\documentclass{article}
\usepackage[dvips]{graphics}
\input{preamble.tex}
\setlength{\oddsidemargin}{.25in}
\setlength{\evensidemargin}{.25in}
\setlength{\textwidth}{6in}
\setlength{\topmargin}{-0.4in}
\setlength{\textheight}{9.5in}
\begin{document}
\lecture{11}{March 15, 2001}{Dan Spielman}{Daniel Preda}
\section*{Quantum Complexity}
Today we will define the quantum classes {\bf EQP} and {\bf BQP} (corresponding to the classical classes {\bf P} and, respectively, {\bf BPP}), and show that $\mathbf{P}\subseteq\mathbf{EQP}$ and $\mathbf{BPP}\subseteq\mathbf{BQP}$. We will also introduce the notion of {\it universal family of quantum gates}.
\paragraph{}
Recall from last lecture:
\begin{definition}
A quantum gate is a unitary matrix U ($U^{\dagger}=U^{-1}$) of size $2^n\times 2^n$, where n is the number of input (and output) wires.
\end{definition}
\paragraph{}
Please note that, besides the wires that carry the desired input state, we also allow extra input wires set to a initially known value (for instance $|0\rangle$ or $|1\rangle$.)
\paragraph{}
In order to define the size of a quantum circuit, we restrict (like in the classical case) the circuit to have only gates with at most 3 inputs (3-bit gates).
\begin{definition}
The size of a quantum circuit is the number of gates it has.
\end{definition}
\paragraph{}
Note: A quantum circuit may have a ``distinguished'' wire that says if the circuit accepts or rejects a certain input. From now on, we will assume that measurement is performed only at the end. It is easy to see that a measurement in the middle of the computation is equivalent to a measurement at the end.
\begin{theorem}
Any quantum circuit $C$ with $n$ inputs and size $f(n)$ is equivalent to a circuit $C'$ with $O(nf(n))$ inputs and of size $O(nf(n))$, where all measurements are performed at the end.
\end{theorem}
\begin{proof}
First of all, notice that, without loss of generality, we may as well assume that we measure in the z-basis (i.e. $(^1_0)$ and $(^0_1)$) (performing an appropriate change of basis if necessary). Because there are at most f(n) gates in the circuit, we can measure at most $O(nf(n))$ qubits (before or after a gate). For each qubit we measure, attach an extra wire which is initially set to $|0\rangle$. Then, replace the measurement with a Controlled-NOT gate (applies a NOT on the second qubit iff the first one is $|1\rangle$) between the qubit to be measured and the extra wire. WLOG, suppose we measure measure the first k qubits. Let the initial state be
\begin{displaymath}
\frac{1}{2^{k/2}}\sum_{i=0}^{2^k-1}{a_i|i\rangle_k} \otimes |\psi{\rangle}_{n-k} \otimes |0\rangle_k
\end{displaymath}
After the application of the Controlled-NOT between each of the first $k$ qubits and the corresponding extra qubit, we end up in the state:
\begin{displaymath}
\frac{1}{2^{k/2}}\sum_{i=0}^{2^k-1}\left ( {a_i|i\rangle_k} \otimes |\psi{\rangle}_{n-k} \otimes {|i\rangle_k} \right )
\end{displaymath}
This state evolves under the unitary operator $U$ that acts on the first $n$ qubits to:
\begin{displaymath}
\frac{1}{2^{k/2}}\sum_{i=0}^{2^k-1}\left [ U \left ( {a_i|i\rangle_k} \otimes |\psi{\rangle}_{n-k} \right ) \otimes {|i\rangle_k} \right ]
\end{displaymath}
Now if we measure the last $k$ qubits, we collapse the state into:
\begin{displaymath}
U \left ( {|i\rangle_k} \otimes |\psi{\rangle}_{n-k} \right )
\end{displaymath}
with probability $|a_i|^2$, for some $i=0...2^k-1$. This is exactly the same state we have ended up in if we had measured the first $k$ qubits, got the same value of $i$, and then evolved under $U$.
