a forward substitution
Algorithm 40 (Forward substitution for lower triangular system)
Step 1. a11x1 = b1
Step 2. a22x2 = −a21x1...
Step n. annxn = bn −∑n−1
j=1anjbj
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Triangular systems of equations II
If A is upper triangular, the systema11 a12 . . . a1n0 a22 . . . a2n...
. . .. . .
...0 . . . 0 ann
x1
...xn
=
b1...bn
is solved by a backward substitution
Algorithm 41 (Backward substitution for upper triangular system)
Step 1. annxn = bn
Step 2. an−1,n−1xn−1 = −an−1,nxn...
Step n. a11x1 = b1 −∑n
j=2a1jbj
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Part III
Systems of linear equations
6 Conditioning
7 Direct methodsTriangular systemsGauss eliminationLU factorizationCholesky factorizationHouseholder method and QR factorizationComputational work
8 Iterative methodsGeneralitiesJacobi, Gauss-Seidel, RelaxationProjection methodsKrylov subspace methods
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Gauss elimination I
Denition 42 (Pivoting matrix)
A pivoting matrix P(i , j), associated with a linear mapping written in a basisE = (ei )
ni=1, is dened as follows
P(i , j) = I − (ei − ej )(ei − ej )H
For A ∈ Kn×n, P(i , j)A is the matrix A with permuted lines i and j , andAP(i , j) is the matrix A with permuted columns i and j . Let us note thatP(i , i) = I .
We now describe the Gauss elimination procedure
Step 1.
Let A = A1 = (a1ij ). Select a nonzero element a1i∗1 of the rst column andpermute the lines 1 and i∗. Let P1 = P(1, i∗) and set
A1 = P1A1 = (a1ij )
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Gauss elimination II
Let introduce the matrix
E1 =
1
− a121a111
. . .
.... . .
− a1n1a111
1
such that
A2 = E1A1 =
a111 a112 . . . a11n0 a222 . . . a22n...
......
0 a2n2 . . . a2nn
Step 2.
We have det(A2) = det(E1P1A1) = det(E1)det(P1)det(A) = ±det(A)(−det(A) if a line permutation has been made, +det(A) if not). Therefore A2
is invertible, and so is the submatrix (A2)ij , 2 6 i , j 6 n. We can then operateas in step 1 for this submatrix for eliminating the subdiagonal elements of
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Gauss elimination III
column 2: introduce a permutation matrix P2 = P(2, i∗), with i∗ > 2, and aline operation matrix E2, and let A2 = P3A2 and A3 = E3A2.
Step k − 1.After k − 1 steps, we have the matrix
Ak = Ek−1Pk−1 . . .E1P1A1 =
ak11 ak12 . . . . . . . . . ak1nak22 . . . . . . . . . ak2n
. . ....
akkk . . . akkn...
...
aknk . . . aknn
After an eventual pivoting with a pivoting matrix Pk , we dene Ak = PkAk andAk+1 = Ek Ak with
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Gauss elimination IV
Ek =
1. . .
1
−akk+1,k
akkk
. . .
.... . .
− aknk
akkk
1
Last step
After n − 1 steps, by we obtain an upper triangular matrix
An = En−1Pn−1 . . .E1P1A
The invertible matrix M = En−1Pn−1 . . .E1P1 is then an invertible matrix suchthat MA is upper triangular.
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Gauss elimination V
Remark. Choice of pivoting
In order to avoid dramatic roundo errors with nite precision computers, weadopt one of the following pivoting strategies.
Partial pivoting. At step k, we select Pk = P(k, i∗) such that|aki∗k | = max
k6i6n|akik |
Total pivoting. At step k, we select i∗ and j∗ such that|aki∗j∗ | = max
i>k,j6n|akij | and we permute lines and columns by dening
Ak = P(k, i∗)AkP(j∗, k).
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Gauss elimination VI
Remark. Computing the determinant of a matrix
The Gauss elimination is an ecient technique for computing the determinantof a matrix. Indeed,
det(A) = det(An)det(M)−1 = ±n∏i=1
anii
where the sign depends on the number of pivoting operations that have beenperformed.
Remark.
In practice, for solving a system Ax = b, we don't compute the matrix M. Werather operate simultaneously on b by computing
Mb = bn = En−1Pn−1 . . .E1P1b
Then, we solve the triangular system MAx = MB, or equivalently Anx = bn.
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Gauss elimination VII
Computational work of Gauss Elimination
O(2
3n3)
For an arbitrary matrix, it seems that this computational work in O(n3) is nearthe optimal that we can expect. That is the reason why Gauss elimination canbe used when no additional information is given on the matrix.
Theorem 43
For A ∈ Kn (inversible or not), there exists at least one invertible matrix Msuch that MA is an upper triangular matrix.
Proof.
For A invertible, the Gauss elimination procedure is a constructive proof for thistheorem. Otherwise, the matrix A is singular if and only there exists a matrixAk with elements akik = 0 for k 6 i 6 n. In this case, we can set Ek = I andPk = I at step k of the Gauss elimination and go to the next step.
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Part III
Systems of linear equations
6 Conditioning
7 Direct methodsTriangular systemsGauss eliminationLU factorizationCholesky factorizationHouseholder method and QR factorizationComputational work
8 Iterative methodsGeneralitiesJacobi, Gauss-Seidel, RelaxationProjection methodsKrylov subspace methods
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
LU factorization I
The LU factorization of a matrix consists in constructing lower and uppertriangular matrices L and U such that A = LU. In fact, this factorization isobtained by the Gauss elimination procedure.Let us consider the Gauss elimination without pivoting, i.e. by letting Ak = Ak .It is possible if at step k, akkk 6= 0. We then let
M = En−1 . . .E1
and obtainMA = U
where U is the desired upper triangular matrixa111 a112 . . . a11n
a222 . . . a22n. . .
...annn
M being a product of lower triangular matrices, it is a lower triangular matrixand so is its inverse M−1. We then have the desired decomposition with
L = M−1 = E−11 . . .E−1n−1
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
LU factorization II
Matrix L = (lij ) is directly obtained from matrices Ek
Ek =
1. . .
1
−lk+1,k
. . ....
. . .
−lnk 1
, E−1k =
1. . .
1
lk+1,k
. . ....
. . .
lnk 1
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
LU factorization III
Theorem 44
Let A ∈ Kn×n be such that the diagonal submatricesa11 . . . a1k...
...ak1 . . . akk
∈ Kk×k are invertible. Then, there exists a lower triangular
matrix L and an upper triangular matrix U such that
A = LU
If we further impose that the diagonal elements of L are equal to 1, thisdecomposition is unique.
Proof.
The condition on the invertibility of submatrices ensures that at step k, thediagonal term akkk is nonzero and therefore that pivoting can be omitted.
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Part III
Systems of linear equations
6 Conditioning
7 Direct methodsTriangular systemsGauss eliminationLU factorizationCholesky factorizationHouseholder method and QR factorizationComputational work
8 Iterative methodsGeneralitiesJacobi, Gauss-Seidel, RelaxationProjection methodsKrylov subspace methods
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Cholesky factorization I
Theorem 45
If A ∈ Rn×n is a symmetric denite positive matrix, there exists at least onelower triangular matrix B = (bij ) ∈ Rn×n such that
A = BBT
If we further impose that the diagonal elements bii > 0, the decomposition isunique.
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Cholesky factorization II
Proof.
We simply show that the diagonal submatrices ∆k = (aij ), 1 6 i , j 6 k, arepositive denite. Therefore, they are invertible and there exists a unique LUfactorization A = LU such that L has unit diagonal terms. Since the ∆k arepositive denite, we have
∏k
i=1uii = det(∆kk) > 0, for all k > 1. We then
dene the diagonal matrix D = diag(√uii ) and we write
A = (LΛ)(Λ−1U) = BC
where B = LΛ and C = Λ−1U have both diagonal terms bii = cii =√uii . The
symmetry of matrix A imposes that BC = CTBT and therefore
CB−T =
1 × . . . ×
1 . . . ×. . .
...1
=
1× 1...
. . .
× . . . × 1
= B−1CT
and this last equality is only possible if CB−T = I ⇒ C = BT . (Prove theuniqueness of the decomposition).
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Part III
Systems of linear equations
6 Conditioning
7 Direct methodsTriangular systemsGauss eliminationLU factorizationCholesky factorizationHouseholder method and QR factorizationComputational work
8 Iterative methodsGeneralitiesJacobi, Gauss-Seidel, RelaxationProjection methodsKrylov subspace methods
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Householder matrices
Denition 46
For v a nonzero vector in Cn, we introduce the following matrix, calledHouseholder matrix associated with v :
H(v) = I − 2vvH
vHv
We will consider, although incorrect, that the identity I is a Householder matrix.
Theorem 47
For x = (xi )ni=1 ∈ Cn, there exists two householder matrices H such that
(Hx)i = 0 for i > 2.
Proof.
Denoting by e1 the rst basis vector of Cn, one veries that the twohouseholder matrices H(v) are associated with the vectors v = x ± ‖x‖2e iαe1,where α ∈ R is the argument of x1 ∈ C, i.e. x1 = |x1|e iα, and we have
H(v)x = ∓‖x‖2e1
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Householder method I
The Householder method for solving Ax = b consists in nding n − 1householder matrices Hin−1i=1
such that Hn−1 . . .H1A is upper triangular.Then, we solve the following triangular system by backward substitution:
Hn−1 . . .H1Ax = Hn−1 . . .H1b
Suppose that Ak = Hk−1 . . .H1A is under the form
Ak =
ak11 ak12 . . . . . . . . . ak1na222 . . . . . . . . . a22n
. . ....
akkk . . . akkn...
...
aknk . . . aknn
Let c = (ci )
n−k+1
i=1∈ Cn−k+1 be the vector with components ci = aki+k−1. There
exists a Householder matrix H(vk), with vk ∈ Cn−k+1, such that H(vk)c has
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Householder method II
zero components except the rst one. Then, we denote vk =
(0vk
)∈ Cn and
we let Hk = H(vk) the householder matrix associated with vk . Let us note that
Hk = H(vk) =
(Ik−1 00 H(vk)
)Performing this operation for k = 1 . . . n − 1, we obtain the desired uppertriangular matrix An = Hn−1 . . .H1A.
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
QR factorization I
The QR factorization is a matrix interpretation of the Householder method.
Theorem 48
For A ∈ Kn×n, there exist a unitary matrix Q ∈ Kn×n and an upper triangularmatrix R ∈ Kn×n such that
A = QR
Moreover, one can choose the diagonal elements of R > 0. Then, if A isinvertible, the corresponding QR factorization is unique.
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
QR factorization II
Proof.
