I'm trying to compute the spectral decomposition of the Hadamard gate but I'm making a mistake somewhere.

Note: I believe (though I may be wrong so correct me if I am) that spectral decomposition is a way to find a diagonalized version of a matrix, additionally I am trying to work out $\sqrt{\operatorname{CNOT}_{12}}$, and based off an exam paper question it said to do so knowing that $\operatorname{CNOT}_{12}=H_2Z_{12}H_2$. I know that $H_2=I \otimes H$, and I thought the best way to do this would be to diagonalize $H_2$ by diagonalizing $H$ then take the square root of the diagonals of $H_2, Z$ to find $\sqrt{\operatorname{CNOT}_{12}}$.

Say we have $$H=\begin{bmatrix} \tfrac{1}{\sqrt{2}} &\tfrac{1}{\sqrt{2}} \\ \tfrac{1}{\sqrt{2}} & -\tfrac{1}{\sqrt{2}} \end{bmatrix}.$$

Now spectral decomposition of this matrix will be $H=\sum_i \lambda_i|\psi_i\rangle \langle\psi_i|$, where $\lambda_i$ corresponds to an eigenvalue and $|\psi_i \rangle$ is its associated eigenvector.

First we find the eigenvalues:

$$\det(H-\lambda I)=\det \begin{bmatrix} \tfrac{1}{\sqrt{2}}-\lambda &\tfrac{1}{\sqrt{2}} \\ \tfrac{1}{\sqrt{2}} & -\tfrac{1}{\sqrt{2}}-\lambda \end{bmatrix}$$ $$=(\tfrac{1}{\sqrt{2}}-\lambda)(-\tfrac{1}{\sqrt{2}}-\lambda)-\tfrac{1}{2}$$ $$=-\tfrac{1}{2}+\lambda^2-\tfrac{1}{2}=-1+\lambda^2\implies \lambda=\pm 1$$

Now we find the eigenvectors:


$$\begin{bmatrix} x\\y \end{bmatrix}=\begin{bmatrix} \tfrac{1}{\sqrt{2}} &\tfrac{1}{\sqrt{2}} \\ \tfrac{1}{\sqrt{2}} & -\tfrac{1}{\sqrt{2}} \end{bmatrix}\begin{bmatrix} x\\y \end{bmatrix}$$

$$\implies \tfrac{x+y}{\sqrt{2}}=x \implies (\sqrt{2}-1)x=y$$

$$\tfrac{x-y}{\sqrt{2}}=y \implies (\sqrt{2}+1)y=x$$

These equations give eigenvectors $$v_1=\begin{bmatrix} 1 \\(\sqrt{2}-1) \end{bmatrix}, v_2=\begin{bmatrix} (\sqrt{2}+1)\\1 \end{bmatrix}$$

The eigenvectors for $\lambda=-1$ are found, similarly, to be $$v_3=\begin{bmatrix} 1 \\(-\sqrt{2}-1) \end{bmatrix}, v_4=\begin{bmatrix} (-\sqrt{2}+1)\\1 \end{bmatrix}$$ But $$H=-\begin{bmatrix} 1 \\(-\sqrt{2}-1) \end{bmatrix}\begin{bmatrix} 1 &(-\sqrt{2}-1) \end{bmatrix}- \begin{bmatrix} (-\sqrt{2}+1)\\1 \end{bmatrix}\begin{bmatrix} (-\sqrt{2}+1)&1 \end{bmatrix}+\begin{bmatrix} 1 \\(\sqrt{2}-1) \end{bmatrix}\begin{bmatrix} 1 &(\sqrt{2}-1) \end{bmatrix}+\begin{bmatrix} (\sqrt{2}+1)\\1 \end{bmatrix}\begin{bmatrix} (\sqrt{2}+1)&1 \end{bmatrix}$$

Doesn't give me a diagonal matrix, where have I gone wrong?

  • $\begingroup$ I don't know why they require you to write $\operatorname{CNOT}_{12}=H_2Z_{12}H_2$. For your method of taking square roots of diagonal elements to work, $H_2$ and $Z_{12}$ would have to be simultaneously diagonalizable (i.e. they at least need to commute). Did you check that? $\endgroup$ Apr 28, 2019 at 7:59
  • $\begingroup$ There is no "diagonalized version" of the matrix. Matrix is a matrix. But matrix can be unitary equivalent to some diagonal matrix. That is, $H=UDU^*$ for some unitary $U$ and diagonal $D$. $\endgroup$
    – Danylo Y
    Apr 28, 2019 at 12:03
  • $\begingroup$ @SanchayanDutta I checked and I see they're not simultaneously diagonalizable (They don't commute ) , this topic you mention was not something we'd not covered so I didn't know I had to check it . So it seems my proposed method won't work. How can we find the square-root of $CNOT_{12}$ then , using the fact that $CNOT_{12}=H_2Z_{12}H_2$ ? $\endgroup$
    – bhapi
    Apr 28, 2019 at 15:13
  • $\begingroup$ @SanchayanDutta I saw you edited your post and I agree that it does seem like a much easier method, it's just that on a past exam paper it said to do it the round-about way, probably exactly because it's round-about and takes more effort I suppose $\endgroup$
    – bhapi
    Apr 28, 2019 at 15:18
  • 1
    $\begingroup$ @SanchayanDutta The exact question (for clarity) is : The controlled phase flip can be converted to the CNOT gate by $H_2Z_{12}H_2=CNOT_{12}$, use this relation and the properties of the gates involved to determine the operation $\sqrt{CNOT_{12}}$ Hint :Split the controlled phase gate into a sequence of two square roots of the same gate first. I don't know if that helps at all . I'm going to keep working on the problem and if I cant get it I'll ask my lecturer in the end and post an answer if no one else does . But if the full question has shed any light please do let me know $\endgroup$
    – bhapi
    Apr 28, 2019 at 15:25

