Take the 2-minute tour ×
Mathematica Stack Exchange is a question and answer site for users of Mathematica. It's 100% free, no registration required.

Let the square nonsingular matrix $M$ is a given convergent matrix. What are the best scalar values for $\alpha$ and $\beta$ (in the real numbers domain), at which the following quadratic matrix equation approximately holds:

\begin{equation} M^2+\alpha M+\beta I\approx0. \end{equation}

In fact, I would like to find the scalar values as closely as possible for the following matrix $M$ (as an example)

Clear["Global`*"]
SeedRandom[12345]
n = 1000; Id = SparseArray[{i_, i_} -> 1., {n, n}, 0.];
A = RandomReal[{-1, 1}, {n, n}];
X = 1/SingularValueList[A, 1][[1]]^2 ConjugateTranspose[A];
M = A.X;

Also I tried to use NMinimize or FindInstance, but I failed. Please note that a rapid approach is more better for me to find more accurate values for the scalar paramteres.

share|improve this question
    
What specifically are you trying to minimize? The sum of squares of elements of M^2+alpha*M+beta*I (Frobenius norm)? Another norm? –  Daniel Lichtblau Aug 20 '13 at 21:57
    
The applied norm is optional (Frobenius or Infinity). Just finding some approximate valaues for the scalars in any fast way is needed. –  Fazlollah Soleymani Aug 21 '13 at 6:40
add comment

2 Answers

One method is to use Mathematica's PseudoInverse[] on the matrix equation that involves the eigenvalues of $M$. To arrive at this matrix equation, multiply each side of $M^2 + \alpha M + \beta I \approx 0$ by an eigenvector $v_1$ of $M$, to get

$$ (\lambda_1^2 + \alpha \lambda_1 + \beta) v_1 \approx 0 \implies \lambda_1^2 + \alpha \lambda_1 + \beta\approx 0,$$ where $\lambda_1$ is the corresponding eigenvalue for $v_1$. Repeat this for all three eigenvectors (assuming that $M$ is semi-simple) to arrive at the following approximate matrix equation

$$\begin{pmatrix}\lambda_1 & 1 \\ \lambda_2 & 1 \\ \lambda_3 & 1 \end{pmatrix} \begin{pmatrix} \alpha \\ \beta\end{pmatrix} \approx - \begin{pmatrix} \lambda_1^2 \\ \lambda_2^2 \\ \lambda_3^2 \end{pmatrix}. $$ To solve this use in Mathematica

{α,β} = PseudoInverse[{{λ1,1},{λ2,1},{λ3,1}}].{λ1^2,λ2^2,λ3^2}

Then all you need are the eigenvalues of $M$, which you can get through

{λ1,λ2,λ3} = Eigenvalues[M]

For a very large matrix it maybe impractical to find all of $M$'s eigenvalues, in which case you can use the Power Method to obtain, for example, the three largest eigenvalues and then use the above method to find α and β. The more eigenvalues you use the better the approximation.

Hope this helps.

share|improve this answer
    
If the OP doesn't have a version with PseudoInverse, then he's not using Mathematica as it was there in v1. But, it was last modified in v5, so at a minimum he should use that version. –  rcollyer Aug 21 '13 at 14:06
    
Running the above with M = DiagonalMatrix[{1,2,3}], I get {4, -(10/3)}. But, plugging this into M.M + α M + β IdentityMatrix[3] gives DiagonalMatrix[{5/3, 26/3, 53/3}]. So, I don't think your method is effective. Although, no solution may exist for the matrix I chose. Anyway to address that? –  rcollyer Aug 21 '13 at 14:19
    
Hi there, my result for this calculation gave {{1/3, 0, 0}, {0, -(2/3), 0}, {0, 0, 1/3}}, which is not entirely terrible. In any case, the bigger the spread of the eigenvalues, the worse the approximation. If M has rank 100 with "strongly" different eigenvalues, then game over, no good approximation exists. On the other hand if you use a rank 2 matrix, such as M = Outer[Times, z, w]+Outer[Times, x, y] , for any lists (vectors) z, w, x and y then the solution will be exact. @rcollyer, thanks for your comments, I've edited my post accordingly. –  Artur Gower Aug 21 '13 at 15:14
add comment

By construction, M is real symmetric and almost certainly positive definite.
The eigenvalues of M are e = (d/d[[1]])^2, where d is the vector of singular values of A.
The eigenvalues of C = M^2 + alpha M + beta are e^2 + alpha e + beta.
The eigenvectors M and C are the same as the left singular vectors of A, say V.
The offdiagonals of V'CV are zero.
The sum of squares of the diagonals of V'CV = the sum of squares of all the elements of C.
But the diagonals of V'CV are just the eigenvalues of C, so we can simplify the problem:
minimize Tr[(e^2 + alpha e + beta)^2].
You already know e, because you constructed it from d, so the solution is immediate:
{alpha = Covariance[e,-e^2]/Variance[e], beta = Mean[-e^2] - alpha*Mean[e]}.
That's the exact least-squares solution.

EDIT - It occurred to me that, although getting the singular values of a 1000 x 1000 matrix takes only a few seconds, if you have to do it many times (e.g., inside a loop) then the total time may be prohibitive, so further analysis may help.

Simulations suggest that if the elements of A are iid Uniform[-1,1] then the expected value of e is close to Reverse[(#/2 + ArcSin[#]/Pi)^2 & [Range@n/n]]. This code

{n = #, y = -(x = (#/2 + ArcSin[#]/Pi)^2 & [Range@n/n])^2;
 N@{a = (n x.y - (Tr@x)(Tr@y))/(n x.x - (Tr@x)^2),
    b = (Tr@y - a*Tr@x)/n}}& /@ {10,100,1000}

gives {alpha,beta} pairs for n = {10, 100, 1000}.

{{10, {-0.939996, 0.104015}}
 {100, {-0.764133, 0.0667884}}
 {1000, {-0.743093, 0.0626846}}}

The pair for n = 1000 should be a good approximation to the actual least-squares coefficients for a sample 1000 x 1000 matrix.

EDIT 2 - I get the same results when the elements of A are normal and logistic as when they're uniform. Has taking n = 1000 put us in asymptoticland? Would someone please check these results?

EDIT 3 - I think what we're seeing is a result of the quarter-circle law that the asymptotic density of the singular values of an n x n matrix whose elements are iid zero-mean random variables with variance 1/n converges to Sqrt[4 - x^2]/Pi, 0 <= x <= 2. The singular values we are dealing with have been normalized -- i.e., divided by the largest one -- so (rounding some statistical corners) their asymptotic density is approximately proportional to Sqrt[1 - x^2], 0 <= x <= 1, and their cumulative distribution is approximately F[x] := 2(x Sqrt[1 - x^2] + ArcSin[x])/Pi. The function I gave in my first edit, say K[p] := p/2 + ArcSin[p]/Pi, is an approximate inverse of F[x]. The actual inverse has no closed form and must be found numerically, say X[p] := Block[{x},x/.FindRoot[F[x]==p,{x,K[p]}]]. For n = 1000, the rms error in approximating the normalized singular values using X[p] is typically much smaller than the rms error using K[p]; e.g., .00095 vs .0091 .

share|improve this answer
add comment

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.