The previous householder construction proves the existence of an uppertriangular matrix
R = Hn−1 . . .H1A
where the Hi are householder matrices. The matrix
Q = (Hn−1 . . .H1)−1 = H−11 . . .H−1n−1 = H1 . . .Hn−1
is unitary (recall that the Hk are unitary and hermitian, i.e. H−1k = HHk = Hk).
This proves this existence of a QR decomposition. Let now denote by αi ∈ Rthe arguments of the diagonal elements rkk = |rkk |e iαk and let D = diag(e iαk ).The matrix Q = QD is still unitary and the matrix R = D−1R is still uppertriangular with all its diagonal elements greater than 0. We then have theexistence of a QR factorization A = QR with rkk > 0. We can then show theuniqueness of this decomposition (let as an exercice).
Remark.
If A ∈ Rn×n, Q,R ∈ Rn×n, with Q an orthogonal matrix.
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Part III
Systems of linear equations
6 Conditioning
7 Direct methodsTriangular systemsGauss eliminationLU factorizationCholesky factorizationHouseholder method and QR factorizationComputational work
8 Iterative methodsGeneralitiesJacobi, Gauss-Seidel, RelaxationProjection methodsKrylov subspace methods
Conditioning Direct methods Iterative methods Triangular systems Gauss elimination LU factorization Cholesky factorization Householder method and QR factorization Computational work
Computational complexity
With classical algorithms...
Algorithm Operations
LU O( 23n3)
Cholesky O( 13n3)
QR O( 23n3)
Conditioning Direct methods Iterative methods Generalities Jacobi, Gauss-Seidel, Relaxation Projection methods Krylov subspace methods
Part III
Systems of linear equations
6 Conditioning
7 Direct methodsTriangular systemsGauss eliminationLU factorizationCholesky factorizationHouseholder method and QR factorizationComputational work
8 Iterative methodsGeneralitiesJacobi, Gauss-Seidel, RelaxationProjection methodsKrylov subspace methods
Conditioning Direct methods Iterative methods Generalities Jacobi, Gauss-Seidel, Relaxation Projection methods Krylov subspace methods
Part III
Systems of linear equations
6 Conditioning
7 Direct methodsTriangular systemsGauss eliminationLU factorizationCholesky factorizationHouseholder method and QR factorizationComputational work
8 Iterative methodsGeneralitiesJacobi, Gauss-Seidel, RelaxationProjection methodsKrylov subspace methods
Conditioning Direct methods Iterative methods Generalities Jacobi, Gauss-Seidel, Relaxation Projection methods Krylov subspace methods
Basic iterative methods I
For the solution of a linear system of equations Ax = b, basic iterative methodsconsist in constructing a sequence xkk≥0 dened by
xk+1 = Bxk + c
from an initial vector x0. Matrix B and vector c are to be dened such that theiterative method converges towards the solution x , i.e.
limk→∞
xk = x
B and c are chosen such that I − B is invertible and such that x is the uniquesolution of x = Bx + c.
Theorem 49
Let B ∈ Kn×n. The following assertions are equivalent
(1) limk→∞ Bk = 0
(2) limk→∞ Bkv = 0 ∀v(3) ρ(B) < 1
(4) ‖B‖ < 1 for at least one subordinate matrix norm ‖ · ‖
Conditioning Direct methods Iterative methods Generalities Jacobi, Gauss-Seidel, Relaxation Projection methods Krylov subspace methods
Basic iterative methods II
Proof.
(1)⇒ (2). ‖Bkv‖ ≤ ‖Bk‖‖v‖ −→k→∞
0
(2)⇒ (3). If ρ(B) ≥ 1, there exists a vector v 6= 0 such that Bv = λv with|λ| ≥ 1 and then Bkv = λkv does not converge towards 0, a contradiction.(3)⇒ (4). Consequence of theorem 36(4)⇒ (1). ‖Bk‖ ≤ ‖B‖k −→
k→∞0.
Theorem 50
The following assertions are equivalent
(i) The iterative method is convergent
(ii) ρ(B) < 1
(iii) ‖B‖ < 1 for at least one subordinate matrix norm ‖ · ‖
Proof.
The iterative method is convergent if and only if limk→∞ ek = 0, withek = xk − x = Bke0. The proof then results from theorem 49.
Conditioning Direct methods Iterative methods Generalities Jacobi, Gauss-Seidel, Relaxation Projection methods Krylov subspace methods
Part III
Systems of linear equations
6 Conditioning
7 Direct methodsTriangular systemsGauss eliminationLU factorizationCholesky factorizationHouseholder method and QR factorizationComputational work
8 Iterative methodsGeneralitiesJacobi, Gauss-Seidel, RelaxationProjection methodsKrylov subspace methods
Conditioning Direct methods Iterative methods Generalities Jacobi, Gauss-Seidel, Relaxation Projection methods Krylov subspace methods
Jacobi, Gauss-Seidel, Relaxation (SOR) I
We decompose A under the form
A = M − N
where M is an invertible matrix and then
Ax = b ⇔ Mx = Nx + b
and we compute the sequence
xk+1 = M−1Nxk + M−1b ≡ Bxk + c
In practice, at each iteration, we solve the system Mxk+1 = Nxk + b. Themethod is then ecient if M have a simple form (diagonal or triangular).
Denition 51
We decompose A = D − E − F where D is the diagonal part of A, −E and −Fits strict lower and upper parts.
Conditioning Direct methods Iterative methods Generalities Jacobi, Gauss-Seidel, Relaxation Projection methods Krylov subspace methods
Jacobi, Gauss-Seidel, Relaxation (SOR) II
Denition 52 (Jacobi)
M = D, N = E + F
Denition 53 (Gauss-Seidel)
M = D − E , N = F
Denition 54 (Successive Over Relaxation (SOR))
M = ω−1D − E , N = ω−1(1− ω)D + F
Conditioning Direct methods Iterative methods Generalities Jacobi, Gauss-Seidel, Relaxation Projection methods Krylov subspace methods
Convergence results I
Theorem 55
Let A a positive denite hermitian matrix, decomposed under the formA = M −N with M invertible. If the matrix (MH + N) is positive denite, thenρ(M−1N) < 1.
Conditioning Direct methods Iterative methods Generalities Jacobi, Gauss-Seidel, Relaxation Projection methods Krylov subspace methods
Convergence results II
Proof.
From theorem 36, we know that it suces to nd a matrix norm for which‖M−1N‖ < 1. We will show this property for the matrix norm subordinate tothe vector norm ‖v‖ =
√vHAv . Let rst note that (MH +N) is hermitian since
(MH + N)H = M + NH = A + N + NH = AH + NH + N = MH + N.
We have‖M−1N‖ = ‖I −M−1A‖ = sup
‖v‖=1
‖v −M−1Av‖
Denoting w = M−1Av , we have, for v such that ‖v‖ = 1,
‖v − w‖2 = 1− vHAw − wHAv + wHAw
= 1− wHMHw − wHMw + wHAw = 1− wH(MH + N)w︸ ︷︷ ︸>0
Therefore ‖v‖ = 1⇒ ‖v −M−1Av‖ < 1. The functionv ∈ Cn 7→ ‖v −M−1Av‖ ∈ R is continuous on the unit sphere, which is acompact set, and therefore the supremum is reached.
Conditioning Direct methods Iterative methods Generalities Jacobi, Gauss-Seidel, Relaxation Projection methods Krylov subspace methods
Convergence results III
Theorem 56 (Sucient condition for convergence of relaxation)
If A is hermitian positive denite, relaxation method converges if 0 < ω < 2.
Proof.
We show that MH + N = 2−ωω
D. Since A is denite positive, we have for the
canonical basis vectors vi , vHi Avi = vHi Dvi > 0. Matrix MH + N is then
hermitian positive denite if and only if 0 < ω < 2, and the proof ends withtheorem 55.
Theorem 57 (Necessary condition for convergence of relaxation)
The spectral radius of the matrix Bω = M−1N of the relaxation method veries
ρ(Bω) ≥ |ω − 1|
and therefore, relaxation method converges only if 0 < ω < 2.
Conditioning Direct methods Iterative methods Generalities Jacobi, Gauss-Seidel, Relaxation Projection methods Krylov subspace methods
Convergence results IV
Proof.
We haveBω = (ω−1D − E)−1(ω−1(1− ω)D + F )
and then
det(Bω) = (1− ω)n =n∏i=1
λi (Bω)
Then
ρ(Bω) ≥
(n∏i=1
λi (Bω)
)1/n
= |1− ω|
Conditioning Direct methods Iterative methods Generalities Jacobi, Gauss-Seidel, Relaxation Projection methods Krylov subspace methods
Part III
Systems of linear equations
6 Conditioning
7 Direct methodsTriangular systemsGauss eliminationLU factorizationCholesky factorizationHouseholder method and QR factorizationComputational work
8 Iterative methodsGeneralitiesJacobi, Gauss-Seidel, RelaxationProjection methodsKrylov subspace methods
Conditioning Direct methods Iterative methods Generalities Jacobi, Gauss-Seidel, Relaxation Projection methods Krylov subspace methods
Projection methods I
We consider a real system of equations Ax = b. Projection techniques consistsin searching an approximate solution x in a subspace V of Rn. Theapproximate solution is then dened by
x ∈ V, b − Ax ⊥ W
where W is a subspace of Rn with the same dimension of V. The approximatesolution is then dened by orthogonality constraints on the residual. x is calleda projection of x onto the subspace V and parallel to subspace W. The caseV =W corresponds to an orthogonal projection and the orthogonalityconstraint is called Galerkin orthogonality. The case V 6=W corresponds to anoblique projection and the orthogonality constraint is called Petrov-Galerkinorthogonality.Let V = (v1, . . . , vm) and W = (w1, . . . ,wm) dene bases of V and W, theapproximation is then dened by x = Vy , with y ∈ Rm such that
WTAVy = WTb ⇒ y = (WTAV )−1WTb
Conditioning Direct methods Iterative methods Generalities Jacobi, Gauss-Seidel, Relaxation Projection methods Krylov subspace methods
Projection methods II
Projection method
Until convergence
1 Select V = (v1, . . . , vm) and W = (w1, . . . ,wm)
2 r = b − Ax
3 y = (WTAV )−1WT r
4 x = x + Vy
Subspaces must be chosen such that WTAV is nonsingular. Two importantparticular choices satises this property.
Theorem 58
WTAV is nonsingular for either one the following conditions
A is positive denite and V =WA is nonsingular and W = AV.