2 Answers 2


Firstly, there's a conceptual error in your calculation of the eigenvectors.

$$\begin{bmatrix} x\\y \end{bmatrix}=\begin{bmatrix} \tfrac{1}{\sqrt{2}} &\tfrac{1}{\sqrt{2}} \\ \tfrac{1}{\sqrt{2}} & -\tfrac{1}{\sqrt{2}} \end{bmatrix}\begin{bmatrix} x\\y \end{bmatrix}$$

$$(\sqrt{2}-1)x=y \tag{1}$$

$$(\sqrt{2}+1)y=x \tag{2}$$

(1) and (2) are equivalent equations. To convince yourself multiply both sides of (1) by $(\sqrt{2}+1)$. The point is that when you solve the eigenvalue equation, you do not get a single eigenvector but rather a linear subspace of eigenvectors. We can only put forth the form of the eigenvectors in that subspace as:

$$\begin{bmatrix}x \\ (\sqrt{2}-1)x\end{bmatrix}=x\begin{bmatrix}1 \\ (\sqrt{2}-1)\end{bmatrix}$$ where $x\in \Bbb C$.

Your other vector $\nu_2$ lies in this same linear subspace. It's equivalent to $(\sqrt{2}+1)\nu_1$.

And again, $\nu_4$ is simply $(-\sqrt{2}+1)\nu_3$.

Secondly, for finding the square root of $\operatorname{CNOT}_{12}$ you could directly diagonalize $\operatorname{CNOT}_{12}$ instead of going about it in such a roundabout fashion (i.e. writing $\operatorname{CNOT}_{12}$ as $H_2Z_{12}H_2$). However, that approach is particularly useful for this specific case; see my update below.

It's a block diagonal matrix, so you can easily diagonalize the individual $2\times 2$ blocks.

Proceed like this:

Spectral decompose the upper left block like $P\Lambda P^{-1}$ and the lower right block like $Q\Lambda'Q^{-1}$. Then $$\operatorname{CNOT}_{12}=\begin{pmatrix}P&0\\0&Q\end{pmatrix}\begin{pmatrix}\Lambda&0\\0&\Lambda'\end{pmatrix}\begin{pmatrix}P^{-1}&0\\0&Q^{-1}\end{pmatrix}.$$

The required square root is simply $$\sqrt{\operatorname{CNOT}_{12}}=\begin{pmatrix}P&0\\0&Q\end{pmatrix}\begin{pmatrix}\sqrt\Lambda&0\\0&\sqrt\Lambda'\end{pmatrix}\begin{pmatrix}P^{-1}&0\\0&Q^{-1}\end{pmatrix}.$$


Initially, I didn't understand the significance of writing the $\operatorname{CNOT}_{12}$ as $H_2Z_{12}H_2$ but the hint in the question paper helped. One obvious fact (as $H_2.H_2=\Bbb I$) is that $$\operatorname{CNOT}_{12}=H_2Z_{12}H_2 = (H_2\sqrt{Z_{12}}H_2)(H_2\sqrt{Z_{12}}H_2),$$ and so $$\sqrt{\operatorname{CNOT}_{12}}=H_2\sqrt{Z_{12}}H_2.$$

Note that $Z_{12}$ is the controlled phase gate for $\phi=\pi$.

Now, an interesting property of the controlled phase gate is:

$$\begin{pmatrix} 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & e^{i\phi/2} \end{pmatrix}.\begin{pmatrix} 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & e^{i\phi/2} \end{pmatrix}= \begin{pmatrix} 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & e^{i\phi} \end{pmatrix}$$

So the square root of the $Z_{12}$ gate i.e. $\sqrt{Z_{12}}$ is $$\begin{pmatrix} 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & e^{i\pi/2} \end{pmatrix}.$$

Now just pre- and post-multiply $\sqrt{Z_{12}}$ with $H_2$'s and you're done! This is indeed simpler than the spectral decomposition method I suggested earlier.

  • 1
    $\begingroup$ Thank you for all your help :) $\endgroup$
    – bhapi
    Apr 28, 2019 at 16:40

There are two points where you are wrong.

First: When you compute the eigenvector of H, which is $2\times 2$ matrix. So there should be one eigenvector corresponding to an eigenvalue, and should be normalized. In your computation for the case of $\lambda=1$, if you make a normalization you will find that there are the same.

Second: If you just want to get the square root of a self-adjoint one, then just make a square root of the eigenvalue is fine. That is for a self-adjoint operator: $$A = \sum_i \lambda_i |\psi_i\rangle\langle\psi_i|$$ then $$\sqrt{A} = \sum_i \sqrt{\lambda_i}|\psi_i\rangle\langle\psi_i| $$


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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