3
$\begingroup$

Per Wikipedia, in Grover's algorithm the probability of observing search string $\omega$ after $r$ iterations is:

$$\left|\begin{bmatrix}\langle \omega | \omega \rangle &\langle \omega | s \rangle\end{bmatrix} (U_sU_\omega )^r \begin{pmatrix} 0 \\ 1\end{pmatrix}\right|^2.$$

How to derive this expression? Shouldn't it just have been $|\langle \omega |(U_sU_\omega)^r|s\rangle|^2$?

$\endgroup$
  • 1
    $\begingroup$ Is not the Wikipedia formula $[(scalar)\times(scalar)\times(matrix)\times(vector)]^2$, which is not, in general, a scalar, much less a probability? $\endgroup$ – Mark S Oct 26 '19 at 14:02
  • 1
    $\begingroup$ @MarkS Yes, I suppose. In my formula, I apply the $U_sU_\omega$ operator $r$ times on the initial state $|s\rangle = \frac{1}{\sqrt N}\sum_{0}^{N-1} |x\rangle$ i.e., $(U_sU_\omega)^r|s\rangle$ and find the probability of getting the search-string state $|\omega\rangle$ upon measurement of the state $|z\rangle = (U_sU_\omega)^r|s\rangle$. The required probability should simply be $|\langle \omega | z\rangle|^2$. $\endgroup$ – Sanchayan Dutta Oct 26 '19 at 14:14
  • 1
    $\begingroup$ @SanchayanDutta They aren't valid quantum gates. The transformation being described by the author is very different from the normal description of quantum circuits in terms of gates. It's a transformation within a specific plane by an $SL(2, \mathbb{R})$ transformation. There's a lot going on here, but two noteworthy aspects are that $SL(2, \mathbb{R})$ is isomorphic with $SU(1,1)$, and, since $\text{Tr}(U_s U_\omega) < 2$, $U_s U_\omega$ is elliptic and conjugate to a rotation. $\endgroup$ – Jonathan Trousdale Oct 30 '19 at 15:20
  • 1
    $\begingroup$ @SanchayanDutta No problem. It's like Grover said in the paper you cited in your related question. "Later on, there were more interpretations: inversion about average, rotation in a two dimensional Hilbert space, antenna array. All of these describe some aspect of the algorithm, but they are very different from the way it was initially invented" (last paragraph before Section 2, emphasis added). This is just another way of looking at his algorithm. $\endgroup$ – Jonathan Trousdale Oct 30 '19 at 15:30
  • 1
    $\begingroup$ There is something very deep about Grover's algorithm, and to quote his paper again "our understanding of this algorithm is still limited." As an aside, thanks for sharing that paper, I really wish more authors would write papers about the mental steps that led them to their breakthroughs. $\endgroup$ – Jonathan Trousdale Oct 30 '19 at 15:31
3
$\begingroup$

The mapping scheme set up by the author in this section is different from that of normal unitary operations. In the context of this equation, the author is referring to the (non-unitary) action of $U_s U_\omega$ on a particular two dimensional subspace of the full Hilbert space, specifically on the plane defined by the (non-orthogonal, non-complete) basis vectors $\lbrace \vert \omega \rangle, \vert s \rangle \rbrace$.

The action of $U_s U_\omega$ in this basis is shown to be $$U_s U_\omega = \begin{bmatrix} 1 & \frac{2}{\sqrt{N}} \\ \frac{-2}{\sqrt{N}} & 1- \frac{4}{N} \end{bmatrix},$$ which is clearly not unitary. $U_s U_\omega$ refers to this specific action in both the referenced equation and the following discussion. In this context, $U_s U_\omega$ is an element of $SL(2, \mathbb{R})$, and the system is set up geometrically such that $U_s U_\omega$ effects rotations within the plane spanned by $\lbrace \vert \omega \rangle, \vert s \rangle \rbrace$.