Conditioning Direct methods Iterative methods Generalities Jacobi, Gauss-Seidel, Relaxation Projection methods Krylov subspace methods
Projection methods III
Theorem 59
Assume that A is symmetric denite positive and V =W. Then, x ∈ V is suchthat Ax − b ⊥ V if and only if
‖x − x‖2A = minx∈V‖x − x‖2A, ‖x‖2A = xTAx
Theorem 60
Let A a nonsingular matrix and W = AV. Then, x ∈ V is such thatAx − b ⊥ W if and only if it minimizes the 2-norm of the residual
‖b − Ax‖2 = minx∈V‖b − Ax‖2
Conditioning Direct methods Iterative methods Generalities Jacobi, Gauss-Seidel, Relaxation Projection methods Krylov subspace methods
Basic one-dimensional projection algorithms I
Basic one-dimensional projection schemes consist in selecting V and W withdimension 1. Let us denote V = spanv and W = spanw. Denotingr = b − Axk the residual at iteration k, the next iterate is dened by
xk+1 = xk + αv , α =(w , r)
(w ,Av)=
wT r
wTAv
Denition 61 (Steepest descent)
We let v = r and w = r . We then have
xk+1 = xk + αr , α =(r , r)
(Ar , r)
If A is symmetric positive denite matrix, xk+1 is the solution of
minα
f (xk + αr), f (x) = ‖x − x‖2A = (x − x ,A(x − x))
We note that −∇f (xk) = A(x − xk) = b − Axk = r , and thereforexk+1 = xk − α∇f (xk). It then corresponds to a steepest descent algorithm forminimizing the convex function f (x), with an optimal choice of step α.
Conditioning Direct methods Iterative methods Generalities Jacobi, Gauss-Seidel, Relaxation Projection methods Krylov subspace methods
Basic one-dimensional projection algorithms II
Theorem 62 (Convergence of steepest descent)
If A is symmetric positive denite matrix, the steepest descent algorithmconverges.
Denition 63 (Minimal residual)
We let v = r and w = Ar . We then have
xk+1 = xk + αr , α =(Ar , r)
(Ar ,Ar)
which is the solution ofminα‖b − A(xk + αr)‖2
Theorem 64
If A is positive denite, minimal residual algorithm converges.
Conditioning Direct methods Iterative methods Generalities Jacobi, Gauss-Seidel, Relaxation Projection methods Krylov subspace methods
Basic one-dimensional projection algorithms III
Denition 65 (Residual norm steepest descent)
We let v = AT r and w = Av = AAT r . We then have
xk+1 = xk + αAT r , α =(Av , r)
(Av ,Av)=‖v‖2
‖Av‖2
which is the solution of
minα
f (xk + αv), f (x) = ‖b − Ax‖2 = (Ax − b,Ax − b)
Note that −∇f (xk) = AT (b − Axk) = AT r = v . It then corresponds to asteepest descent algorithm on convex function f (x), with an optimal choice ofstep α.
Theorem 66
If A is nonsingular, residual norm steepest descent algorithm converges.
Conditioning Direct methods Iterative methods Generalities Jacobi, Gauss-Seidel, Relaxation Projection methods Krylov subspace methods
Part III
Systems of linear equations
6 Conditioning
7 Direct methodsTriangular systemsGauss eliminationLU factorizationCholesky factorizationHouseholder method and QR factorizationComputational work
8 Iterative methodsGeneralitiesJacobi, Gauss-Seidel, RelaxationProjection methodsKrylov subspace methods
Conditioning Direct methods Iterative methods Generalities Jacobi, Gauss-Seidel, Relaxation Projection methods Krylov subspace methods
Krylov subspace methods
Krylov subspace methods are projection methods which consists in deningsubspace V as the m-dimensional Krylov subspace of matrix A, associated withr0 = b − Ax0, where x0 is an initial guess. This Krylov subspace is dened by
V = Km(A, r0) = spanr0,Ar0, . . . ,Am−1r0
The dierent Krylov subspace methods dier from the choice of space W andfrom the choice of a preconditioner. First class of methods consisting in takingW = Km(A, r0) or W = AKm(A, r0). Second class of methods consisting intaking W = Km(AT , r0).
A complete reference about iterative methods
Yousef Saad.Iterative Methods for Sparse Linear Systems.SIAM, 2003.
Jacobi Givens-Householder QR Power iterations Krylov
Part IV
Eigenvalue problems
9 Jacobi method
10 Givens-Householder method
11 QR method
12 Power iterations
13 Methods based on Krylov subspaces
Jacobi Givens-Householder QR Power iterations Krylov
Eigenvalue problems
The aim is to present dierent techniques for nding the eigenvalues andeigenvectors (λi , vi ) of a matrix A:
Avi = λivi
Jacobi Givens-Householder QR Power iterations Krylov
Part IV
Eigenvalue problems
9 Jacobi method
10 Givens-Householder method
11 QR method
12 Power iterations
13 Methods based on Krylov subspaces
Jacobi Givens-Householder QR Power iterations Krylov
Jacobi method I
Jacobi method allows to nd all the eigenvalues of a symmetric matrix A. It iswell adapted to full matrices.There exists an orthogonal matrix O such that OTAO = diag(λ1, . . . , λn),where the λi are the eigenvalues of A, distinct or not. The Jacobi methodconsists in constructing a sequence of elementary orthogonal matrices (Ωk)k≥1such that the sequence (Ak)k≥1, dened by
Ak+1 = ΩTk AkΩk = (Ω1 . . .Ωk)TAk(Ω1 . . .Ωk) = OT
k AOk
converges towards the diagonal matrix diag(λ1, . . . , λn) (with an eventualpermutation).Each transformation Ak → Ak+1 consists in eliminating two symmetricextra-diagonal terms by a rotation. Let A = Ak and B = Ak+1. The matrix Ωk
is selected as follows
Ωk = I + (cos(θ)− 1)(epeTp + eqe
Tq ) + sin(θ)epe
Tq − sin(θ)eqe
Tp
where θ ∈ (−π/4, π/4)\0 is the unique angle such that bpq = bqp = 0. θ issolution of
cotan(2θ) =aqq − app2apq
Jacobi Givens-Householder QR Power iterations Krylov
Jacobi method II
Theorem 67 (Convergence of eigenvalues)
The sequence (Ak)k≥1 obtained with the Jacobi method converges and
limk→∞
Ak = diag(λσ(i))
where σ is a permutation of 1, ..., n.
Theorem 68 (Convergence of eigenvectors)
We suppose that all eigenvalues of A are distinct. Then, the sequence (Ok)k≥1in the Jacobi method converges to an orthogonal matrix whose columns forman orthonormal set of eigenvectors of A.
Jacobi Givens-Householder QR Power iterations Krylov
Part IV
Eigenvalue problems
9 Jacobi method
10 Givens-Householder method
11 QR method
12 Power iterations
13 Methods based on Krylov subspaces
Jacobi Givens-Householder QR Power iterations Krylov
Givens-Householder method I
Givens-Householder method is adapted to the research of selected eigenvaluesof a symmetric matrix A, such as the eigenvalues lying in a given interval.Two steps
1 Determine an orthogonal matrix P such that PTAP is tridiagonal, withthe Householder method.
2 Compute the eigenvalues of a tridiagonal symmetric matrix with theGivens method.
Theorem 69
For a symmetric matrix A, there exists an orthogonal matrix P, product of n− 2Householder matrices Hk such that PTAP is tridiagonal: P = H1H2 . . .Hn−2
HT1 AH1 =
× × 0 0 . . .× × × × . . .0 × × × . . .0 × × × . . .
.
.
....
.
.
.
, HT2 H
T1 AH1H2 =
× × 0 0 . . .× × × 0 . . .0 × × × . . .0 0 × × . . .
.
.
....
.
.
.
...
Jacobi Givens-Householder QR Power iterations Krylov
Part IV
Eigenvalue problems
9 Jacobi method
10 Givens-Householder method
11 QR method
12 Power iterations
13 Methods based on Krylov subspaces
Jacobi Givens-Householder QR Power iterations Krylov
QR method I
The most commonly used method to compute the whole set of eigenvalues ofan arbitrary matrix A, even nonsymmetric.
QR algorithm
Let A1 = A. For k ≥ 1, perform until convergence
Ak = QkRk (QR factorization)
Ak+1 = RkQk
All matrices Ak are similar to matrix A. Under certain conditions, the matrixAk converges towards a triangular matrix which is the Schur form of A, whosediagonal terms are the eigenvalues of A.
Jacobi Givens-Householder QR Power iterations Krylov
Part IV
Eigenvalue problems
9 Jacobi method
10 Givens-Householder method
11 QR method
12 Power iterations
13 Methods based on Krylov subspaces
Jacobi Givens-Householder QR Power iterations Krylov
Power iterations method I
Power iteration method allows the capture of the dominant (largest magnitude)eigenvalue and associated eigenvector of a real matrix A.
Power iteration algorithm
Start with an arbitrary normalized vector x (0) and compute the sequence
x (k+1) =Ax (k)
‖Ax (k)‖
andβ(k+1) = (Ax (k), x (k))
Theorem 70
If the dominant eigenvalue is real and of multiplicity 1, the sequences (x (k))k≥0and (β(k))k≥0 respectively converge towards the dominant eigenvector andeigenvalue.
Jacobi Givens-Householder QR Power iterations Krylov
Power iterations method II
Proof.
Let us prove the convergence of the method when A is symmetric. Then, thereexists an orthonormal basis of eigenvectors (v1, . . . , vn), associated witheigenvalues (λ1, . . . , λn). Let us consider that |λ1| > |λi | for all i > 1. Theinitial vector x (0) can be decomposed on this basis: x (0) =
∑n
i=1aivi and then,
since Avi = λivi ,
x (k) =Ax (k−1)
‖Ax (k)‖=
Akx (0)
‖Akx (0)‖
Akx (0) =n∑i=1
aiλki vi = a1λ
k1w
(k), w (k) =
(v1 +
n∑i=2
aia1
(λiλ1
)k
vi
)and since w (k) → v1, we obtain
x (k) =a1λ
k1w
(k)
‖a1λk1w (k)‖−→k→∞
sign(a1λk1)v1, β(k) −→
k→∞(Av1, v1) = λ1
Let us note that for general matrices, a proof using the Jordan form can beused.
Jacobi Givens-Householder QR Power iterations Krylov
Power iterations method III
Exercice. Power method with deation
Under certain conditions, Power method with deation allows to compute thewhole set of eigenvalues of a matrix. See exercices.
Denition 71 (Inverse power method)
For an invertible matrix A, applying the power method to matrix A−1 allows toobtain the eigenvalue of A with smallest magnitude and the associatedeigenvector (if the smallest magnitude eigenvalue is of multiplicity 1).
Denition 72 (Shifted inverse power method)
The shifted inverse power method consists in applying the inverse powermethod to the shifted matrix Aσ = (A− σI ). It allows the capture of theeigenvalue (and associated eigenvector) which is the closest from the value σ.Indeed, if we denote by (vi , λi ) the eigenpairs of matrix A, Aσ has foreigenpairs (vi , λi − σ). Therefore the inverse power method on Aσ willconverge towards the eigenvalue (λi − σ) such that |λi − σ| = minj |λj − σ|.
