This question already has an answer here:

I have a question about the Kullback-Leibler divergence.

Can someone explain why the "distance" between the blue density and the "red" density is smaller than the distance between the "green" curve and the "red" one?

Graph of three pdfs


marked as duplicate by kjetil b halvorsen, Peter Flom Apr 12 '17 at 13:30

This question has been asked before and already has an answer. If those answers do not fully address your question, please ask a new question.

  • $\begingroup$ I don't have the answer... but I am not sure that, in this context, it does really make sense to consider an inverse gaussian distribution with mean 1 and variance 3 or 5... $\endgroup$ – user3016 Feb 2 '11 at 16:14
  • 5
    $\begingroup$ Careful! The KL divergence is not a true "distance" because it is asymmetric. In each case, which of the two possible values have you computed? $\endgroup$ – whuber Feb 2 '11 at 17:21
  • 1
    $\begingroup$ What distributions are these? Both the Gamma and the Inverse Gaussian take two parameters. The red one clearly is not a Gamma with a shape parameter of 0.85. Through trial and error it looks like the Gamma has a scale of 1 and shape of 1/0.85 while the Inverse Gaussians have means of 1 and scale parameters as given. Is this correct? $\endgroup$ – whuber Feb 2 '11 at 17:39
  • 1
    $\begingroup$ Information-theoretic interpretation of KL divergence is in answer here -- stats.stackexchange.com/questions/1028/… $\endgroup$ – Yaroslav Bulatov Feb 3 '11 at 3:21
  • 2
    $\begingroup$ Also, KL-divergence is not symmetric, so to remove ambiguity it is better to say KL(A,B) or KL(B,A) instead of "distance between A and B" $\endgroup$ – Yaroslav Bulatov Feb 3 '11 at 3:28

Because I compute slightly different values of the KL divergence than reported here, let's start with my attempt at reproducing the graphs of these PDFs:

PDFs for red, blue, and green

The KL distance from $F$ to $G$ is the expectation, under the probability law $F$, of the difference in logarithms of their PDFs. Let us therefore look closely at the log PDFs. The values near 0 matter a lot, so let's examine them. The next figure plots the log PDFs in the region from $x=0$ to $x=0.10$:

Log PDFs in the interval from 0 to 0.10

Mathematica computes that KL(red, blue) = 0.574461 and KL(red, green) = 0.641924. In the graph it is clear that between 0 and 0.02, approximately, log(green) differs far more from log(red) than does log(blue). Moreover, in this range there is still substantially large probability density for red: its logarithm is greater than -1 (so the density is greater than about 1/2).

Take a look at the differences in logarithms. Now the blue curve is the difference log(red) - log(blue) and the green curve is log(red) - log(green). The KL divergences (w.r.t. red) are the expectations (according to the red pdf) of these functions.

Log PDFs for red:blue and red:green, over the interval from 0 to 0.04

(Note the change in horizontal scale, which now focuses more closely near 0.)

Very roughly, it looks like a typical vertical distance between these curves is around 10 over the interval from 0 to 0.02, while a typical value for the red pdf is about 1/2. Thus, this interval alone should add about 10 * 0.02 /2 = 0.1 to the KL divergences. This just about explains the difference of .067. Yes, it's true that the blue logarithms are further away than the green logs for larger horizontal values, but the differences are not as extreme and the red PDF decays quickly.

In brief, extreme differences in the left tails of the blue and green distributions, for values between 0 and 0.02, explain why KL(red, green) exceeds KL(red, blue).

Incidentally, KL(blue, red) = 0.454776 and KL(green, red) = 0.254469.


Specify the distributions

red = GammaDistribution[1/.85, 1];
green = InverseGaussianDistribution[1, 1/3.];
blue = InverseGaussianDistribution[1, 1/5.];

Compute KL

(* Numeric integation between specified endpoints. *)
kl[pF_, qF_, l_, u_] := Module[{p, q},
   p[x_] := PDF[pF, x];
   q[x_] := PDF[qF, x];
   NIntegrate[p[x] (Log[p[x]] - Log[q[x]]), {x, l, u},  
    Method -> "LocalAdaptive"]
(* Integration over the entire domain. *)
kl[pF_, qF_] := Module[{p, q},
   p[x_] := PDF[pF, x];
   q[x_] := PDF[qF, x];
   Integrate[p[x] (Log[p[x]] - Log[q[x]]), {x, 0, \[Infinity]}]

kl[red, blue]
kl[red, green]
kl[blue, red, 0, \[Infinity]]
kl[green, red, 0, \[Infinity]]

Make the plots

plot[{f_, u_, r_}] := 
  Plot[Evaluate[f[#, x] & /@ {blue, red, green}], {x, 0, u}, 
   PlotStyle -> {{Thick, Darker[Blue]}, {Thick, Darker[Red]}, 
     {Thick, Darker[Green]}},
   PlotRange -> r,
   Exclusions -> {0},
   ImageSize -> 400
  plot[f], {f, {{PDF, 4, {Full, {0, 3}}}, {Log[PDF[##]] &, 
     0.1, {Full, Automatic}}}}
  ] // TableForm

Plot[{Log[PDF[red, x]] - Log[PDF[blue, x]], 
  Log[PDF[red, x]] - Log[PDF[green, x]]}, {x, 0, 0.04}, 
 PlotRange -> {Full, Automatic}, 
 PlotStyle -> {{Thick, Darker[Blue]}, {Thick, Darker[Green]}}]
  • $\begingroup$ can you give link to Mathematica source? $\endgroup$ – Yaroslav Bulatov Feb 3 '11 at 3:24
  • $\begingroup$ @Yaroslav I added it to the end. $\endgroup$ – whuber Feb 3 '11 at 5:05
  • 1
    $\begingroup$ @Whuber : Wahou ! Thank you very much. I am going to prepare a coffee and then I focus on your answer ! $\endgroup$ – user3016 Feb 3 '11 at 6:59
  • $\begingroup$ @whuber +1, excellent and detailed answer as always :) $\endgroup$ – mpiktas Feb 3 '11 at 7:56
  • 3
    $\begingroup$ @Marco Yes. It's a good idea also to plot the pdf for the density against which you are integrating (which you had already done for this question). You might find it helpful to read the section on "Graphical Moments" at quantdec.com/envstats/notes/class_06/properties.htm . $\endgroup$ – whuber Feb 3 '11 at 14:09

KL divergence measures how difficult it is to fake one distribution with another one. Assume that you draw an i.i.d. sample of size $n$ from the red distribution and that $n$ is large. It may happen that the empirical distribution of this sample mimicks the blue distribution. This is rare but this may happen... albeit with a probability which is vanishingly small, and which behaves like $\mathrm{e}^{-nH}$. The exponent $H$ is the KL divergence of the blue distribution with respect to the red one.

Having said that, I wonder why your KL divergences are ranked in the order you say they are.

  • 7
    $\begingroup$ +1. Maybe the ranking depends on the order in which H was computed. E.g., one could be KL(green, red) and the other could be KL(red, blue). Morever, tail behavior can have a profound effect on the value: what we really need to see are plots of the logarithms of the densities, not the densities themselves. $\endgroup$ – whuber Feb 2 '11 at 17:32