Mentioned in

Convolution as smoothing

2Gurkenglas

1Maxwell Peterson

5Gurkenglas

1Maxwell Peterson

2Donald Hobson

New Comment

and it doesn't much matter if you change the kernel each time

That's counterintuitive. Surely for every there's an that'll get you anywhere? If , .

Yes, that sounds right - such an exists. And expressing it in fourier series makes it clear. So the “not much” in “doesn’t much matter” is doing a lot of work.

I took his meaning as something like "reasonably small changes to the distributions in don’t change the qualitative properties of ". I liked that he pointed it out, because a common version of the CLT stipulates that the random variables must be identically distributed, and I really want readers here to know: No! That isn’t necessary! The distributions can be different (as long as they’re not too different)!

But it sounds like you’re taking it more literally. Hm. Maybe I should edit that part a bit.

The fourier transform as a map between function spaces is continuous, one-to-one and maps gaussians to gaussians, so we can translate "convoluting nice distribution sequences tends towards gaussians" to "multiplying nice function sequences tends towards gaussians". The pointwise logarithm as a map between function spaces is continuous, one-to-one and maps gaussians to parabolas, so we can translate further to "nice function series tend towards parabolae", which sounds more almost always false than usually true.

In the second space, functions are continuous and vanish at infinity.

This doesn't work if the fourier transform of some function is somewhere negative, but then multiplying the function sequence has zeroes.

In the third space, functions are continuous and diverge down at infinity.

Hm. What you're saying sounds reasonable, and is an interesting way to look at it, but then I'm having trouble reconciling it with how widely the central limit theorem applies in practice. Is the difference just that the space of functions is much larger than the space of probability distributions people typically work with? For now I've added an asterisk telling readers to look down here for some caution on the kernel quote.

Suppose you take a bunch of differentiable functions, all of which have a global maximum at 0, and add them pointwise.

Usually you will get a single peak at 0 towering above the rest. The only special case is if In the neighbourhood of 0, the function is approximately parabolic. (Its differentiable.) You take the exponent, this squashes everything but the highest peak down to nearly 0. (In relative terms). The highest peak turns into a sharp spiky Gaussian . You take the inverse Fourier transform and get a shallow Gaussian.

Even if you are unlucky enough to start with several equally high peaks in your 's then you still get something thats kind of a Gaussian. This is the case of a perfectly multimodal distribution, something 0 except on exact multiples of a number. The number of heads in a million coin flips forms a Gaussian out of dirac deltas at the integers.

But the condition of having a maximum at 0 in the Fourier transform is weaker than always being positive. If then

Convolutions smooth out hard sharp things into nice smooth things. (See previous post for why convolutions are important). Here are some convolutions between various distributions:

(For all five examples, all the action in the functions f and g is in positive regions - i.e. f(x) > 0 only when x > 0, and g too - though this isn't necessary.)

Things to notice:

## Understanding convolution as smoothing tells us about the central limit theorem

From Glen_b on stackoverflow:

(a kernel is like a distribution; italics are mine.)

Up till now I've just called it "smoothing", but convolution on most (all?) probability distributions is better described as a smoothing-symmetricking-hillshaping operation. The graphics above are examples. And remember: The central limit theorem is a statement about convolutions on probability distributions! So here's a restatement of the basic central limit theorem:

Or you could look at it the other way around: The central limit theorem describes what comes out on the other side of repeated convolutions. If someone asked you what the behavior of D∗:=d1∗d2∗d3∗⋯∗dn was,

you could answer with the contents of the central limit theorem.Because your answer was the central limit theorem, it would contain information about: how varying the shapes of the component distributions di affected the shape of D∗; how many distributions di you needed to get close to some target distribution H; and what the di needed to look like in order for H to look Gaussian. It would also contain the same kind of information for the moments of H, like the mean. And the variance. It would describe how to use the skew (which corresponds to the third moment) to evaluate how many distributions di you needed for H to look Gaussian (the Berry-Esseen theorem).

These information, together with various additional conditions and bounds, constitute a network of facts about what kinds of probability distributions {d1,...,dn} cause d1∗d2∗⋯∗dn to look Gaussian, and how Gaussian it will look. We call this network of facts "the central limit theorem", or sometimes - to emphasize the different varieties introduced by relaxing or strengthing some of the conditions - the central limit theorems.

*: While this kind of thing is true for many distributions, see Gurkenglas in the comments for why we shouldn't take "it doesn't much matter if you change the kernel each time" too literally.