Jacobi Givens-Householder QR Power iterations Krylov
Part IV
Eigenvalue problems
9 Jacobi method
10 Givens-Householder method
11 QR method
12 Power iterations
13 Methods based on Krylov subspaces
Jacobi Givens-Householder QR Power iterations Krylov
Methods based on Krylov subspaces
A complete reference for the solution of eigenvalue problems
Yousef Saad.Numerical Methods For Large Eigenvalue Problems.SIAM, 2011.
Fixed point Monotone operators Dierential calculus Newton method
Part V
Nonlinear equations
14 Fixed point theorem
15 Nonlinear equations with monotone operators
16 Dierential calculus for nonlinear operators
17 Newton method
Fixed point Monotone operators Dierential calculus Newton method
Solving nonlinear equations
The aim is to introduce dierent techniques for nding the solution u of anonlinear equation
A(u) = b, u ∈ K ⊂ V
where K is a subset of a vector space V and A : K → V is a nonlinear mapping.We will equivalently consider the nonlinear equation
F (u) = 0, u ∈ K ⊂ V
where F : K → V .
Fixed point Monotone operators Dierential calculus Newton method
Innite dimensional framework
Denition 73
A Banach space V is a complete normed vector space. That means that this isa vector space (on complex or real elds) equipped with a norm ‖ · ‖ and suchthat every Cauchy sequence with respect to this norm has a limit in V .
Denition 74
A Hilbert space is a Banach space V whose norm ‖ · ‖ is associated with anscalar (or hermitian) product (·, ·), with ‖v‖2 = (v , v).
Example 75
V = Rn equipped the natural euclidian scalar product is a nite-dimensionalHilbert space.V = Cn equipped the natural hermitian product is a nite-dimensional Hilbertspace on complex eld.
Fixed point Monotone operators Dierential calculus Newton method
Part V
Nonlinear equations
14 Fixed point theorem
15 Nonlinear equations with monotone operators
16 Dierential calculus for nonlinear operators
17 Newton method
Fixed point Monotone operators Dierential calculus Newton method
Fixed point theorem I
We here consider nonlinear problems under the form
T (u) = u, u ∈ K ⊂ V (1)
where T : K → V is a nonlinear operator.
Denition 76
A solution u of the equation T (u) = u is called a xed point of mapping T .
We are interested in the existence of a solution to equation (1) and in thepossibility of approaching this solution by the following sequence (uk)k≥0dened by
uk+1 = T (uk)
Remark.
Let us note that nonlinear equations F (u) = 0 can be recasted (in dierentways) in the form (1), by letting
T (u) = F (u) + u, T (u) = αF (u) + u, . . .
Fixed point Monotone operators Dierential calculus Newton method
Fixed point theorem II
Denition 77
Let V be a Banach space endowed with a norm ‖ · ‖. A mappingT : K ⊂ V → V is said
contractive if there exists a constant α, with 0 ≤ α < 1, such that
‖T (u)− T (v)‖ ≤ α‖u − v‖ ∀u, v ∈ K
α is called the contractivity constant.
non-expansive if
‖T (u)− T (v)‖ ≤ ‖u − v‖ ∀u, v ∈ K
Lipschitz continuous if there exists a constant β ≥ 0 such that
‖T (u)− T (v)‖ ≤ β‖u − v‖ ∀u, v ∈ K
β is called the Lipschitz-continuity constant.
Fixed point Monotone operators Dierential calculus Newton method
Fixed point theorem III
Theorem 78 (Banach xed-point theorem)
Assume that K is a closed set in a Banach space V and that T : K → K is acontractive mapping with contractivity constant α. Then, we have thefollowing results:
There exists a unique u ∈ K such that T (u) = u
For any u0 ∈ K, the sequence (uk)k≥0 in K, dened by uk+1 = T (uk),converges to u, i.e.
‖u − uk‖ −→k→∞
0
Fixed point Monotone operators Dierential calculus Newton method
Fixed point theorem IV
Proof.
Let us prove that uk is a Cauchy sequence. We have
‖uk+1 − uk‖ = ‖T (uk )− T (uk−1)‖ ≤ α‖uk − uk−1‖ ≤ αk‖u1 − u0‖
for m ≥ k ≥ 1, we then have
‖um − uk‖ ≤m−1∑i=k
‖ui+1 − ui‖ ≤m−1∑i=k
αi‖u1 − u0‖ = ‖u1 − u0‖αk
m−1−k∑i=0
αi
=αk (1− αm−k )
1− α‖u1 − u0‖ ≤
αk
1− α‖u1 − u0‖
Since α ∈ [0, 1), ‖um − uk‖ → 0 as m, k →∞, and therefore, uk is a Cauchy sequence.Since the sequence uk is Cauchy in a Banach space V , it converges to some u ∈ V andsince K is closed, the limit u ∈ K . In the relation uk+1 = T (uk ), we take the limit k →∞and obtain u = T (u), by continuity of T . Then, u is a xed point of T .For the uniqueness, suppose that u1 and u2 are two xed points. Then we have
‖u2 − u1‖ = ‖T (u2)− T (u1)‖ ≤ α‖u2 − u1‖
which is possible only if u2 = u1.
Fixed point Monotone operators Dierential calculus Newton method
Fixed point theorem V
Example 79
Let V = R and T (x) = ax + b. If a 6= 1, the sequence xk+1 = T (xk) ischaracterized by
xk = axk−1 + b = akx0 +1− ak
1− ab
If |a| < 1, xk converges to b1−a , which is the unique xed point of T . If |a| > 1,
the sequence diverges. Let us note that
|T (x)− T (x)| = |a||x − x |
and therefore, we have that T is a contractive mapping if |a| < 1.
Fixed point Monotone operators Dierential calculus Newton method
Part V
Nonlinear equations
14 Fixed point theorem
15 Nonlinear equations with monotone operators
16 Dierential calculus for nonlinear operators
17 Newton method
Fixed point Monotone operators Dierential calculus Newton method
Nonlinear equations with monotone operators I
We consider the application of the xed point theorem to the analysis ofsolvability of a class of nonlinear equations
A(u) = b u ∈ V
where V is a Hilbert space and A : V → V is a Lipschitz continuous andstrictly monotone operator.
Denition 80 (Monotone operator)
A mapping A : V → V on a Hilbert space V is said
monotone if(A(u)− A(v), u − v) ≥ 0 ∀u, v ∈ V
strictly monotone if
(A(u)− A(v), u − v) > 0 ∀u, v ∈ V , u 6= v
strongly monotone if there exists a constant α > 0 such that
(A(u)− A(v), u − v) ≥ α‖u − v‖2 ∀u, v ∈ V
α is called the strong monotonicity constant.
Fixed point Monotone operators Dierential calculus Newton method
Nonlinear equations with monotone operators II
Theorem 81
Let V be a Hilbert space and A : V → V a strongly monotone and Lipschitzcontinuous operator, with monotonicity constant α and Lipschitz-continuityconstant β. Then, for any b ∈ V , there exists a unique u ∈ V such that
A(u) = b
Moreover, if A(u1) = b1 and A(u2) = b2, then
‖u1 − u2‖ ≤1
α‖b1 − b2‖
which means that the solution depends continuously on the right-hand side b.
Fixed point Monotone operators Dierential calculus Newton method
Nonlinear equations with monotone operators III
Proof.
The equation A(u) = b is equivalent to Tγ(u) = u, withTγ(u) = u − γ(A(u)− b) for any γ 6= 0. The idea is to prove that there existsa γ such that Tγ : V → V is contractive. The application of Banach xedpoint theorem will then give the existence and uniqueness of a xed point of u,and therefore, the existence and uniqueness of a solution to A(u) = b. We have
‖Tγ(w)− Tγ(v)‖2 = ‖(w − v)− γ(A(w)− A(v))‖2
= ‖w − v‖2 − 2γ(A(w)− A(v),w − v) + γ2‖A(w)− A(v)‖2
≤ (1− 2γα + γ2β2)‖w − v‖2
For γ2 < 2α/β2, we have (1− 2γα + γ2β2) < 1, and Tγ is a contraction.Now if A(u1) = b1 and A(u2) = b2, we have A(u1)− A(u2) = b1 − b2 and
α‖u1−u2‖2 ≤ (A(u1)−A(u2), u1−u2) = (b1−b2, u1−u2) ≤ ‖b1−b2‖‖u1−u2‖
where the second inequality is the Cauchy-Schwartz inequality satised by theinner product of a Hilbert space. This proves the continuity of the solution uwith respect to b.
Fixed point Monotone operators Dierential calculus Newton method
Part V
Nonlinear equations
14 Fixed point theorem
15 Nonlinear equations with monotone operators
16 Dierential calculus for nonlinear operators
17 Newton method
Fixed point Monotone operators Dierential calculus Newton method
Fréchet and Gâteaux derivatives I
Let F : K ⊂ V →W be a nonlinear mapping, where K is a subset of a normedspace V and W a normed space. We denote by L(V ,W ) the set of linearapplications from V to W .
Denition 82 (Fréchet derivative)
F is Fréchet-dierentiable at u if and only if there exists A ∈ L(V ,W ) suchthat
F (u + v) = F (u) + Av + o(‖v‖) as ‖v‖ → 0
A is denoted F ′(u) and is called the Fréchet derivative of F at u. If F isFréchet-dierentiable at all points in K , we denote by F ′ : K ⊂ V → L(V ,W )the Fréchet derivative of F on K .
Property 83
If F admits a Fréchet derivative F ′(u) at u, then F is continuous at u.
Fixed point Monotone operators Dierential calculus Newton method
Fréchet and Gâteaux derivatives II
Denition 84 (Gâteaux derivative)
F is Gâteaux-dierentiable at u if and only if there exists A ∈ L(V ,W ) suchthat
limt→0
F (u + tv)− F (u)
t= Av ∀v ∈ V (2)
A is denoted F ′(u) and is called the Gâteaux derivative of F at u. If F isGâteaux-dierentiable at all points in K , we denote by F ′ : K ⊂ V → L(V ,W )the Gâteaux derivative of F on K .
Property 85
If a mapping F is Fréchet-dierentiable, it is also Gâteaux dierentiable andthe derivatives F ′ coincide.Conversely, if a mapping F is Gâteaux-dierentiable at u and if F is continuousat u or if the limit in (2) is uniform with v such that ‖v‖ = 1, then F is alsoFréchet-dierentiable and the two derivatives coincide.
