Jacobi eigenvalue algorithm

From Wikipedia, the free encyclopedia

The Jacobi eigenvalue algorithm is a numerical procedure for the calculation of all eigenvalues and eigenvectors of a real symmetric matrix.

Contents 1 Description 2 Convergence 3 Cost 4 Algorithm 4.1 Notes 4.2 Example 5 Applications for real symmetric matrices 6 Generalizations 7 References

[edit] Description

Let and let denote the matrix with components

describes a rotation with angle φ in the plane spanned by the k. and l. unit vector. We have , i.e. is orthogonal.

For a real symmetric matrix S let . This matrix differs from S only in rows and columns k and l where ( if for short) :

and

Define

(the sum of squares of all components),

(the sum of squares of all diagonal components),

(the sum of squares of all off-diagonal components).

| | S | | _F is the Frobenius norm of S and Γ (S ) is called off-diag norm since Γ(S) = 0 if and only if S is diagonal.

The Frobenius norm does not change under unitary rotations, so we have | | S' | | _F = | | S | | _F . If we choose φ according to

then S_kl' = 0 and one finds . Thus the off-norm decreases and hence S' becomes "more" diagonal than S. In order to optimize this effect, S_kl should be the largest off-diagonal component. Such an element is called a pivot and the rotated matrix S^J is called a Jacobi rotation.

The Jacobi eigenvalue method repeatedly performs Jacobi rotations until the matrix becomes almost diagonal. Then the elements in the diagonal are approximations of the (real) eigenvalues of S.

[edit] Convergence

If p = S_kl is a pivot element, then by definition for . Since S has exactly 2 N := n ( n - 1) off-diag elements, we have or . This implies or , i.e. the sequence of Jacobi rotations converges at least linearly by a factor (1 − 1 / N)^{1 / 2} to a diagonal matrix.

A number of N Jacobi rotations is called a sweep; let S^σ denote the result. The previous estimate yields

,

i.e. the sequence of sweeps converges at least linearly with a factor ≈ e^{1 / 2} .

However the following result of Schönhage yields locally quadratic convergence. To this end let S have m distinct eigenvalues λ₁,...,λ_m with multiplicities ν₁,...,ν_m and let d > 0 be the smallest distance of two different eigenvalues. Let us call a number of

Jacobi rotations a Schönhage-sweep. If S^s denotes the result then

.

Thus convergence becomes quadratic as soon as .

[edit] Cost

Each Jacobi rotation can be done in n steps when the pivot element p is known. However the search for p requires inspection of all N ≈ ½ n² off-diag elements. We can reduce this to n steps too if we introduce an additional index array with the property that m_i is the index of the largest element in row i, (i = 1, … , n - 1) of the current S. Then (k, l) must be one of the pairs (i,m_i) . Since only columns k and l change, only m_k and m_l must be updated, which again can be done in n steps. Thus each rotation has O(n) cost and one sweep has O(n³) cost which is equivalent to one matrix multiplication. Additionally the m_i must be initialized before the process starts, this can be done in n² steps.

Typically the Jacobi method converges within numerical precision after a small number of sweeps. Note that multiple eigenvalues reduce the number of iterations since N_S < N.

[edit] Algorithm

The following algorithm is a description of the Jacobi method in math-like notation. It calculates a vector e which contains the eigenvalues and a matrix E which contains the corresponding eigenvectors, i.e. e_i is an eigenvalue and the column E_i an orthonormal eigenvector for e_i , i = 1, … , n.

procedure jacobi(S ∈ Rn×n; out e ∈ Rn; out E ∈ Rn×n)
  var
    i, k, l, m, state ∈ N
    s, c, t, p, y ∈ R
    ind ∈ Nn
    changed ∈ Ln

  function maxind(k ∈ N) ∈ N ! index of largest element in row k
    m := k+1
    for i := k+2 to n do
      if │Sil│ > │Sim│ then m := l endif
    endfor
    return m
  endfunc

  procedure update(k ∈ N; t ∈ R) ! update ek and its status
    y := ek; ek := y+t
    if changedk and (y=ek) then changedk := false; state := state−1
    elsif (not changedk) and (y≠ek) then changedk := true; state := state+1
    endif
  endproc

  procedure rotate(k,l,i,j ∈ N) ! perform rotation of Sij, Skl
    ┌   ┐    ┌     ┐┌   ┐
    │Skl│    │c  −s││Skl│
    │   │ := │     ││   │
    │Sij│    │s   c││Sij│
    └   ┘    └     ┘└   ┘
  endproc