\end{proof}
\begin{definition}
{\bf EQP}=\{$L$ $\mid$ $\exists$ a family of quantum circuits $\{Q_n\}_n$, where $Q_n$ takes $n$ inputs, and a ptime turing machine $M$ that on input $1^n$ outputs $Q_n$ such that:
\begin{displaymath}
w \in L \Leftrightarrow Pr \left [ Q_{\mid w \mid }(w) = 1 \right ] = 1
\end{displaymath}
\begin{displaymath}
w \notin L \Leftrightarrow Pr \left [ Q_{\mid w \mid }(w) = 1 \right ] = 0 \ \}
\end{displaymath}
\end{definition}
\begin{definition}
{\bf BQP}=\{$L$ $\mid$ $\exists$ a family of quantum circuits $\{Q_n\}_n$, where $Q_n$ takes $n$ inputs, and a ptime turing machine $M$ that on input $1^n$ outputs $Q_n$ such that:
\begin{displaymath}
w \in L \Leftrightarrow Pr \left [ Q_{\mid w \mid }(w) = 1 \right ] \geq 2/3
\end{displaymath}
\begin{displaymath}
w \notin L \Leftrightarrow Pr \left [ Q_{\mid w \mid }(w) = 1 \right ] \leq 1/3 \ \}
\end{displaymath}
\end{definition}
Before proving that $\mathbf{P} \subseteq \mathbf{EQP}$ and $\mathbf{BPP} \subseteq \mathbf{BQP}$, there are some remarks to be made:\\
It can be proved that the transition amplitudes of the quantum gates need only be accurate to $poly(\log T)$ bits of precision for a quantum circuit of size $T$.\\
Because any quantum gate $U$ is unitary, $U^{-1}$ is also unitary and hence every quantum gate is reversible. As a parentheses, there has been extended work in the field of reversible computation well before the emergence of quantum computing. The Landauer's principle states that when a bit of information is erased, it must either be transfered to another bit or be dissipated as $k_B \ln 2$ entropy ($k_B$ is Boltzmann's constant, approx. $1.38\times 10^{-23} J/K$). Hence, erasing one bit creates heat in an amount of $k_BT \ln 2$, where $T$ is the temperature.
\begin{theorem}
$\mathbf{P} \subseteq \mathbf{EQP}$
\end{theorem}
\begin{proof}
Suppose we have a classical circuit $C$ that computes $f:\{0, 1\}^n \rightarrow \{0, 1\}^m$, has only AND, OR or NOT gates and has size $s(n)$. We construct a quantum circuit $C'$ of size $O(s(n)+m)$ computing $Q_f:\{0, 1\}^{n+m} \rightarrow \{0, 1\}^{n+m}$ that maps each pair $(x, y)$ into $(x, y\oplus f(x))$, $x\in\{0, 1\}^n$, $y\in\{0, 1\}^m$, using at most $s(n)$ additional qubits, initially set to $|0\rangle$.
\paragraph{}
First we replace each classical gate with its quantum version:\\
NOT:
\begin{center}
\includegraphics{not.eps}
\end{center}
The Control is done here on first bit being $|0\rangle$. The last qubit ends up in state $|\bar{a}\rangle$\\
AND:
\begin{center}
\includegraphics{and.eps}
\end{center}
The Control is done here on $|1\rangle$. The last qubit ends up in state $|a\wedge b\rangle$\\
OR:
\begin{center}
\includegraphics{or.eps}
\end{center}
The Control is done here on $|1\rangle$. The last qubit ends up in state $|a\vee b\rangle$. Note that OR can also be built from NOT and AND.\\
\paragraph{}
We design the quantum circuit as follows: we use one wire for each input (``input'' qubits), one for each (classical) gate (initially set to $|0\rangle$) (``gate'' qubits), and one for each output (initially set to $|0\rangle$) (``output'' qubits). Then, using the quantum gates described above, and following the description of the classical circuit, we assign to each of the ``gate'' qubits the value of the output bit of that classical gate. When we finished all the gates, we just attach Controlled-NOTs between the ``gate'' qubits corresponding to the classical output bits, and the ``output'' qubits, thus copying the output into the ``output'' bits.
\paragraph{}
If we wish to return the ``gate'' qubits to $|0\rangle$, we could apply all the gates in reverse order (note that a Control applied twice is identity). This is useful because otherwise the ``output'' qubits remain entangled with the ``gate'' qubits and this can mess up further computations on the ``input'' and ``output'' qubits.
\end{proof}
\begin{theorem}
$\mathbf{BPP} \subseteq \mathbf{BQP}$
\end{theorem}
\begin{proof}
The idea is simple: use same proof as above, but also add some ``random'' qubits initially set to $|0\rangle$. Then apply Hadamards on all those qubits, turning them into $1/\sqrt{2}(|0\rangle + |1\rangle)$. Whenever we want to flip a coin, do a Controlled-NOT on one of those qubits (use different qubits for different coin tosses). At the end, measure all the ``random'' qubits, and obtain $(|0\rangle$ or $|1\rangle)$ with probability $1/2$. By an argument similar to that used in Theorem 3, this is equivalent to measuring them in the middle of the computation and using the random value right away.
\end{proof}
\begin{theorem}
Any quantum gate can be built from quantum gates with at most 3-bit inputs (and even with at most 2-bit inputs)
\end{theorem}
\paragraph{}
We may attempt to prove this theorem next lecture. For now, we will focus on finding a family of at most 3-bit input gates that can be used as the above theorem claims. Clearly, what we would like is to be able to generate any 3-bit input gate.
\begin{definition}
We define a ``universal family of quantum gates'' to be a family of gates of dimension of up to $8\times8$ such that any $8\times8$ unitary matrix is near a matrix generated by this family.