Fixed point Monotone operators Dierential calculus Newton method
Convex functions I
Denition 86
A subset K of a vector space V is said convex if
∀u, v ∈ K , ∀t ∈ [0, 1], tu + (1− t)v ∈ K
Denition 87
A function J : K → R, dened on a convex set K of V , is said
convex if for all u, v ∈ K
J(tu + (1− t)v) ≤ tJ(u) + (1− t)J(v) ∀t ∈ [0, 1]
strictly convex if for all u, v ∈ K with u 6= v ,
J(tu + (1− t)v) < tJ(u) + (1− t)J(v) ∀t ∈ (0, 1)
Fixed point Monotone operators Dierential calculus Newton method
Convex functions II
Theorem 88
Let J : K ⊂ V → R be Gateaux-dierentiable. The following statements areequivalent:
(1) J is convex
(2) J(v) ≥ J(u) + (J ′(u), v − u), for all u, v ∈ K
(3) J ′ is monotone, i.e. (J ′(v)− J ′(u), v − u) ≥ 0 , for all u, v ∈ K
Theorem 89
Let J : K ⊂ V → R be Gateaux-dierentiable. The following statements areequivalent:
(1) J is strictly convex
(2) J(v) > J(u) + (J ′(u), v − u), for all u, v ∈ K with u 6= v
(3) J ′ is strictly monotone, i.e. (J ′(v)− J ′(u), v − u) > 0 , for all u, v ∈ Kwith u 6= v
Fixed point Monotone operators Dierential calculus Newton method
Convex functions III
Denition 90
A function J : K ⊂ V → R is said strongly convex if it is Gateaux-dierentiableand if its Gâteaux derivative is strongly monotone, i.e. if there exists a constantα > 0 such that
(J ′(v)− J ′(u), v − u) ≥ α‖u − v‖2
Fixed point Monotone operators Dierential calculus Newton method
Convex optimization I
Theorem 91
Let K be a closed convex subset of an Hilbert space V . Assume thatJ : K → R be a convex and Gâteaux dierentiable mapping. Then, there existsu ∈ K such that
J(u) = infv∈K
J(v) (3)
if and only if there exists u ∈ K such that
(J ′(u), v − u) ≥ 0 ∀v ∈ K (4)
When K is a linear subspace, the last inequality reduces to
(J ′(u), v) = 0 ∀v ∈ K (5)
Fixed point Monotone operators Dierential calculus Newton method
Convex optimization II
Proof.
Assume (3). Then ∀v ∈ K and ∀t ∈ [0, 1],
J(u) ≤ J(tv + (1− t)u) ≤ tJ(v) + (1− t)J(u)
and thenJ(u + t(v − u))− J(u)
t≥ J(v)− J(u) ∀t ∈ (0, 1]
Taking the limit t → 0+, we obtain
(J′(u), v − u) ≥ J(v)− J(u) ≥ 0
Now, assume (4). Since J is convex, we have ∀v ∈ K
J(v) ≥ J(u) + (J′(u), v − u) ≥ 0
Finally, if K is a subspace, then for all v ∈ K , u ± v ∈ K and therefore
(J′(u),±v) ≥ 0 ⇒ (J′(u), v) = 0 ∀v ∈ K
Fixed point Monotone operators Dierential calculus Newton method
Part V
Nonlinear equations
14 Fixed point theorem
15 Nonlinear equations with monotone operators
16 Dierential calculus for nonlinear operators
17 Newton method
Fixed point Monotone operators Dierential calculus Newton method
Newton method I
Let U and V be two Banach spaces and F : U → V a Fréchet-dierentiablefunction. We want to solve
F (u) = 0
The Newton method consists in constructing a sequence unn∈N by solvingsuccessive linearized problems. At iteration n, we introduce the linearization Fof F at un, dened by
F (v) = F (un) + F ′(un)(v − un)
and we dene un+1 such that F (un+1) = 0. The Newton iterations are thendened as follows.
Newton iterations
Start from an initial guess u0 and compute the sequence unn∈N dened by
un+1 = un − F ′(un)−1F (un)
Fixed point Monotone operators Dierential calculus Newton method
Newton method II
Theorem 92 (local convergence of Newton method)
Assume u∗ is solution of F (u∗) = 0 and assume that F ′(u∗)−1 exists and is acontinuous linear map from V to U. Assume that F ′ is locally Lipschitzcontinuous at u∗, i.e.
‖F ′(u)− F ′(v)‖ ≤ L‖u − v‖ ∀u, v ∈ N(u∗)
where N(u∗) is a neighborhood of u∗. Then, there exists δ > 0 such that if‖u0 − u∗‖ ≤ δ, the sequence unn≥1 of the Newton method is well-denedand converges to u∗. Moreover, there exists a constant M < 1/δ such that
‖un+1 − u∗‖ ≤ ‖un − u∗‖2
and‖un − u∗‖ ≤ (Mδ)2
n
/M
Proof.
See [Atkinson & Han (2009, section 5.4)]
Fixed point Monotone operators Dierential calculus Newton method
Newton method for nonlinear systems of equations I
Let F : Rm → Rm and consider the nonlinear system of equations
F (u) = 0
The iterations of the Newton method are dened by
un+1 = un − F ′(un)−1F (un)
where F ′(un) ∈ Rm×m is called the tangent matrix at un.In algebraic notations, F (u) and F ′(u) can be expressed as follows:
u =
a1...am
, F (u) =
F1(a1, . . . , am)...
Fn(a1, . . . , am)
, F ′(u) =
∂F1∂a1
(u) . . . ∂F1∂am
(u)...
...∂Fm∂a1
(u) . . . ∂Fm∂am
(u)
Fixed point Monotone operators Dierential calculus Newton method
Modied Newton method
One iteration of the (full) Newton method can be written as a linear system ofequations
Anδn = −F (un), δn = un+1 − un
where An = F ′(un). In order to avoid the computation of the tangent matrixF ′(un) at each iteration, we can use modied Newton iterations where An isonly an approximation of F ′(un). For example, we could update An when theconvergence is too slow or after every k iterations:
An = F ′(um) for n = mk + j , j ∈ 0, . . . , k − 1
Remark.
The convergence of the modied Newton method is usually slower that (full)Newton method but more iterations can be performed for the samecomputation time.
Interpolation Best approximation Orthogonal polynomials
Part VI
Interpolation / Approximation
18 InterpolationLagrange interpolationHermite interpolationTrigonometric interpolation
19 Best approximationElements on topological vector spacesGeneral existence resultsExistence and uniqueness of best approximationBest approximation in Hilbert spaces
20 Orthogonal polynomialsWeighted L2 spacesClassical orthogonal polynomials
Interpolation Best approximation Orthogonal polynomials
Introduction
Principle of approximation
The aim is to replace a function f , known exactly or approximately, by anapproximating function p which is more convenient for numerical computation.
The most commonly used approximating functions p are polynomials, piecewisepolynomials or trigonometric polynomials.There are several ways of dening the approximating function among a givenclass of functions: interpolation, projection, ...
Interpolation Best approximation Orthogonal polynomials Lagrange interpolation Hermite interpolation Trigonometric interpolation
Part VI
Interpolation / Approximation
18 InterpolationLagrange interpolationHermite interpolationTrigonometric interpolation
19 Best approximationElements on topological vector spacesGeneral existence resultsExistence and uniqueness of best approximationBest approximation in Hilbert spaces
20 Orthogonal polynomialsWeighted L2 spacesClassical orthogonal polynomials
Interpolation Best approximation Orthogonal polynomials Lagrange interpolation Hermite interpolation Trigonometric interpolation
Part VI
Interpolation / Approximation
18 InterpolationLagrange interpolationHermite interpolationTrigonometric interpolation
19 Best approximationElements on topological vector spacesGeneral existence resultsExistence and uniqueness of best approximationBest approximation in Hilbert spaces
20 Orthogonal polynomialsWeighted L2 spacesClassical orthogonal polynomials
Interpolation Best approximation Orthogonal polynomials Lagrange interpolation Hermite interpolation Trigonometric interpolation
Preliminary denitions
We denote by Pn(I ) the space of polynomials of degre n dened on the closedinterval I ⊂ R:
Pn(I ) = v : I → R; v(x) =n∑i=0
vixi , vi ∈ R
We denote by C(I ) the space of continuous functions f : I → R. C(I ) is aBanach space when equipped with the norm
‖f ‖C(I ) = supx∈I|f (x)|
We denote by f (i) the i-th derivative of f . We denote by Cm(I ) the space of mtimes dierentiable functions f such that all its derivatives f (i) of order i ≤ mare continuous. Cm(I ) is a Banach space when equipped with the norm
‖f ‖Cm(I ) = maxi≤m‖f (i)‖C(I )
Interpolation Best approximation Orthogonal polynomials Lagrange interpolation Hermite interpolation Trigonometric interpolation
Lagrange interpolation
Let f ∈ C([a, b]) be a continuous function dened on the interval [a, b]. Weintroduce a set of n + 1 distinct points xini=0 on [a, b], such that
a ≤ x0 < . . . < xn ≤ b
The Lagrange interpolation pn ∈ Pn of f is the unique polynomial of degree nsuch that
pn(xi ) = f (xi ) for all i ∈ 0, . . . , n
We can represent pn as follows:
pn(x) =n∑i=0
f (xi )`i (x), `i (x) =n∏j=0
j 6=i
x − xjxi − xj
where the `ini=0 form a basis of Pn, called the Lagrange interpolation basis. Itis the unique basis of functions satisfying the interpolation conditions
`i (xj ) = δij ∀i , j ∈ 0, . . . , n
Interpolation Best approximation Orthogonal polynomials Lagrange interpolation Hermite interpolation Trigonometric interpolation
Lagrange interpolation
Theorem 93
Assume f ∈ Cn+1([a, b]). Then, for x ∈ [a, b], there exists ξx ∈ [a, b] such that
f (x)− pn(x) =ωn(x)
(n + 1)!f (n+1)(ξx), ωn(x) =
n∏i=0
(x − xi )
Inuence of the interpolation grid: Function wn(x) on [−1, 1]
Gauss-Legendre grid (blue), Uniform grid (red), Random grid (black)
−1 −0.5 0 0.5 1−0.08
−0.06
−0.04
−0.02
0
0.02
0.04
0.06
0.08
n = 5 −1 −0.5 0 0.5 1−0.5
0
0.5
1
1.5
2
Interpolation Best approximation Orthogonal polynomials Lagrange interpolation Hermite interpolation Trigonometric interpolation
Lagrange interpolation
Theorem 93
Assume f ∈ Cn+1([a, b]). Then, for x ∈ [a, b], there exists ξx ∈ [a, b] such that
f (x)− pn(x) =ωn(x)
(n + 1)!f (n+1)(ξx), ωn(x) =
n∏i=0
(x − xi )
Inuence of the interpolation grid: Function wn(x) on [−1, 1]
Gauss-Legendre grid (blue), Uniform grid (red), Random grid (black)
−1 −0.5 0 0.5 1−6
−5
−4
−3
−2
−1
0
1
2x 10
−3
n = 11 −1 −0.5 0 0.5 1−0.5
0
0.5
1
1.5
2
Interpolation Best approximation Orthogonal polynomials Lagrange interpolation Hermite interpolation Trigonometric interpolation
Lagrange interpolation: a famous example...
