Let $X$ be a discrete random variable with Laplacian distribution with mean $0$ and scale $\lambda$, as

$$ p(X) = \frac{1}{2\lambda} \exp\left(-\frac{|x|}{2\lambda}\right), \\ X \in \{0,1,...,255\}. $$

Suppose we have $N$ observations $\{x_1,x_2,...,x_N\}$ for $X$ and want to estimate the $\lambda$. One solution is to use Maximum likelihood method which simply estimates $\lambda$ as the mean of $x_i$.

My confusion is about another solution employed in a C++ code. Apparently the implemented approach is to estimate $\lambda$ such that minimizes the conditional entropy. Here is the this code:

You can skip this code. I have represented it as a mathematical expression in the following

double BestLambda, BestEntropy;
BestEntropy = LARGE;
for (double lambda = 0; lambda <= 10.0; lambda += 0.01) {
    // Form Q vector
    double sum = 0.0;
    for (int i = 0; i < 255; i++) {
        pQVec[i] = exp( -lambda * abs(i) );
        sum += pQVec[i];
    //Normalize pQVec
    for(int i = 0; i < 255; i++)
        pQVec[i] = pQVec[i] / sum;

    // Calculate conditional entropy
    double EntropySum = 0.0;
    for (int i = 0; i < 255; i++) {
        if (XFreq[i] > 0) {
            EntropySum -= (XFreq[i] * log(pQVec[i])/log(2.0));

    // Compare against best entropy
    if (EntropySum < BestEntropy) {
        BestEntropy = EntropySum;
        BestLambda = lambda;

In the mathematical terms, this program estimates $\lambda$ in this way,

$$ \DeclareMathOperator*{\argmin}{\arg\!\min} \hat\lambda = \argmin_{\lambda} \left\{ - \sum\limits_{i=1}^{255} f(x=i)\log_2(p_{\lambda}(x=i)) \right\}, $$

where $f(x=i)$ is the frequency of $x$'s with value $i$ in the samples and $p_\lambda(x)$ is a Laplacian distribution with mean 0 and scale $\lambda$.

While this code is implemented and published by the guys at Stanford and supported by their highly-cited paper, I have doubt if their approach in estimating $\lambda$ is accurate. I have studied a little bit on Information Theory and cannot figure out what we are actually minimizing in this expression.

Shouldn't we minimize the distance between two distributions? Using Kullback–Leibler divergence we could estimate $\lambda$ as

$$ \hat\lambda = \argmin_{\lambda} \left\{ D(f(x)||p_{\lambda}(x) ) \right\} = \argmin_{\lambda} \left\{ \sum_{i=1}^{255} \left[ f(x=i)\log_2\left(\frac{f(x=i)}{p_{\lambda}(x=i)}\right) \right] \right\} . $$


First, for the shake of exactness, a discrete random variable cannot have a Laplace distribution which is a continuous distribution for continuous rv's. If your variable is by nature discrete, then you should look to the discrete analogue of the Laplace distribution. But I assume that you have a continuous rv but naturally only a finite discrete sample.

Then, regarding your question, what the code does is exactly minimizing KL distance. Look at the argmin operation you propose: you have the logarithm of a ratio - so it can be additively broken into a positive part (which does not depend on lambda and so is irrelevant for the minimization procedure), and a negative part, which depends on lambda and which is exactly what the code attempts to minimize: you won't get the value of the minimized distance/divergence, but you will get the lambda that minimizes it.

  • $\begingroup$ Wow thanks! Didn't expected to be that simple. And yes, the random variable is continuous and we have samples from a finite domain. $\endgroup$ – Isaac Jul 24 '13 at 9:25

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.