  ! init e, E, and arrays ind, changed
  E := I; state := n
  for k := 1 to n do indk := maxind(k); ek := Skk; changedk := true endfor
  while state≠0 do ! next rotation
    m := 1 ! find index (k,l) of pivot p
    for k := 2 to n−1 do
      if │Sk indk│ > │Sm indm│ then m := k endif
    endfor
    k := m; l := indm; p := Skl
    ! calculate c = cos φ, s = sin φ
    y := (el−ek)/2; t := │y│+√(p2+y2)
    s := √(p2+t2); c := t/s; s := p/s; t := p2/t
    if y<0 then s := −s; t := −t endif
    Skl := 0.0; update(k,−t); update(l,t)
    ! rotate rows and columns k and l
    for i := 1 to k−1 do rotate(i,k,i,l) endfor
    for i := k+1 to l−1 do rotate(k,i,i,l) endfor
    for i := l+1 to n do rotate(k,i,l,i) endfor
    ! rotate eigenvectors
    for i := 1 to n do
      ┌   ┐    ┌     ┐┌   ┐
      │Eki│    │c  −s││Eki│
      │   │ := │     ││   │
      │Eli│    │s   c││Eli│
      └   ┘    └     ┘└   ┘
    endfor
    ! rows k, l have changed, update rows indk, indl
    indk := maxind(k); indl := maxind(l)
  loop
endproc

[edit] Notes

1. The logical array changed holds the status of each eigenvalue. If the numerical value of e_k or e_l changes during an iteration, the corresponding component of changed is set to true, otherwise to false. The integer state counts the number of components of changed which have the value true. Iteration stops as soon as state = 0. This means that none of the approximations has recently changed its value and thus it is not very likely that this will happen if iteration continues. Here it is assumed that floating point operations are optimally rounded to the nearest flointing point number.

2. The upper triangle of the matrix S is destroyed while the lower triangle and the diagonal are unchanged. Thus it is possible to restore S if necessary according to

for k := 1 to n−1 do ! restore matrix S
  for l := k+1 to n do Skl := Slk endfor
endfor

3. The eigenvalues are not necessarily in descending order. This can be achieved by a simple sorting algorithm.

for k := 1 to n−1 do
  m := k
  for l := k+1 to n do
    if el > em then m := l endif
  endfor
  if k ≠ m then swap em,ek; swap Em,Ek endif
endfor

[edit] Example

Let

Then jacobi produces the following eigenvalues and eigenvectors after 3 sweeps (19 iterations) :

e₁ = 2585.25381092892231

e₂ = 37.1014913651276582

e₃ = 1.4780548447781369

e₄ = 0.1666428611718905

[edit] Applications for real symmetric matrices

When the eigenvalues (and eigenvectors) of a symmetric matrix are known, the following values are easily calculated.

Singular values

The singular values of a (square) matrix A are the square roots of the (non negative) eigenvalues of A^TA. In case of a symmetric matrix S we have of S^TS = S², hence the singular values of S are the absolute values of the eigenvalues of S

2-Norm and spectral radius

The 2-norm of a matrix A is the norm based on the euclidian vectornorm, i.e. the largest value | | Ax | | ₂ when x runs through all vectors with | | x | | ₂ = 1. It is the largest singular value of A. In case of a symmetric matrix it is largest absolute value of its eigenvectors and thus equal to its spectral radius.

Condition number

The condition number of a nonsingular matrix A is defined as cond(A) = | | A | | ₂ * | | A ^{− 1} | | ₂. In case of a symmetric matrix it is the absolute value of the quotient of the largest and smallest eigenvalue. Matrices with large condition numbers can cause numerically unstable results : small perturbation can result in large errors. Hilbert matrices are the most famous ill-conditioned matrices. For example the fourth order Hilbert matrix has a condition of 15514, while for order 8 it is 2.7 * 10⁸.