Runge function f (x) = 1
1+x2on [−5, 5]
Uniform grid: n = 5, 11, 19
−5 0 5−0.2
0
0.2
0.4
0.6
0.8
1
1.2
−5 0 5−0.2
0
0.2
0.4
0.6
0.8
1
1.2
−5 0 5−1
0
1
2
3
4
5
6
7
8
9
Gauss-Legendre grid: n = 5, 11, 19
−5 0 50
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
−5 0 50
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
−5 0 50
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Interpolation Best approximation Orthogonal polynomials Lagrange interpolation Hermite interpolation Trigonometric interpolation
Part VI
Interpolation / Approximation
18 InterpolationLagrange interpolationHermite interpolationTrigonometric interpolation
19 Best approximationElements on topological vector spacesGeneral existence resultsExistence and uniqueness of best approximationBest approximation in Hilbert spaces
20 Orthogonal polynomialsWeighted L2 spacesClassical orthogonal polynomials
Interpolation Best approximation Orthogonal polynomials Lagrange interpolation Hermite interpolation Trigonometric interpolation
Hermite polynomial interpolationFirst order interpolation
First order Hermite polynomial interpolation consists in interpolating a functionf (x) and its derivative f ′(x).Assume f ∈ C 1([a, b]). We introduce a set of n + 1 distinct points xini=0 on[a, b], with
a ≤ x0 < . . . < xn ≤ b
The hermite interpolant p2n+1 ∈ P2n+1 of f is uniquely dened by the followinginterpolation conditions:
p2n+1(xi ) = f (xi ), p′2n+1(xi ) = f ′(xi ), 0 ≤ i ≤ n
Interpolation Best approximation Orthogonal polynomials Lagrange interpolation Hermite interpolation Trigonometric interpolation
General Hermite polynomial interpolationHigher order interpolation
Hermite interpolation can be generalized for the interpolation of higher orderderivatives. At a given point xi , it interpolates the function and its derivativesup to the order mi ∈ N. Let N =
∑n
i=0(mi + 1)− 1. A generalized Hermite
interpolant pN ∈ PN is uniquely dened by the following conditions
p(j)N (xi ) = f (j)(xi ), 0 ≤ j ≤ mi , 0 ≤ i ≤ n
Theorem 94
Assume f ∈ CN+1([a, b]). Then, for x ∈ [a, b], there exists ξx ∈ [a, b] such that
f (x)− pN(x) =ωN(x)
(N + 1)!f (N+1)(ξx), ωN(x) =
n∏i=0
(x − xi )mi
Interpolation Best approximation Orthogonal polynomials Lagrange interpolation Hermite interpolation Trigonometric interpolation
Part VI
Interpolation / Approximation
18 InterpolationLagrange interpolationHermite interpolationTrigonometric interpolation
19 Best approximationElements on topological vector spacesGeneral existence resultsExistence and uniqueness of best approximationBest approximation in Hilbert spaces
20 Orthogonal polynomialsWeighted L2 spacesClassical orthogonal polynomials
Interpolation Best approximation Orthogonal polynomials Lagrange interpolation Hermite interpolation Trigonometric interpolation
Trigonometric polynomials
A trigonometric polynomial is dened as follows
pn(x) = a0 +n∑j=1
(aj cos(jx) + bj sin(jx)) , x ∈ [0, 2π)
pn is said of degree n if |an|+ |bn| 6= 0. An equivalent notation is as follows:
pn(x) =n∑
j=−n
cjeijx ,
witha0 = c0, aj = cj + c−j , bj = i(cj − c−j )
or equivalently (under a polynomial-like form)
pn(x) =n∑
j=−n
cjzj = z−n
2n∑k=0
ck−nzk , z = e ix
Interpolation Best approximation Orthogonal polynomials Lagrange interpolation Hermite interpolation Trigonometric interpolation
Trigonometric interpolation
We introduce 2n + 1 distinct interpolation points xj2nj=0 in [0, 2π). Classically,we use uniformly distributed points
xj = j2π
2n + 1, 0 ≤ j ≤ 2n
The trigonometric interpolant of degree n of function f is dened by thefollowing conditions
pn(xj ) = f (xj ), 0 ≤ j ≤ 2n
It can be equivalently reformulated as an interpolation problem in the complexplane: nd cknk=−n such that
2n∑k=0
ck−nzkj = znj f (xj ), 0 ≤ j ≤ 2n
where we have introduce complex points zj = e ixj .
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Part VI
Interpolation / Approximation
18 InterpolationLagrange interpolationHermite interpolationTrigonometric interpolation
19 Best approximationElements on topological vector spacesGeneral existence resultsExistence and uniqueness of best approximationBest approximation in Hilbert spaces
20 Orthogonal polynomialsWeighted L2 spacesClassical orthogonal polynomials
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
The problem of the best approximation
The aim is to nd the best approximation p of a function f in a set offunctions K (e.g. a polynomial space, piecewise polynomial space, ...)
minp∈K‖f − p‖
The obtained best approximation p depends on the norm selected formeasuring the error (e.g. L2-norm, L∞-norm, ...).
We will rst introduce some general results about optimization problems
infv∈K
J(v)
by giving some general conditions on the set K and the function J for theexistence of a minimizer.
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
An rst comprehensive case: extrema of real-valued functions I
Consider a real-valued continuous function J ∈ C([a, b]). The problem is tond a minimizer of J
infv∈[a,b]
J(v)
The classical result of Weierstrass states that a continuous function on a closedinterval K = [a, b] has a minimum in K (and a maximum). We recall the mainsteps of a typical proof in order to obtain more general requirements on K andJ.
1 We denote byα = inf
v∈KJ(v)
By denition of the inmum, there exists a sequence vn ⊂ K such thatlimn→∞ J(vn) = α.
2 K is a closed and bounded interval in R, and therefore it is a compact set.Therefore, from the sequence vn ⊂ K , we can extract a subsequencevnk which converges to some v∗ ∈ K ,
vnk −→k→∞v∗
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
An rst comprehensive case: extrema of real-valued functions II
3 Using the continuity of J, we obtain
J(v∗) = limk→∞
J(vnk ) = α
which proves that v∗ is a minimizer of J in K .
Now we come back on the dierent points of the proof in order to generalizethe existence result for functionals J dened on a subset K of a Banach spaceV .
1 The existence of a minimizing sequence vn ⊂ K is the denition of theinmum.
2 In an innite-dimensional Banach space V , a bounded sequence does notnecessarily admits a converging subsequence. However, for a reexiveBanach space V , there exists a weakly convergent subsequence. We thensuppose that V is a reexive Banach space and K ⊂ V is a bounded set.In order for K to contain the limit of this subsequence, K has to be weaklyclosed.
3 Finally, we want the weak limit of the subsequence to be a minimizer of J.We could then impose to J to be continuous with respect to a weak limit.However, this condition is too restrictive and it is sucient to impose thatJ is weakly lower semi continuous (allowing discontinuities).
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Part VI
Interpolation / Approximation
18 InterpolationLagrange interpolationHermite interpolationTrigonometric interpolation
19 Best approximationElements on topological vector spacesGeneral existence resultsExistence and uniqueness of best approximationBest approximation in Hilbert spaces
20 Orthogonal polynomialsWeighted L2 spacesClassical orthogonal polynomials
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Elements on topological vector spaces I
In the following, V denotes a normed space, i.e. a vector space equipped witha norm ‖ · ‖.
Denition 95 (Strong convergence on V )
A sequence vn ⊂ V is said to converge strongly to v ∈ V if
limn→∞
‖vn − v‖ = 0
It is denotedvn → v
Denition 96 (Cauchy sequence)
A sequence vn ⊂ V is Cauchy if
limn→∞
supi,j≥n‖vi − vj‖ = 0
or equivalently, if ∀ε > 0, there exists n ∈ N such that for all i , j ≥ n,‖vi − vj‖ ≤ ε.
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Elements on topological vector spaces II
Denition 97 (Closed set)
A subset K ⊂ V is said to be closed if it contains all the limits of itsconvergent sequences:
vn ⊂ K and vn → v ⇒ v ∈ K
The closure K of a set K is the union of this set and of the limits of allconverging sequences in K .
Denition 98 (Compact set)
A subset K of a normed space V is said to be (sequentially) compact if everysequence vnn∈N contains a subsequence vnk k∈N converging to an elementin K .A set K whose closure K is compact is said relatively compact.
Denition 99 (Banach space)
A Banach space is a complete normed vector space, i.e. a normed vector spacesuch that every Cauchy sequence in V has a limit in V .
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Elements on topological vector spaces III
Denition 100 (Dual of a normed space V )
The dual space of a normed space V is set space V ′ = L(V ,R) of linearcontinuous maps from V to R. V ′ is a Banach space for the norm
‖L‖ = supv∈V :‖v‖≤1
|L(v)| = supv∈V
|L(v)|‖v‖ , L ∈ V ′
Denition 101 (Reexive normed space)
A normed space V is said reexive if V ′′ = V , where V ′′ = (V ′)′ is the dual ofthe dual of V , also called bidual of V .
Denition 102 (Strong convergence on V ′)
A sequence Ln ⊂ V ′ is said to converge strongly to L ∈ V ′ if
limn→∞
‖Ln − L‖ = 0
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Elements on topological vector spaces IV
The dual space can be used to dene a new topology on V , called the weaktopology. The notions of convergence, closure, continuity... can be redenedwith respect to this new topology.
Denition 103 (Weak convergence on V )
A sequence vn ⊂ V is said to converge weakly to v ∈ V if
limn→∞
L(v − vn) = 0 ∀L ∈ V ′
It is denotedvn v
Denition 104 (Weakly closed set in V )
A subset K ⊂ V is said to be weakly closed if it contains all the limits of itsweakly convergent sequences:
vn ⊂ K and vn v ⇒ v ∈ K
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Elements on topological vector spaces V
Denition 105 (Weakly compact set)
A subset K of a normed space V is said to be weakly compact if everysequence vnn∈N contains a subsequence vnk k∈N weakly converging to anelement in K .A set K whose closure in the weak topology is weakly compact is said weaklyrelatively compact.
Theorem 106 (Reexive Banach spaces and converging bounded sequences)
A Banach space V is reexive if and only if every bounded sequence in V has asubsequence weakly converging to an element in V .
Let us note that the above theorem could be reformulated as follows: a Banachspace is reexive if and only if the unit ball is relatively compact in the weaktopology.
Theorem 107
A set K in V is bounded and weakly closed if and only if it is weakly compact.
