Is $$\det(I+ ABB^*A^*C^{-1})=\det(I+ B^*A^*ABC^{-1})$$ where $I$ is identity matrix, $A,B,C$ are complex valued matrices. And $C$ is $(I+X)$ where $X$ is PSD.

I know that this makes $ABB^*A^*$ and $B^*A^*AB$ positive semi/definite. And I know for square matrices $X,Y,Z$ we have $\det(XYZ)=\det(YXZ)$ so $\det (ABB^*A^*C^{-1})=\det(B^*A^*ABC^{-1})$. And also $\det (I + ABB^*A^*C^{-1})=\det(I + B^*A^*ABC^{-1})$. Please comment if this is correct. I have a feeling I'm missing something since I didn't use properties of $C$.

Simulations seem to hold. At least when I used $A=\pmatrix{1&1\\ 0&0},\ B=I$ and $C=\pmatrix{2&0\\ 0&2}$ gave 2 in both sides.

Thanks a lot in advance.

PS. Not home work.

  • 1
    $\begingroup$ Why don't you try a few instances of $A,B,C$ to see whether the hypothesis is true? For simplicity, you may set $B=I$. $\endgroup$ – user1551 Oct 6 '13 at 18:29

Your test example is too specific: $B = I$, $C = 2I$, $C^{-1} = (1/2)I$. Of course that everything here commutes.

I tried the following random one (in Python3):

import numpy as np

A = np.array([[3,2,6],[4,1,5]])
B = np.array([[1,2],[3,4],[5,6]])
C = np.array([[2,1],[1,3]])

AB = np.dot(A, B)
Cinv = np.linalg.inv(C)

M1 = np.identity(2) + np.dot(np.dot(AB, AB.transpose()), Cinv)
M2 = np.identity(2) + np.dot(np.dot(AB.transpose(), AB), Cinv)
print("A =\n", A)
print("B =\n", B)
print("AB =\n", AB)
print("C =\n", C)
print("C^{-1} =\n", Cinv)
print("M1 =\n", M1)
print("M2 =\n", M2)
print("det M1 =", np.linalg.det(M1))
print("det M2 =", np.linalg.det(M2))

The output says that your statement is wrong:

A =
 [[3 2 6]
 [4 1 5]]
B =
 [[1 2]
 [3 4]
 [5 6]]
AB =
 [[39 50]
 [32 42]]
C =
 [[2 1]
 [1 3]]
C^{-1} =
 [[ 0.6 -0.2]
 [-0.2  0.4]]
M1 =
 [[ 1744.    535. ]
 [ 1451.2   446.6]]
M2 =
 [[  869.2   808.6]
 [ 1123.6  1047.8]]
det M1 = 2478.4
det M2 = 2204.8

Your proof is wrong because

$$\det\left( \prod_k M_k \right) = \det\left( \prod_k M_{p(k)} \right),$$

for all permutations $p$, but, generally,

$$\det\left( I + \prod_k M_k \right) \ne \det\left( I + \prod_k M_{p(k)} \right).$$

Even if you limit $A$ and $B$ to be square, the statement won't hold. For example, remove the third column/row from $A$ and $B$ above:

A = np.array([[3,2],[4,1]])
B = np.array([[1,2],[3,4]])
C = np.array([[2,1],[1,3]])

The result is:

det M1 = 172.0
det M2 = 151.0

As user1551 told you: try some random example first. Not "nice" ones like your example above, but completely random ones. If a statement is wrong, chances are that a random test will sink it.

  • $\begingroup$ Thank you. Could you please explain why $$\det\left( I + \prod_k M_k \right) \ne \det\left( I + \prod_k M_{p(k)} \right).$$ in positive semi-definite case? I know $\det(I+AB)=\det(I+BA)$ when $A,B$ are PSD. $\endgroup$ – triomphe Oct 11 '13 at 18:45
  • $\begingroup$ @MLT This formula works for general (not just PSD) $A$ and $B$ of appropriate orders (it's called Sylvester's determinant theorem. However, it doesn't hold for more than two matrices. What I wrote means that such identity is not always correct. For some matrices $M_k$, however many of them, you will have equality there. But not for any set of three or more matrices. $\endgroup$ – Vedran Šego Oct 11 '13 at 19:14

Edit: Try $A=\pmatrix{1&1\\ 0&0},\ B=I,\ C=\pmatrix{2&1\\ 1&2}$ and $C^{-1}=\frac13\pmatrix{2&-1\\ -1&2}$. The two determinants are $\frac53$ and $\frac73$, which are not equal. For questions like this, you may first do a few numerical experiments with a computer to see if the hypothesis can be easily falsified.

  • $\begingroup$ I changed the definition of $C$. Thanks $\endgroup$ – triomphe Oct 7 '13 at 19:35
  • $\begingroup$ @MLT The statement is still false. See my new edit. $\endgroup$ – user1551 Oct 8 '13 at 16:28
  • $\begingroup$ How do you think I can fix it? please, Thank you. $\endgroup$ – triomphe Oct 8 '13 at 16:42
  • 1
    $\begingroup$ @MLT I am sorry. Except the trivial conditions (e.g. $AB$ is normal or $C^{-1}$ commute with $AB$ or $B^\ast A^\ast$, nothing came to mind immediately. $\endgroup$ – user1551 Oct 9 '13 at 15:01

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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