[2] Solving systems of first order differential equations [3] Symmetric and Orthonormal Matrices
[1] Diagonal factorization
Recall: if Ax = λx, then the system ẏ = Ay has a general solution of the form y = c1 eλ1 t x1 + c2 eλ2 t x2 , where the λi are eigenvalues with corresponding eigenvectors xi . I’m never going to see eigenvectors without putting them into a matrix. And I’m never going to see eigenvalues without putting them into a matrix. Let’s look at an example from last class.
5 2 1 1 A= . We found that this had eigenvectors and . 2 5 1 −1 I’m going to form a matrix out of these eigenvectors called the eigenvector matrix S: 1 1 S= 1 −1
Then lets look at what happens when we multiply AS, and see that we can factor this into S and a diagonal matrix Λ:
5 2 1 1 7 3 1 1 7 0 = = 2 5 1 −1 7 −3 1 −1 0 3
A S S Λ
We call matrix Λ with eigenvalues λ on the diagonal the eigenvalue matrix.
So we see that AS = SΛ, but we can multiply both sides on the right by S and we get a factorization A = SΛS −1 . We’ve factored A into 3 pieces. −1
1 Properties of Diagonalization
• A2 = SΛS −1 SΛS −1 = SΛ2 S −1
• A−1 = (SΛS −1 )−1 = (S −1 )−1 Λ−1 S −1 = SΛ−1 S −1
Diagonal matrices are easy to square and invert because you simply square or invert the elements along the diagonal!
[2] Solving systems of first order differential
equations The entire reason we are finding eigenvectors is to solve differential equations. Let’s express our solution to the differential equation in terms of S and Λ:
eλ1 t 0 c1 y = x 1 x2 0 eλ2 t c2
S eΛt c
What determines c? Suppose we have an initial condition y(0). Plugging
this into our vector equation above we can solve for c:
y(0) = SIc −1 S y(0) = c
The first line simply expresses our initial condition as a linear combination of the eigenvectors, y(0) = c1 x1 + c2 x2 . The second equation just multiplies the first by S −1 on both sides to solve for c in terms of y(0) and S −1 , which we know, or can compute from what we know.
Steps for solving a differential equation
Step 0. Find λi and xi .
2 Step 1. Use the initial condition to compute the parameters:
c = S −1 y(0)
Step 2. Multiply c by eΛt and S:
y = SeΛt S −1 y(0).
[3] Symmetric and Orthonormal Matrices
In our example, we saw that A was symmetric (A = AT ) implied that the eigenvectors were perpendicular, or orthogonal. Perpendicular and orthogo- nal are two words that mean the same thing. Now, the eigenvectors we chose
1 1 and 1 −1 √ had length 2. If we make them unit length, we can choose eigenvectors that are both orthogonal and unit length. This is called orthonormal. Question: Are the unit length vectors also eigenvectors? √ √ 1/√2 1/√2 and 1/ 2 −1/ 2
Yes! If Ax = λx, then
x x A =λ . ||x|| ||x|| It turns out that finding the inverse of a matrix whose columns are or- thonormal is extremely easy! All you have to do is take the transpose!
If the inverse exists, it is unique, so S T must be the inverse!
If we set θ = π/4 we get
1 1 −1 √ , 2 1 1 but what we found was 1 1 1 √ . 2 1 −1 Fortunately we can multiply the second column by negative 1, and it is still and eigenvector. So in the 2 by 2 case, we can always choose the eigenvectors of a symmetric matrix so that the eigenvector matrix is not only orthonormal, but also so that it is a rotation matrix! In general, a set of vectors x1 , . . . , xn is said to be orthonormal if the dot product of any vector with itself is 1:
xi · xi = xTi xi = 1,
and the dot product of any two vectors that are not equal is zero:
xi · xj = xTi xj = 0,
when i 6= j. This tells us that the matrix product:
We’ve seen that 2 by 2 orthonormal eigenvector matrices can be chosen to
be rotation matrices. Let’s look at a 3 by 3 rotation matrix: 2 2 −1 1 S = −1 2 2 3 2 −1 2
As an exercise, test that all vector dot products are zero if the vectors are not equal, and are one if it is a dot product with itself. This is a particularly nice matrix because there are no square roots! And this is also a rotation matrix! But it is a rotation is 3 dimensions. Find a symmetric matrix A whose eigenvector matrix is S. All we have to do is choose any Λ with real entries along the diagonal, and then A = SΛS T is symmetric! Recall that (AB)T = B T AT . We can use this to check that this A is in fact symmetric:
AT = (SΛS T )T = S T T ΛT S T = SΛS T
This works because transposing a matrix twice returns the original ma- trix, and transposing a diagonal matrix does nothing! In physics and engineering this is called the principal axis theorem. In math, this is the spectral theorem.
5 Why is it called the principal axis theorem? An ellipsoid whose principal axis are along the standard x, y, and z axes can be written as the equation ax2 + by 2 + cz 2 = 1, which in matrix form is a 0 0 x x y z 0 b 0 y = 1 0 0 c z
However, what you consider a general ellipsoid, the 3 principal direction
can be pointing in any direction. They are orthogonal direction though! And this means that we can get back to the standard basis elements by applying a rotation matrix S whose columns are orthonormal. Thus our equation for a general ellipsoid is: T x a 0 0 x S y 0 b 0 S y = 1 z 0 0 c z a 0 0 x x y z S T 0
b 0 S y = 1 0 0 c z
6 M.I.T. 18.03 Ordinary Differential Equations 18.03 Extra Notes and Exercises c Haynes Miller, David Jerison, Jennifer French and M.I.T., 2013