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Lower semicontinuity I
Denition 108 (Lower semicontinuity)
A function J : V → R is lower semicontinuous (l.s.c.) if
vn ⊂ K and vn → v ∈ K ⇒ J(v) ≤ lim infn→∞
J(vn)
Denition 109 (Weak lower semicontinuity)
A function J : V → R is weakly lower semicontinuous (w.l.s.c.) if
vn ⊂ K and vn v ∈ K ⇒ J(v) ≤ lim infn→∞
J(vn)
Proposition 110
Continuity implies lower semicontinuity (but the converse statement is nottrue)
Weak lower semicontinuity implies lower semicontinuity (but the conversestatement is not true)
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Lower semicontinuity II
Example 111
Let us prove that the norm function ‖.‖ : v ∈ V 7→ ‖v‖ ∈ R in a normed spaceV is w.l.s.c.Let vn ⊂ V be a weakly convergent sequence with vn v . There exists alinear form L ∈ V ′ such that L(v) = ‖v‖ and ‖L‖ = 1 (Corollary of theGeneralized Hahn-Banach theorem). We then have
L(vn) ≤ ‖L‖‖vn‖ = ‖vn‖
and therefore‖v‖ = L(v) = lim
n→∞L(vn) ≤ lim inf
n→∞‖vn‖
If V is an inner product space, we have a simpler proof. Indeed,
‖v‖2 = (v , v) = limn→∞
(vn, v) ≤ lim infn→∞
‖v‖‖vn‖
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Part VI
Interpolation / Approximation
18 InterpolationLagrange interpolationHermite interpolationTrigonometric interpolation
19 Best approximationElements on topological vector spacesGeneral existence resultsExistence and uniqueness of best approximationBest approximation in Hilbert spaces
20 Orthogonal polynomialsWeighted L2 spacesClassical orthogonal polynomials
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
General existence results I
We introduce the problem
infv∈K
J(v) (π)
Theorem 112
Assume V is a reexive Banach space. Let K ⊂ V denote a bounded andweakly closed set. Let J : V → R denote a weakly l.s.c. function. Then,problem (π) has a solution in K.
Proof.
Denote α = infv∈K J(v) and vn ⊂ K a minimizing sequence such thatlimn→∞ J(vn) = α. Since K is bounded, vn is a bounded sequence in areexive Banach space and therefore, we can extract a subsequence vnk weakly converging to some u ∈ V . Since K is weakly closed, u ∈ K . Since J isw.l.s.c.
J(u) ≤ lim infk→∞
J(vnk ) = α
and therefore, u ∈ K is a minimizer of J.
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
General existence results II
We now remove the boundedness of the set K by adding a coercivity conditionon J.
Denition 113
A functional J : V → R is said coercive if
J(v)→ +∞ as ‖v‖ → ∞
Theorem 114
Assume V is a reexive Banach space. Let K ⊂ V denote a weakly closed set.Let J : V → R denote a weakly l.s.c. and coercive function. Then, the problem(π) has a solution in K.
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
General existence results III
Proof.
Pick an element v0 ∈ K with J(v0) <∞ and let K0 = v ∈ K ; J(v) ≤ J(v0).Since J is coercive, K0 is bounded. Moreover, K0 is weakly closed. Indeed, ifvn ⊂ K0 is such that vn v∗, then v∗ ∈ K (since K is weakly closed) andJ(v∗) ≤ lim infn J(vn) ≤ J(v0), and therefore v∗ ∈ K0. The optimizationproblem is then equivalent to the optimization problem
infv∈K0
J(v)
of a w.l.s.c. function on a bounded and weakly closed set. Theorem 112 allowsto conclude on the existence of a minimizer.
Lemma 115 (Convex closed sets are weakly closed)
A convex and closed set K ⊂ V is weakly closed.
Lemma 116 (Convex l.s.c. functions are w.l.s.c.)
A convex and l.s.c. function is also w.l.s.c.
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
General existence results IV
For convex sets and convex functions, theorems 112 and 114 can then bereplaced by the following theorem.
Theorem 117
Assume V is a reexive Banach space. Let K ⊂ V denote a convex and closedset. Let J : V → R denote a convex l.s.c. function. Then, if either (i) K isbounded, or (ii) J is coercive on K, then the minimization problem (π) has asolution in K. Moreover, if J is strictly convex, this solution is unique.
Proof.
The existence simply follows from theorems 112 and 114 and from properties115 and 116. It remains to prove the uniqueness if J is strictly convex. Assumethat u1, u2 ∈ K are two solutions such that u1 6= u2. We haveJ(u1) = J(u2) = minv∈K J(v). Since K is convex, αu1 + (1− α)u2 ∈ K forα ∈ (0, 1), and by strict convexity of J, we have
J(αu1 + (1− α)u2) < αJ(u1) + (1− α)J(u2) = minv∈K
J(v)
which contradicts the fact that u1 and u2 are solutions.
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
General existence results V
In the case of a non reexive Banach space V (e.g. V = C([a, b])) the abovetheorems do not apply. However, the reexivity is used for the extraction of aweakly convergent subsequence from a bounded sequence in K . In fact, we justneed the completeness of the set K and not of the space V . In particular, fornite-dimensional subset K , we have.
Theorem 118
Assume V is a normed space. Let K ⊂ V denote a nite-dimensional convexand closed set. Let J : V → R denote a convex l.s.c. function. Then, if either(i) K is bounded, or (ii) J is coercive on K, then the minimization problem (π)has a solution in K. Moreover, if J is strictly convex, this solution is unique.
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Part VI
Interpolation / Approximation
18 InterpolationLagrange interpolationHermite interpolationTrigonometric interpolation
19 Best approximationElements on topological vector spacesGeneral existence resultsExistence and uniqueness of best approximationBest approximation in Hilbert spaces
20 Orthogonal polynomialsWeighted L2 spacesClassical orthogonal polynomials
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Existence and uniqueness of best approximation I
We apply the general results about optimization on the following bestapproximation problem. For a given element u ∈ V , where V is a normedspace, we want to nd the elements in a subset K ⊂ V which are the closest tou. The problem writes
infv∈K‖u − v‖
Denoting J(v) = ‖u − v‖, the problem can then be written under the forminfv∈K J(v).
Property 119
Function J(v) = ‖u − v‖ is convex, continuous (and hence w.l.s.c), andcoercive.
We then have the two existence results.
Theorem 120
Let V be a reexive Banach space and K ⊂ V a closed convex subset. Thenthere exists a best approximation u ∈ K verifying
‖u − u‖ = minv∈K‖u − v‖
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Existence and uniqueness of best approximation II
Theorem 121
Let V be a normed space and K ⊂ V a nite-dimensional closed convexsubset. Then there exists a best approximation u ∈ K verifying
‖u − u‖ = minv∈K‖u − v‖
For the uniqueness of the best approximation, we have to look at the propertiesof the norm.
Theorem 122
I there exists a p > 1 such that v 7→ ‖v‖p is strictly convex, then a solution uof the best approximation problem is unique.
Example 123
If V is a Hilbert space equipped with the inner product (·, ·) andassociated norm ‖ · ‖, v 7→ ‖v‖2 is a strictly convex function.
If V = Lp(Ω) with p ∈ (1,+∞), v 7→ ‖v‖pLp(Ω) is strictly convex.
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Part VI
Interpolation / Approximation
18 InterpolationLagrange interpolationHermite interpolationTrigonometric interpolation
19 Best approximationElements on topological vector spacesGeneral existence resultsExistence and uniqueness of best approximationBest approximation in Hilbert spaces
20 Orthogonal polynomialsWeighted L2 spacesClassical orthogonal polynomials
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Best approximation in Hilbert spaces I
Let V be a Hilbert space equipped with inner product (·, ·) and associatednorm ‖ · ‖.
Lemma 124
Let K be a closed convex set in Hilbert space V . u ∈ K is a bestapproximation of u ∈ V if and only if
(u − u, v − u) ≤ 0 ∀v ∈ K
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Best approximation in Hilbert spaces II
Proof.
First suppose that u ∈ K is a best approximation of u ∈ V . Then,
‖u − u‖2 ≤ ‖w − u‖2
∀w ∈ K . By selecting w = u + α(v − u), with α ∈ (0, 1) and v ∈ K , we have
0 ≥ ‖u − u‖2 − ‖(u − u) + α(v − u)‖2
= −α2(v − u, v − u)− 2α(u − u, v − u)
for all α ∈ (0, 1). That implies (u − u, v − u) ≥ 0 ∀v ∈ K .Conversely, if (u − u, v − u) ≥ 0, ∀v ∈ K , then
‖v − u‖2 = ‖(v − u) + (u − u)‖2
= ‖v − u‖2 + 2(v − u, u − u) + ‖u − u‖2
≥ ‖u − u‖2
for all v ∈ K .
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Best approximation in Hilbert spaces III
Corollary 125
Let K be a closed convex set in Hilbert space V . For any u ∈ V , the bestapproximation in K is unique.
Proof.
Let u1, u2 ∈ K be two best approximations of u ∈ V . Then,(u − u1, u2 − u1) ≤ 0 and (u − u2, u1 − u2) ≤ 0. Additionning theseinequalities, we obtain
(u2 − u1, u2 − u1) = ‖u2 − u1‖2 ≤ 0
and therefore u1 = u2.
We then conclude with the following theorem.
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Best approximation in Hilbert spaces IV
Theorem 126
Let K ⊂ V be a nonempty closed convex set in Hilbert space V . For anyu ∈ V , there exists a unique best approximation u ∈ K dened by
‖u − u‖ = minv∈K‖u − v‖
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Best approximation in Hilbert spaces V
Remark.
Let us give another classical proof for the existence of a best approximation,which uses the inner product structure of the space V . Let unn∈N ⊂ K be aminimizing sequence such that limn→∞ ‖u − un‖ = α = infv∈K ‖u − v‖. Usingthe parallelogram law satised by the norm ‖.‖ in an inner product space, wehave
2‖u − un‖2 + 2‖u − um‖2 = ‖un − um‖2 + ‖2u − un − um‖2
Since K is convex, we have (un + um)/2 ∈ K and therefore
‖un − um‖2 = 2‖u − un‖2 + 2‖u − um‖2 − 4‖u − (un + um)/2‖2
≤ 2‖u − un‖2 + 2‖u − um‖2 − 4α2 −→m,n→∞
0
which proves that un is a Cauchy sequence. Since V is complete, un ⊂ Kconverges to an element u ∈ V and since K is closed, u ∈ K .
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Best approximation in Hilbert spaces: Projection I
Denition 127 (Projector on a convex set)
The best approximation u ∈ K of u ∈ V in a closed convex set K is called theprojection of u onto K and is denoted
u = PK (u)
where PK : V → K is called the projection operator of V onto K .