\end{definition}
\paragraph{}
We now define what ``near'' means:
\begin{definition}
Let
\begin{displaymath}
|M|=\sum_{i, j}{|m_{i, j}|}
\end{displaymath}
\end{definition}
\paragraph{}
We want a family $\cal{F}$ such that $\forall U, \epsilon >0, \ \exists U^{'}$ generated by gates in family such that $|U-U^{'}|<\epsilon$ and only need $poly \left ( 1/\epsilon \right )$ gates to construct $U^{'}$ (actually it can be proved that we can do this even with $polylog \left ( 1/\epsilon \right )$ gates)
\begin{theorem}
Let $v$ be a unit vector, and $U$, $U^{'}$ two unitary matrices such that $|U-U^{'}|<\epsilon$. Then the angle between $Uv$ and $U^{'}v$ is at most $\epsilon$, for $\epsilon$ small.
\end{theorem}
\begin{proof}
Let $A=U-U^{'}$. The length of the vector $Uv-U^{'}v$ is:
\begin{displaymath}
\parallel Av \parallel = \sqrt{\sum_i{\mid \sum_j{A_{ij}v_j} \mid^2}} \leq \sum_i{\mid \sum_j{A_{ij}v_j} \mid} \leq \sum_{ij}{\mid A_{ij}v_j \mid}
\end{displaymath}
But $\mid v_j \mid \leq 1$, so we get:
\begin{displaymath}
\parallel Av \parallel \leq \sum_{ij}{\mid A_{ij} \mid} \leq \epsilon
\end{displaymath}
This completes the proof, as $\parallel Uv \parallel = \parallel U^{'}v \parallel =1$
\end{proof}
Now we can show that if we have $m$ gates, each with error $\epsilon$, the errors add up:
\begin{theorem}
Let $U_i$, and respectively $U^{'}_i$, $i=1...m$ be unitary matrices of $N\times N$. If, $\forall i$, $|U_i-U^{'}_i|<\epsilon$, then $\forall v$, the angle between $U_1U_2...U_mv$ and $U^{'}_1U^{'}_2...U^{'}_mv$ is at most $m\epsilon$, for $\epsilon$ small.
\end{theorem}
\begin{proof}
We'll prove by induction. We already proved the base case. Inductive step:\\
Suppose the angle between $U^{'}_1..U^{'}_kU_{k+1}...U_mv$ and $U_1...U_mv$ is at most $k\epsilon$. The angle between $U^{'}_{k+1}...U_mv$ and $U_{k+1}...U_mv$ is at most $\epsilon$ and because unitary matrices preserve the angle, we have that the angle between $U^{'}_1..U^{'}_kU^{'}_{k+1}...U_mv$ and $U^{'}_1..U^{'}_kU_{k+1}...U_mv$ is at most $\epsilon$. Thus, the angle between $U^{'}_1..U^{'}_kU^{'}_{k+1}...U_mv$ and $U_{k+1}...U_mv$ is at most $(k+1)\epsilon$.
\end{proof}
\paragraph{}
The above theorems prove that if $\epsilon=o(1/m)$ then we have almost infinite precision. Now we will try to define a ``universal'' family of 1-bit gates.
\begin{fact}
Any $2\times 2$ unitary matrix can be decomposed as:
\begin{displaymath}
\left(\begin{array}{cc}e^{i\delta}&0\\0&e^{i\delta} \end{array}\right)\cdot
\left(\begin{array}{cc}e^{i\alpha/2}&0\\0&e^{-i\alpha/2} \end{array}\right)\cdot
\left(\begin{array}{cc}\cos\theta/2&\sin\theta/2\\-\sin\theta/2&\cos\theta/2 \end{array}\right)\cdot
\left(\begin{array}{cc}e^{i\beta/2}&0\\0&e^{-i\beta/2} \end{array}\right)\cdot
\end{displaymath}
\end{fact}
This is true as any element of $SU(2)$ is generated by rotations and phase shifts. The first matrix in the above expansion is just an overall phase.
\paragraph{}
All we need to do is to find a value $\theta_0$ such that, $\forall \theta \in [0, 2\pi]$, $\mid k\theta_0-\theta\mid mod \ 2\pi \leq \epsilon$, and $k$ is bounded by (and can be computed in) some polynomial in $1/\epsilon$.
\begin{claim}
$2\pi/\sqrt{2}$ is a good value for $\theta_0$ above.
\end{claim}
\paragraph{}
The above result can be intuitively understood by observing that multiples of $\theta_0$ land uniformly on the circle. In {\it ``Quantum complexity theory''} (available at http://http.cs.berkeley.edu/~vazirani/), E. Bernstein and U. Vazirani prove that $\theta_0=2\pi\sum_{i=1}^{\infty}{2^{-2^{i}}}$ is a good value for $\theta_0$ and they give a $polylog(1/\epsilon)$ algorithm that outputs a $k$ bounded by a polynomial in $1/\epsilon$.
\paragraph{}
If we want to generalize the above results for 3-bit gates, the idea (which may be expanded next lecture) is to create gates Control-Control-Rotation and Control-Control-Phase, with the same $\theta_o$ as above.
\end{document}