Thursday, December 22, 2016

Simpler Expected Analysis of Randomized Selection Algorithm.

In this post I present a simpler proof to derive expected asymptotic bounds for the Randomized Selection algorithm -- Las Vegas version to be more precise. Given a set $S = \{ a_1\,a_2\,a_3\ldots a_n \}$ of keys and an integer $k \in [1,n]$ the selection problem asks to report the $k^{th}$ largest key in the set $S$. Recall that this problem can be solved optimally using $O(n)$ comparison operations using the BFPRT (Blum, Floyd, Pratt, Rivest and Tarjan 1973) algorithm. However a correct implementation of BFPRT algorithm may take a lot of code. On the other hand randomized algorithms for many deterministic algorithms are often simpler to implement. The algorithm chooses a random pivot for partitioning the input and recursively picks on of the partition until the partition (sub problem) size is just one. Following is an implementation of RandSelect in around 25 lines.

Our goal here is the analyze the expected run-time and expected number of steps for the problem size converge to unity -- more precisely the expected value of the variable rand_walk_steps in the following code. I have purposefully kept the analysis very rigorous to make it accessible to everyone.

TL;DR "Expected number of comparisons in RandSelect is $\Theta(n)$ and expected number of rand_walk_steps is $\Theta(\log(n))$".

// Randomized selection algorithm to pick k^th largest element
// in a mutable array.
template < typename T > 
T RandSelect(T *array, size_t n, size_t k) {
  assert(k > 0 && k <= n); 
  size_t problem_size = n, select = k;

  // What is the expected value of rand_walk_steps ??
  size_t rand_walk_steps = 0;  

  while (problem_size > 1) {
    size_t i = 0, j = problem_size - 1;
    size_t ridx = RandArrayIndex(problem_size);
    const T pivot = array[ridx];

    // Partition the sub problem with pivot.
    while (i <= j) {
      if (array[i] <= pivot) {
        i++;
      } else if (array[j] > pivot) {
        j--;
      } else {
        std::swap(array[i], array[j]);
      }   
    }   

    problem_size = (i >= select) ? i : problem_size - i;
    array = (i >= select) ? array : &array[i];
    select = (i >= select) ? select : select - i;
    ++rand_walk_steps;
  }
  return array[0];
}


Let $X_i$ be the random variable corresponding to problem size in $i^{th}$ step ($1 \geq i \geq n-1$) in the algorithm -- each step corresponds to one iteration of the outer most while loop. If the algorithm terminates in $j$ steps then all $X_k$ with $k>j$ are set to $0$.
$$
\begin{array}{lll}
X_i &=& \text{Random variable for problem size in }i^{th}\text{ iteration of the outer most}\textsf{ while}\text{ loop} \\
X_i && \left\{ \begin{array}{lr} \in [1,X_{i-1}-1] & \textsf{If } X_{i-1} > 1 \\
= 0 & \textsf{Otherwise}
\end{array} \right. \\
X_0 &=& n \,\,\,\,\,\,\, \text{Initial problem size and }\,\,\, E[X_0] = n \\
&& \\
\end{array}
$$

Lemma-1: If $X_{i-1} >1$ Then $E[X_i | X_{i-1}] = \frac{X_{i-1}}{2}$

Proof: For a given $X_{i-1} > 1$ the random variable $X_{i}$ is uniformly distributed in the range $[1,X_{i-1}-1]$. Hence the expected value is simply the average $\frac{1 + (X_{i-1}-1)}{2} = \frac{X_{i-1}}{2}$ $\Box$.

Lemma-2: $E[X_i] = \frac{n}{2^i}$

Proof:
$$
\begin{array}{llll}
&E[X_i|X_{i-1}] &=& \frac{X_{i-1}}{2}\,\,\,\,\,\text{From Lemma-1}\\
&E[\,E[X_i|X_{i-1}]\,] &=& E[\frac{X_{i-1}}{2}] \\
\Rightarrow & E[X_i] &=& E[\frac{X_{i-1}}{2}] = \frac{E[X_{i-1}]}{2}\,\,\,\,\,\text{for any two random variables } X,Y\,\,E[E[X|Y]] = E[X]\\
&&=&\frac{1}{2}\times E[X_{i-1}] = \frac{1}{2^2}\times E[X_{i-2}] \ldots \,\,\,\,\, \text{Apply recursively} \\
&&\ldots& \\
&&=& \frac{1}{2^i}\times E[X_{i-i}] = \frac{E[X_0]}{2^i} = \frac{n}{2^i} \,\, \Box\\
\end{array}
$$

Theorem-1: Expected number of comparisons done by RandSelect is $\Theta(n)$.

