Saturday, December 27, 2014

Randomized algorithms: Verifying matrix multiplication

Another great example of randomized algorithms in the introductory chapters of the book "Probability and Computing" is verifying matrix multiplication. Suppose we have 3 matrices of compatible dimensions $A, B$, and $C$, and we want to verify that
\[A\cdot B = C\]

For simplicity, let's assume that all the matrices are square, and of dimension $n \times n$. The straightforward way to do the verification is to explicitly multiple the matrices $A$ and $B$ together, an operation that is of the order of magnitude of $O(n^3)$.

As an aside, we can do better than $O(n^3)$ for matrix multiplication for larger matrices. Wikipedia has an excellent writeup on Strassen's algorithm which accomplishes matrix multiplication in $O(n^{2.8074})$ through divide and conquer. First the matrices $A$, $B$, and $C$ are padded with zero columns and rows so that their dimensions are of the form $2^m \times 2^m$. The matrices are then divided into equally sized  block matrices of the form:
\[ A = \left[ \begin{array}{cc} A_{1,1} & A_{1,2} \\ A_{2,1} & A_{2,2} \end{array} \right] \]
\[ B = \left[ \begin{array}{cc} B_{1,1} & B_{1,2} \\ B_{2,1} & B_{2,2} \end{array} \right] \]
\[ C = \left[ \begin{array}{cc} C_{1,1} & C_{1,2} \\  C_{2,1} & C_{2,2} \end{array} \right] \]
Then
\[ C_{i,j} = \Sigma_{k=1}^2 A_{i,k} B_{k,j} \]
The trick of the Strassen's algorithm is decreasing the number of required multiplications (8) by defining the auxiliary matrices $M_i$ such that:
\[ \begin{eqnarray}
M_1 & = & (A_{1,1}+A_{2,2})(B_{1,1}+B_{2,2}) \\
M_2 & = & (A_{2,1}+A_{2,2}) B_{1,1} \\
M_3 & = & A_{1,1}(B_{1,2}-B_{2,2}) \\
M_4 & = & A_{2,2}(B_{2,1}-B_{1,1}) \\
M_5 & = & (A_{1,1}+A_{1,2})B_{2,2}\\
M_6 & = & (A_{2,1} - A_{1,1})(B_{1,1}+B_{1,2}) \\
M_7 & = & (A_{1,2} - A_{2,2})(B_{2,1}+B_{2,2})
\end{eqnarray}
\]

which have $7$ multiplications instead, and expressing the result in terms of the  $M$ matrices as:
\[
\begin{eqnarray}
C_{1,1} & = & M_1 + M_4 - M_5 + M_7\\
C_{1,2} & = & M_3 + M_5 \\
C_{2,1} & = & M_2 + M_4 \\
C_{2,2} & = & M_1 - M_2 + M_3 + M_6
\end{eqnarray}
\]

The algorithm continues to divide matrices $A$, $B$, and $C$ into smaller blocks in a similar manner until there is only one element, and the multiplication becomes straightforward.

There are faster algorithms for matrix multiplication, such as the Coppersmith–Winograd algorithm with complexity of $O(n^{2.375477})$ and yet faster ones with $O(n^{2.3728639})$, but what if these are not fast enough?

Randomized algorithms come to the rescue. The idea is to choose a vector $r$ of dimension $n$, whose elements $r_i \in \{0,1\}$ and compute the quantities $ABr$ and $Cr$. If $ABr \ne Cr$ then $AB \ne C$ and the verification of matrix multiplication is done.

On the other hand if $ABr = Cr$ then either $AB=C$ or we might have a false positive. The false positive occurs when $(AB-C)r=0$, with the first term in the equation $AB-C$ not zero. The book computes the probability of this happening as less than $\frac{1}{2}$. The argument goes as follows.

Define the first term as a matrix $D=AB-C$. Since $D \ne 0$ in the false positive case, it must contain a non-zero element somewhere in the matrix. Assume this element is $d_{11}$.  We can then write $Dr=0$ as the set of $n$ simultaneous equations:

\[ \Sigma_{j=1}^{n} d_{ij}r_j=0 \]

Or for $r_1$ in the false positive case

\[ r_1 = -\frac{\Sigma_{j=2}^{n} d_{1j}r_j}{d_{11}} \]

Using the principle of deferred decisions we can figure out the probability of the false positive. We start by selecting $r_i$ for $i \in \{2, n\}$ randomly and independently from the set $\{0,1\}$, and we pause before selecting $r_1$ from the same set.  Since $r_1$ can have one of two values $\{0,1\}$, the probability of selecting $r_i$ that satisfies the prior equation to give us the false positive case is $\frac{1}{2}$

Since a false positive probability of $\frac{1}{2}$ is not good enough, we can repeat the algorithm with a new random vector $r$. After $k$ repetitions of the algorithm the false positive probability drops to $\frac{1}{2^k}$, which we can make small to our liking with many repetitions. The cost of the verification algorithm is on the order of $O(kn^2)$ multiplications, which is better than the fastest direct algorithm for explicit matrix multiplication.



No comments :

Post a Comment