Proposition 128
The projection operator is monotone
(PK (v)− PK (u), v − u) ≥ 0 ∀u, v ∈ V
and non expansive
‖PK (v)− PK (u)‖ ≤ ‖v − u‖ ∀u, v ∈ V
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Best approximation in Hilbert spaces: Projection II
Proof.
From the characterizations of PK (u) ∈ K and PK (v) ∈ K , we have respectively
(PK (u)− u,PK (v)− PK (u)) ≥ 0, (PK (v)− v ,PK (u)− PK (v)) ≥ 0
Adding these inequalities, we obtain
(v − u,PK (v)− PK (u)) ≥ (PK (v)− PK (u),PK (v)− PK (u)) ≥ 0
and
‖PK (v)− PK (u)‖2 ≤ (v − u,PK (v)− PK (u)) ≤ ‖v − u‖‖PK (v)− PK (u)‖
We now introduce the following particular case when K is a subspace of V .
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Best approximation in Hilbert spaces: Projection III
Theorem 129 (Projection on linear subspaces)
Let K be a complete subspace of V . Then, for any u ∈ V , there exists aunique best approximation u = PK (u) ∈ K characterized by
(u − PK (u), v) = 0 ∀v ∈ K
Proof.
We have(u − u,w − u) ≤ 0 ∀w ∈ K
and since K is a subspace, for all v ∈ K , w = u ± v ∈ K , and therefore
±(u − u, v) ≤ 0 ∀v ∈ K
In the case where K is a subspace, u − PK (u) is orthogonal to K , andtherefore, PK is called an orthogonal projection operator.
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Best approximation in Hilbert spaces: Projection IV
Let us consider that we know an orthonormal basis ϕini=1 of K = Kn. Theprojection PKn (u) is characterized by
PKn (u) =n∑i=1
(ϕi , u)ϕi
Example 130 (Least square approximation by polynomials)
Let V = L2(−1, 1) and Kn = Pn(−1, 1) the space of polynomials of degree lessthan n. An orthonormal basis of Kn is given by the Legendre polynomialsLini=0 dened by
Li (x) =√
(2i + 1)/21
2i i !
d i
dx i
((x2 − 1)i
)
Interpolation Best approximation Orthogonal polynomials Elements on topological vector spaces General existence results Existence and uniqueness of best approximation Best approximation in Hilbert spaces
Best approximation in Hilbert spaces: Projection V
Example 131 (Least square approximation by trigonometric polynomials)
Let V = L2(0, 2π) and Kn the space of trigonometric polynomials of degreeless than n. The best approximation un = PKn (u) is characterized by
un(x) = a0/2 +n∑j=1
(aj cos(jx) + bj sin(jx))
with
aj =1
(cos(jx), cos(jx))(u(x), cos(jx)) =
1
π
∫2π
0
u(x) cos(jx)dx , j ≥ 0
bj =1
(sin(jx), sin(jx))(u(x), sin(jx)) =
1
π
∫2π
0
u(x) sin(jx)dx , j ≥ 1
Note that un tends to the well-known Fourier series expansion of u.
Interpolation Best approximation Orthogonal polynomials Weighted L2 spaces Classical orthogonal polynomials
Part VI
Interpolation / Approximation
18 InterpolationLagrange interpolationHermite interpolationTrigonometric interpolation
19 Best approximationElements on topological vector spacesGeneral existence resultsExistence and uniqueness of best approximationBest approximation in Hilbert spaces
20 Orthogonal polynomialsWeighted L2 spacesClassical orthogonal polynomials
Interpolation Best approximation Orthogonal polynomials Weighted L2 spaces Classical orthogonal polynomials
Part VI
Interpolation / Approximation
18 InterpolationLagrange interpolationHermite interpolationTrigonometric interpolation
19 Best approximationElements on topological vector spacesGeneral existence resultsExistence and uniqueness of best approximationBest approximation in Hilbert spaces
20 Orthogonal polynomialsWeighted L2 spacesClassical orthogonal polynomials
Interpolation Best approximation Orthogonal polynomials Weighted L2 spaces Classical orthogonal polynomials
Weighted L2 spaces
Let I ⊂ R and ω : I → R be a weight function which is integrable on I andalmost everywhere positive. We introduce the weighted function space
L2ω(I ) = v : I → R; v is measurable on I ,
∫I
|v(x)|2ω(x)dx < +∞
L2ω(I ) is a Hilbert space for the inner product
(u, v) =
∫I
u(x)v(x)ω(x)dx
and associated norm
‖u‖ =
(∫I
u(x)2ω(x)dx
)1/2
Two functions u, v ∈ L2ω(I ) are said orthogonal if (u, v) = 0.
Interpolation Best approximation Orthogonal polynomials Weighted L2 spaces Classical orthogonal polynomials
Part VI
Interpolation / Approximation
18 InterpolationLagrange interpolationHermite interpolationTrigonometric interpolation
19 Best approximationElements on topological vector spacesGeneral existence resultsExistence and uniqueness of best approximationBest approximation in Hilbert spaces
20 Orthogonal polynomialsWeighted L2 spacesClassical orthogonal polynomials
Interpolation Best approximation Orthogonal polynomials Weighted L2 spaces Classical orthogonal polynomials
Classical orthogonal polynomials I
A system of orthonormal polynomials pnn≥0, with pn ∈ Pn(I ), can beconstructed by applying the Gram-Schmidt procedure to the basis ofmonomials 1, x , x2, . . .. For a given interval I and weight function ω, it leadsto a uniquely dened system of polynomials.In the following table, we indicate classical families of polynomials for dierentinterval domains I and weight functions.
Classical orthogonal polynomials
I ω(x) pn
(−1, 1) (1+x)a−1(1−x)b−1
2a+b−1B(a,b)Jacobi
(−1, 1) 1
2Legendre
(−1, 1) (1−x2)−1/2
B(1/2,1/2)Chebyshev of rst kind
(−1, 1) (1−x2)1/2
4B(3/2,3/2)Chebyshev of second kind
R 1√2πexp(− x2
2) Hermite
(0,+∞) 1
Γ(a)xaexp(−x) Laguerre
Interpolation Best approximation Orthogonal polynomials Weighted L2 spaces Classical orthogonal polynomials
Classical orthogonal polynomials II
Γ denotes the Euler Gamma function dened by
Γ(a) =
∫ ∞0
xa−1exp(−x)dx
B(a, b) denotes the Euler Beta function dened by
B(a, b) =Γ(a)Γ(b)
Γ(a + b)
Remark.
The given weight functions are such that∫I
ω(x)dx = 1
It then denes a measure µ with density ω (dµ(x) = ω(x)dx) and with unitarymass. Equivalently, µ (resp. ω) can be interpreted as the probability law (resp.probability density function) of a random variable.
Interpolation Best approximation Orthogonal polynomials Weighted L2 spaces Classical orthogonal polynomials
Classical orthogonal polynomials III
Exercice.
Construct by the Gram-Schmidt procedure the orthonormal polynomials ofdegree n = 0, 1, 2 on the interval I = (0, 1) and for the weight functionω(x) = log(1/x).
Basic quadrature formulas Gauss quadrature
Part VII
Numerical integration
21 Basic quadrature formulas
22 Gauss quadrature
Basic quadrature formulas Gauss quadrature
Numerical integration
Given a function f : Ω→ R, the aim is to approximate the value of the integral
I (f ) =
∫Ω
f (x)dx
using evaluations of the function
I (f ) ≈n∑
k=1
f (xk)ωk
or eventually of the function and its derivatives
I (f ) ≈n∑
k=1
f (xk)ωk +n∑
k=1
f ′(xk)ωk + . . .
These approximations are called quadrature formulas. A quadrature formula issaid of interpolation type if it uses only evaluations of the function.
Basic quadrature formulas Gauss quadrature
Integration error and precision
We denote by In(f ) the quadrature formula.
Denition 132
A quadrature formula have a degree of precision k if it integrates exactly allpolynomials of degree less or equal to k
In(f ) = I (f ) ∀f ∈ Pk(Ω)
In(f ) 6= I (f ) for some f ∈ Pk+1(Ω)
Basic quadrature formulas Gauss quadrature
Part VII
Numerical integration
21 Basic quadrature formulas
22 Gauss quadrature
Basic quadrature formulas Gauss quadrature
Basic quadrature formulas
Rectangle formula (precision degree 1)∫ b
a
f (x)dx ≈ (b − a)f (a + b
2)
Trapezoidal formula (precision degree 1)∫ b
a
f (x)dx ≈ (b − a)f (a) + f (b)
2
Simpson formula (precision degree 3)∫ b
a
f (x)dx ≈ (b − a)
6(f (a) + 4f (
a + b
2) + f (b))
...
Basic quadrature formulas Gauss quadrature
Composite quadrature formulas
In order to compute
I (f ; Ω) =
∫Ω
f (x)dx ,
we divide the domain Ω into m subdomains Ωmi=1 such that
I (f ; Ω) =m∑i=1
I (f ; Ωi )
and we introduce a basic quadrature formula on each subdomain
I (f ; Ω) ≈m∑i=1
In(f ; Ωi )
Basic quadrature formulas Gauss quadrature
Part VII
Numerical integration
21 Basic quadrature formulas
22 Gauss quadrature
Basic quadrature formulas Gauss quadrature
Gauss quadrature I
We want to approximate the weighted integral of a function f
Iw (f ) =
∫ b
a
f (x)w(x)dx
where w(x)dx denes a measure of integration. A Gauss quadrature formulawith n points is dened by
Iw (f ) ≈ Iwn (f ) =n∑i=1
ωi f (xi )
with points and weights such that it integrates exactly all polynomialsf ∈ P2n−1(a, b). The xi (resp. ωi ) are called Gauss points (resp. Gaussweights) associated with the present measure. We introduce the function spaceL2w (a, b) and its natural inner product
(f , g)w =
∫ b
a
f (x)g(x)w(x)dx
Basic quadrature formulas Gauss quadrature
Gauss quadrature II
Theorem 133
In(f ) = I (f ) for all f ∈ P2n−1(a, b) if and only if
the points xi are such that the polynomialzn(x) =
∏n
i=1(x − xi ) ∈ Pn(a, b) is orthogonal to Pn−1(a, b), i.e.
(zn(x), p(x))w = 0 ∀p ∈ Pn−1(a, b)
the weights are dened by ωi = I (Li ), where Li is the Lagrange interpolantat xi , dened by Li (x) =
∏n
j=1,j 6=i (x − xj )/(xi − xj )
Corollary 134
The n Gauss points of a n-points Gauss quadrature are the n roots of thedegree n orthogonal polynomial.
For (a, b) = (−1, 1) and w(x) = 1, the xi are the n roots of the degree nLegendre polynomial.
For (a, b) = (−∞,∞) and w(x) = exp(−x2), the xi are the n roots of thedegree n Hermite polynomial.
...