Proof: The number comparisons done by the algorithm in each iteration is $\leq 3\times X_i$. We at most $2$ comparisons in the inner if..else branch and $1$ comparison to test $i <= j$, so a total of $3$ comparisons in the each iteration of the inner most while loop. The loop runs at most $X_i$ times in the $i^{th}$ iteration of the outer most while loop. Finally we need $1$ comparison to test the termination of the while loop. Let $T$ be the random variable corresponding the total number of comparisons in the algorithm. So we need to find $E[T]$.
$$
\begin{array}{ll}
&T \leq 3(X_0 + X_1 + X_2 + \ldots X_{n-1}) + 1 & \text{Total number of comparisons in the Algorithm. }\\
&&\\
\Rightarrow& \sum_{i=0}^{n-1} X_i \leq T \leq 3\sum_{i=0}^{n-1} X_{i} + 1 & \\
& \sum_{i=0}^{n-1} E[X_i] \leq E[T] \leq 3\sum_{i=0}^{n-1} E[X_{i}] + 1 & \text{Linearity of expectation.} \\
&&\\
& \sum_{i=0}^{n-1} \frac{n}{2^i} \leq E[T] \leq 3\sum_{i=0}^{n-1} \frac{n}{2^i} + n & \text{Apply Lemma-2.} \\
& 2n\left(1-\frac{1}{2^n}\right) \leq E[T] \leq 6n\left(1-\frac{1}{2^n}\right) +n & \text{Use: } \sum_{i=0}^{n-1}\frac{1}{2^i} = \frac{1-1/2^n}{1-1/2}.\\
&&\\
\Rightarrow &n \leq E[T] < 6n + 1 = 6n +1 & \text{Use: for } n\geq 1\,\,\,,\frac{1}{2} \leq \left(1-\frac{1}{2^n}\right) < 1 \\ &&\\ \Rightarrow & E[T] = \Theta(n)\,\,\Box & \\ \end{array} $$







We now return the question of the expected number of steps for the problem size to reach unity (i.e. expected number of iterations of the outer most while loop). This is also referred as the expected number of recursive calls in Motwani and Raghavan's book. However the following analysis is much simpler than the application of a calculus based lemma based on random walks in the book. Let $Y_{i}$ be a random variable corresponding to the number of iterations starting with a problem size of $X_{i} > 1$.
$$
\begin{array}{lll}
Y_{i} &:=&\text{random variable for the number of steps with problem size }\,\, X_{i} >1 \\
Y_i &=& 1 + Y_{i+1} \,\,\,\,\,\, \text{(Recursive formulation)}\\
Y = Y_0 &=& \underbrace{1 + 1 + 1 \dots + 1}_{k \text{-times}} + Y_{k} \,\,\,\,\,\, \text{(Expand } k \text{ times)}\\
&& \\
\end{array}
$$

Theorem-2: Expected number of steps for the problem size to reach unity, $E[Y] = \Theta(\log(n))$.

Proof:
$$
\begin{array}{ll}
&Y_i \leq X_{i}\,\,\,\, \text{ (Number of steps are at most problem size at } i^{th} \text{ step which is } X_i)\\
& \\
\Rightarrow &E[Y_i] \leq E[X_i] \,\,\,\, \text{( Let } Z = X_i-Y_i\,\,\,\,,\,\,\,\, E[Z] = E[X_i]-E[Y_i] \geq 0 \text{)}\\
&E[Y_i] \leq E[X_i] = \frac{n}{2^i} \,\,\,\, \text{ (Apply Lemma-2)} \\
&Y = Y_0 = \underbrace{1 + 1 + 1 \dots + 1}_{t \text{-times}} + Y_{t} \\
\Rightarrow & E[Y] = \underbrace{E[1] + E[1] \dots + E[1]}_{t \text{-times}} + E[Y_{t}] = t + E[Y_{t}] \\
&E[Y] = t + E[Y_{t}] \leq t + E[X_{t}] \leq t + \frac{n}{2^t} \\
&\\
\Rightarrow &t \leq E[Y] \leq t + \frac{n}{2^t} \,\,\,\, \text{After } t \text{ steps }\\
& \\
&\text{When } t = \log_2(n) \text{ the above inequality becomes the following } \\
& \log_2(n) \leq E[Y] \leq \log_2(n) + \frac{n}{2^{log_2(n)}} = \log_2(n) +1 \\

\Rightarrow & E[Y] =\Theta(log(n)) \,\, \Box \\
\end{array}
$$




This completes our expected analysis of the RandSelect algorithm. The analysis just uses elementary probability theory without relying on a Calculus based random walk theorem in the "Randomized Algorithms" book by Motwani and Raghavan -- (Theorem-1.3 on page-15).