The mapping for the action of $U_s U_\omega$ is given by $$U_s U_\omega : a \vert \omega \rangle + b \vert s \rangle \mapsto \begin{bmatrix} \vert \omega \rangle & \vert s \rangle \end{bmatrix} U_s U_\omega \begin{bmatrix} a \\ b \end{bmatrix},$$ where $\begin{bmatrix} \vert \omega \rangle & \vert s \rangle \end{bmatrix}$ is an $N \times 2$ matrix. The author uses this mapping to rotate $\vert s \rangle$ into $\vert \tilde s \rangle$ (the latter designation introduced here for clarity) according to $$\vert s \rangle = 0 \vert \omega \rangle + 1 \vert s \rangle \xrightarrow{(U_s U_\omega)^r} \begin{bmatrix} \vert \omega \rangle & \vert s \rangle \end{bmatrix} (U_s U_\omega)^r \begin{bmatrix} 0 \\ 1 \end{bmatrix} = \vert \tilde s \rangle, $$ where the transformed vector, $\vert \tilde s \rangle$, is $\vert s \rangle$ rotated by $(U_s U_\omega)^r$ within the desired plane. The inner product of the rotated vector $\vert \tilde s \rangle$ with the dual vector of the marked state $\vert \omega \rangle$ is evaluated, using the Jordan normal form provided in the article, as $$\langle \omega \vert \tilde s \rangle = \begin{bmatrix} \langle \omega \vert \omega \rangle & \langle \omega \vert s \rangle \end{bmatrix} U_s U_\omega \begin{bmatrix} 0 \\ 1 \end{bmatrix}$$ $$ = \begin{bmatrix} 1 & \frac{1}{\sqrt{N}} \end{bmatrix} \begin{bmatrix} -i & i \\ e^{i t} & e^{-i t} \end{bmatrix} \begin{bmatrix} e^{i \, 2rt} & 0 \\ 0 & e^{-i \, 2rt} \end{bmatrix} \left( \tfrac{i}{2 \cos{t}} \right) \begin{bmatrix} e^{-i t} & -i \\ -e^{i t} & -i \end{bmatrix} \begin{bmatrix} 0 \\ 1 \end{bmatrix},$$ where $\langle \omega \vert s \rangle = \tfrac{1}{\sqrt{N}}$ from earlier in the article and $t \equiv \arcsin \tfrac{1}{\sqrt{N}}$. Carrying through the multiplication leads to $$\langle \omega \vert \tilde s \rangle = \tfrac{1}{2\cos t} \left(\tfrac{1}{\sqrt{N}}(e^{i(2rt+t)}+e^{-i(2rt+t)}) - i (e^{i\,2rt}-e^{-i\,2rt}) \right)$$ $$=\tfrac{1}{\cos t} \left(\sin t \cos(2rt+t) + \sin 2rt \right).$$ Using the angle sum identity for $\cos(2rt+t)$ gives $$\langle \omega \vert \tilde s \rangle = \tfrac{1}{\cos t} \left( \sin t (\cos t \cos 2rt - \sin t \sin 2rt ) + \sin 2rt \right)$$ $$=\sin t \cos 2rt + \sin 2rt \left( \tfrac{1-\sin{}^2t}{\cos t} \right) = \sin ((2r+1)t),$$ which is the identity the author was working towards. The probability of observing $\omega$ after $r$ iterations is then simply $$\vert \langle \omega \vert \tilde s \rangle \vert^2 = \sin{}^2((2r+1)t),$$ as desired. Note, however, that in this case the norm on the LHS could be replaced by parentheses, since the derivation above shows that $\langle \omega \vert \tilde s \rangle$ is guaranteed to be real. This is not surprising since the operator in the relevant plane is in $SL(2,\mathbb{R})$.

While not directly responsive to the question, there is some interesting group/representation theory underpinning the math above. Without going into detail, it seems worth noting that $\vert \text{Tr}(U_s U_\omega) \vert <2$ implies that $U_s U_\omega$ is elliptic and therefore conjugate to a rotation, and that $SL(2,\mathbb{R}) \cong SU(1,1) \cong Sp(2,\mathbb{R})$.

| improve this answer | |
$\endgroup$
  • 1
    $\begingroup$ Very clear! I saw $(\langle\omega|\omega\rangle \langle\omega|s\rangle)$ as a scalar product of two inner products, which leads to nonsense. Better maybe would have been for Wiki to write $[\langle\omega|\omega\rangle \langle\omega|s\rangle]$ to emphasize a $2\times 1$ matrix $\endgroup$ – Mark S Oct 29 '19 at 13:17
  • $\begingroup$ Agreed, the notation used in the Wikipedia article is susceptible to confusion at multiple levels. $\endgroup$ – Jonathan Trousdale Oct 29 '19 at 16:44
2
$\begingroup$

Another answer already addressed how the specific expression in the OP means and how to derive it. Here I'll show another way to derive the same result (very similar to the derivation in this other answer on another question, though with different notation).