Rank

A matrix A has rank r if it has r columns which are linearily independent while the remaining columns are linearily dependent on these. Equivalently r is the dimension of the range of A. Furthermore it is the number of nonzero singular values.

In case of a symmetric matrix r is the number of nonzero eigenvalues. Unfortunately because of rounding errors numerical approximations of zero eigenvalues may not be zero (it may also happen that a numerical approximation is zero while the true value is not). Thus one can only calculate the numerical rank by making a decision which of the eigenvalues are close enough to zero.

Pseudoinverse

The pseudo inverse of a matrix A is the unique matrix X = A ⁺ for which AX and XA are symmetric and for which AXA = A, XAX = X holds. If A is nonsingular, then 'A ⁺ = A ^{− 1}.

When procedure jacobi (S, e, E) is called, then the relation S = E^TDiag(e)E holds where Diag(e) denotes the diagonal matrix with vector e on the diagonal. Let e ⁺ denote the vector where e_i is replaced by 1 / e_i if and by 0 if e_i is (numerically close to) zero. Since matrix E is orthogonal, it follows that the pseudo-inverse of S is given by S ⁺ = E^TDiag(e ⁺ )E.

Least squares solution

If matrix A does not have full rank, there may not be a solution of the linear system Ax = b. However one can look for a vector x for which | | Ax − b | | ₂ is minimal. The solution is x = A ⁺ b. In case of a symmetric matrix S as before, one has x = S ⁺ b = E^TDiag(e ⁺ )Eb.

Matrix exponential

From S = E^TDiag(e)E one finds expS = E^TDiag(expe)E where exp e is the vector where e_i is replaced by expe_i. In the same way f(S) can be calculated in an obvious way for any (analytic) function f.

Linear Differential Equations

The differential equation x' = A x', x(0) = a has the solution x(t) = exp (t A) a. For a symmetric matrix S it follows that x(t) = E^TDiag(expte)Ea. If is the expansion of a by the eigenvectors of S, then .

Let W^s be the vector space spanned by the eigenvectors of S which correspond to a negative eigenvalue and W^u analogously for the positive eigenvalues. If then i.e. the equilibrium point 0 is attractive to x(t). If then , i.e. 0 is repulsive to x(t). W^s and W^u are called stable and unstable manifolds for S. If a has components in both manifolds, then one component is attracted and one component is repelled. Hence x(t) approaches W^u as .

[edit] Generalizations

The Jacobi Method has been generalized to complex hermitian matrices, general nonsymmetric real and complex matrices as well as block matrices.

Since singular values of a real matrix are the squares of the eigenvalues of the symmetric matrix S = A^TA it can also be used for the calculation of these values. For this case, the method is modified in such a way that S must not be explicitly calculated which reduces the danger of round-off errors. Note that JSJ^T = JA^TAJ^T = JA^TJ^TJAJ^T = B^TB with .

The Jacobi Method is also well suited for parallelism.

[edit] References

• Jacobi, Über ein leichtes Verfahren, die in der Theorie der Säkularstörungen vorkommenden Gleichungen numerisch aufzulösen, Crelle's Journal 30 (1846), 51–94
• Schönhage, Zur quadratischen Konvergenz des Jacobi-Verfahrens, Numer. Math. 6 (1964), 410–412
• Rutishauser, The Jacobi Method for Real Symmetric Matrices, Numer. Math. 9 (1966), 1–10
• Sameh, "On Jacobi and Jacobi-like algorithms for a parallel computer", Math. Comp. 25 (1971), 579–590
• Veselic, "On a class of Jacobi-like procedures for diagonalizing arbitrary real matrices", Numer. Math. 33 (1979), 157–172
• Veselic, "A quadratically convergent Jacobi-like method for real matrices with complex conjugate eigenvalues", Numer. Math. 33 (1979), 425–435
• Shroff, "A parallel algorithm for the eigenvalues and eigenvectors of a general complex matrix", Numer. Math. 58 (1991), 779–805