Thursday, November 24, 2016

Integer Division Algorithm with no Division or Multiplication Operators.

I was working through a number theory problem today and I needed an efficient algorithm to find a largest integer $n$ s.t $a\times n \leq b$ for any given integers $a, b$ (w.l.o.g $a < b$). Clearly $n$ is the quotient when $b$ is divided by $a$. Following is an algorithm which can compute the quotient and reminder in $\Theta(\log(a) + \log(b))$ operations without using multiplication or division operators. In fact if the number of bits in $a$ and $b$ are known then the algorithm runs in $\Theta(\log(a) - \log(b))$ operations.

typedef unsigned int uint;

// Integer division of b with a without using division
// or multiplication operators.
div_t integer_div(uint a, uint b) {
  div_t result;
  uint bit_diff = bit_count(a) - bit_count(b);
  uint prod = 0, prod_shift = b << bit_diff, shift = 1 << bit_diff;

  result.quot = 0;
  do {
    if (prod + prod_shift <= a) {
      prod += prod_shift;
      result.quot += shift;
    }
    shift >>= 1;
    prod_shift >>= 1;
  } while (shift);
  result.rem = a - prod;

  return result;
}

// utility function to find number of bits
// required to represent the value of a.
uint bit_count(uint a) {
  unsigned int i = 0;
  do {
    ++i;
  } while ((a = a >> 1));
  return i;
}

The above algorithm was tested on $4294967293$ pairs of $(a,b), a,b \in [1, INT\_MAX]$ and compared with std::div -- running all these combinations just take around $2$ minutes on a mac book pro. It makes sense because that $\log(a) \leq 32$ (word size on the machine) hence we were just running around four billion constant operations. Typically in all algorithm analysis the word size is considered a $O(1)$, so as long as the inputs bit-widths are within the word size of the machine the runtime of the algorithm can be considered $O(1)$.


Notice that the algorithm is greedy in nature and will provide a correctness next.

Let $c=\sum_{i=0}^{\log(a)-\log(b)}t_i\times 2^{i}$ and $t_{\log(a)-\log(b)}t_{\log(a)-\log(b)-1}\ldots t_0$ the bit representation of $c$. The algorithm tries to pick the powers of $2$ in a greedy manner (i.e. $2^{\log(a)-\log(b)}\geq 2^{\log(a)-\log(b)-1} \geq \ldots 2^{0}$). Consider the choice of $i^{th}$ bit of $c$, if $2^{i}\times a \leq b$ then any $i$-bit integer with a $0$ in the $i^{th}$ bit is strictly less than $2^{i}$. Since $2^{i-1} + 2^{i-2} + \ldots 2^{0} = \frac{2^{i-1+1} -1}{2-1} = 2^{i}-1 < 2^{i}$. We can use induction of $i$ to complete the proof. Hence we can claim that if we greedily pick the powers of $2$ (i.e ${2^i\,\,| \log(a)-\log(b)\leq i \leq 0 }$) it will yield the largest integer $c$ s.t $a\times c \leq b$. The integer division algorithm is just corollary of this basic fact.









Saturday, June 11, 2016

What kind of strings have Kolmogorov complexity equal to their length ?

I have been doing some reading about Information Theory and Kolmogorov complexity lately. I just thought about the following problem. Consider a language of all C programs (strings) which print themselves, formally Let $$L_{self} = \{ w\,\, | w\,\, \mbox{is a C program which prints itself} \}$$.
Following is an example of a C program which prints itself (on a side-note the trick behind the construction is to use the ASCII values for quotes and newlines). Clearly the Kolmogorov complexity of each w is less than or equal to the length of w -- given that we have constructed a program of length same as the string w.
I'm wondering if this is indeed optimal ? -- that is given a w, if its possible to produce a program to generate w and size strictly less than |w|.

char qt=34;
char nl=10;
char *str="char qt=34;%cchar nl=10;%cchar *str=%c%s%c;%cint main() {%c printf(str,nl,nl,qt,str,qt,nl,nl,nl,nl);%c}%c";
int main() {
 printf(str,nl,nl,qt,str,qt,nl,nl,nl,nl);
}

BTW, compile the program with
gcc -w 
to suppress warnings about header inclusion.