Using the notation $\mathbb P_\psi\equiv|\psi\rangle\!\langle\psi|$, we have $U_s\equiv 2\mathbb P_s-I$ and $U_\omega\equiv I-2\mathbb P_\omega$, and thus $$U_s U_\omega=-I+2(\mathbb P_s+\mathbb P_\omega)-4\mathbb P_s\mathbb P_\omega.\tag{AA}$$

Consider the decomposition of $|s\rangle$ in a basis containing $|\omega\rangle$: $$|s\rangle=\sin\alpha |\omega\rangle+\cos\alpha e^{i\phi}|\omega_\perp\rangle,\tag{A}$$ for $\alpha,\phi\in\mathbb R$, and with $|\omega_\perp\rangle$ a state orthogonal to $|\omega\rangle$ ${}^{(1)}$. The state $|\omega_\perp\rangle$ here is uniquely determined by $|s\rangle,|\omega\rangle$, and the orthogonality requirement. Expanding the various elements of (AA) we then have $$I=\mathbb P_\omega + \mathbb P_{\omega_\perp} + \mathbb P_\perp,$$ $$\mathbb P_s=\sin^2(\alpha)\,\mathbb P_\omega + \cos^2(\alpha)\mathbb P_{\omega_\perp} + \sin\alpha\cos\alpha(e^{-i\phi}|\omega\rangle\!\langle\omega_\perp|+e^{i\phi}|\omega_\perp\rangle\!\langle\omega|),$$ $$\mathbb P_s\mathbb P_\omega=\sin^2(\alpha) \mathbb P_\omega + \sin\alpha\cos\alpha e^{i\phi}|\omega_\perp\rangle\!\langle\omega|,$$ where we defined $\mathbb P_\perp$ as the projector onto the space orthogonal to $|\omega\rangle$ and $|\omega_\perp\rangle$. Putting everything together we get $$U_s U_\omega = -\mathbb P_\perp + \cos(2\alpha)(\mathbb P_\omega+\mathbb P_{\omega_\perp}) +\sin(2\alpha)(e^{-i\phi}|\omega\rangle\!\langle\omega_\perp|-e^{i\phi}|\omega_\perp\rangle\!\langle\omega|).\tag B$$ This is telling us that $U_s U_\omega$ is really a rotation matrix in $\mathrm{span}(|\omega\rangle,|\omega_\perp\rangle\}$ (which we should have expected from it being a product of two reflections), and acts trivially (modulo minus sign) on the rest of the space.

If we assume to be working on states in $\mathrm{span}(|\omega\rangle,|\omega_\perp\rangle\}$ (which contains $|s\rangle$ by definition of $|\omega_\perp\rangle$), we can then ignore the $\mathbb P_\perp$ term. We can furthermore ease the notation by relabeling $|\omega\rangle\to|0\rangle$ and $|\omega_\perp\rangle\to|1\rangle$. Then, $$U_s U_\omega = \begin{pmatrix}\cos(2\alpha) & e^{-i\phi}\sin(2\alpha) \\ -e^{i\phi}\sin(2\alpha) & \cos(2\alpha)\end{pmatrix}\in\mathrm{SU}(2)\tag C$$ Note that all of this, going from (AA) to (C), doesn't have that much to do with Grover's algorithm or anything quantum: we are simply showing how the product of two reflections with respect to vectors with internal angle $\alpha$, amounts to a rotation by an angle of $2\alpha$.

The conclusion is now at hand: $U_s U_\omega$ is a rotation matrix in this space. From this observation, it is not hard to see that the matrix powers can be written as: $$(U_s U_\omega)^r = \begin{pmatrix}\cos(2r\alpha) & e^{-i\phi}\sin(2r\alpha) \\ -e^{i\phi}\sin(2r\alpha) & \cos(2r\alpha)\end{pmatrix}.$$ Finally, $$(U_s U_\omega)^r \begin{pmatrix}\sin\alpha \\ e^{i\phi}\cos\alpha\end{pmatrix} = \begin{pmatrix}\sin((2r+1)\alpha) \\ e^{i\phi}\cos((2r+1)\alpha)\end{pmatrix}.$$ The final probability after $r$ steps is therefore $\sin^2((2r+1)\alpha)$.

An interesting observation is that at no point in all of this we are using the fact that the dynamic happens in a high-dimensional space. Indeed, we could have identified $|\omega\rangle\sim|0\rangle$ and $|\omega_\perp\rangle\sim|1\rangle$ from the beginning and thought in terms of the evolution of a single qubit, and no generality would have been lost. In other words, Grover's algorithm really only involves an evolution in an effective two-dimensional space. You can then visualise what's going on in the Bloch sphere, and if you do that you find the state evolving along the geodesic path connecting initial and target point (which can also be understood as due to the fact that Grover's algorithm is indeed nothing but a discretisation of such geodesic path).

See also this answer on math.SE for a different way to prove this result.


${}^{(1)}$ I could write this decomposition more explicitly, being $|s\rangle$ in the OP a specific state (the balanced superposition over all states). However, I don't want to use properties of this particular $|s\rangle$, in order to make the result hold equally for any choice of $|s\rangle$.

| improve this answer | |
$\endgroup$

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.