Posts tagged with Gaussian

Fun coursera on virology.

• Viruses are so numerous (10³⁰) and filling up everywhere. It gives this Boltzmann flavour of ‘enough stuff” to really do statistics on.

• Viruses are just a bundle of `{proteins, lipids, nucleic acids}` with a shell. It’s totally value-free, no social Darwinism or “survival of the fittest” being imbued with a moral colour. Just a thing that happened that can replicate.
• Maybe this is just because I was reading about nuclear spaces (⊂ topological vector spaceand white-noise processes that I think of this. Viruses have a qualitatively different error structure than Gaussian. Instead of white-noise it’s about if they can get past certain barriers, like:
• survive out in the air/water/cyanide
• bind to a DNA
• adapt to the host’s defences
• … it seems like a mathematician or probabilist could use the viral world of errors to set out different assumptions for a mathematical object that would capture the broad features of this world that’s full of really tiny things but very different to gas particles.
• Did I mention that I love how viral evolution is totally value-neutral and logic-based?
• Did I mention how I love that these things are everywhere all the time, filling up the great microspace my knowledge had left empty between man > animals > plants > > bacteria > > minerals?

### going the long way

What does it mean when mathematicians talk about a bijection or homomorphism?

Imagine you want to get from `X` to `X′` but you don’t know how. Then you find a "different way of looking at the same thing" using ƒ. (Map the stuff with ƒ to another space `Y`, then do something else over in `image ƒ`, then take a journey over there, and then return back with ƒ ⁻¹.)

The fact that a bijection can show you something in a new way that suddenly makes the answer to the question so obvious, is the basis of the jokes on www.theproofistrivial.com.

In a given category the homomorphisms `Hom` ∋ ƒ preserve all the interesting properties. Linear maps, for example (except when `det=0`) barely change anything—like if your government suddenly added another zero to the end of all currency denominations, just a rescaling—so they preserve most interesting properties and therefore any linear mapping to another domain could be inverted back so anything you discover over in the new domain (`image of ƒ`) can be used on the original problem.

All of these fancy-sounding maps are linear:

They sound fancy because whilst they leave things technically equivalent in an objective sense, the result looks very different to people. So then we get to use intuition or insight that only works in say the spectral domain, and still technically be working on the same original problem.

Pipe the problem somewhere else, look at it from another angle, solve it there, unpipe your answer back to the original viewpoint/space.

` `

For example: the Gaussian (normal) cumulative distribution function is monotone, hence injective (one-to-one), hence invertible.

By contrast the Gaussian probability distribution function (the “default” way of looking at a “normal Bell Curve”) fails the horizontal line test, hence is many-to-one, hence cannot be totally inverted.

So in this case, integrating once `∫[pdf] = cdf` made the function “mathematically nicer” without changing its interesting qualities or altering its inherent nature.

` `

Or here’s an example from calc 101: u-substitution. You’re essentially saying “Instead of solving this integral, how about if I solve a different one which is exactly equivalent?” The `→ƒ` in the top diagram is the u-substitution itself. The “main verb” is doing the integral. U-substituters avoid doing the hard integral, go the long way, and end up doing something much easier.

$\dpi{200} \bg_white \large \text{Problem: integrate } \int {8x^7 - 6x^2 \over x^8 - 2x^3 + 13587} \ \mathrm{d}x \\ \\ \rule{13cm}{0.4pt} \\ \\ \text{\textsc{Clever person:} \textit{How about instead I integrate} } \int {1 \over u} \ \mathrm{d}u \text{ \textit{?}} \\ \\ \\ \text{\textsc{Question asker:} \textit{Huh?}} \\ \\ \\ \text{\textsc{Clever person:} \textit{They're equivalent, you see? Watch!} } \\ \\ \text{\small{(applies basis isomorphism }} \phi: x \mapsto u \\ \text{\small{ as well as chain rule for }} \mathrm{d} \circ \phi: \mathrm{d}x \mapsto \mathrm{d}u \text{\small{)}} \\ \\ \text{ \small{(gets easier integral)}} \\ \\ \text{ \small{(does easier integral)}} \\ \\ \text{ \small{(laughs)}} \\ \\ \text{ \small{(transforms it back }} \phi^{-1}: u \mapsto x \text{\small{)}} \\ \\ \text{ \small{(laughs again)}} \\ \\ \text{\textsc{Question asker:} \textit{Um.}} \\ \\ \text{ \small{(thinks)}} \\ \\ \text{ \textit{Unbelievable. That worked. You must be some kind of clever person.}}$

` `

Or in physics—like tensors and Schrödinger solving and stuff.

Physicists look for substitutions that make the computation they have to do more tractable. Try solving a Schrödinger PDE for hydrogen’s first electron `s¹`in `xyz` coordinates (square grid)—then try solving it in spherical coordinates (longitude & latitude on expanding shells). Since the natural symmetry of the `s¹` orbital is spherical, changing basis to polar coords makes life much easier.

` `

Likewise one of the goals of tensor analysis is to not be tied to any particular basis—so long as the basis doesn’t trip over itself, you should be free to switch between bases to get different jobs done. Terry Tao talks about something like this under the keyword “spending symmetry”—if you use up your basis isomorphism, you need to give it back before you can use it again.

"Going the long way" can be easier than trying to solve a problem directly.

## Climate Statistics

• httpness: (studying statistics) Can there be a different standard deviation up and down?
• isomorphisms: Yes. it's called a semideviation. (Or a quasinorm.) There are a lot of people who argue that semideviations and quasinorms are more natural than standard deviation and norms.
• httpness: So that's not a normal distribution?
• isomorphisms: Whatever distribution you're using, there are different measures of dispersion on that -- standard deviation, downside risk / semideviation, interquartile range, kurtosis, etc.
• httpness: I was just thinking about temperatures. The standard deviation changes depending on the time of year, and the chance of unseasonably warm or cold days changes too.
• httpness: Here's an example of what I mean. let's say during the summer there _is_ a standard deviation and it's the same up and down. But at another time of year there could be more chance of a very warm day, and at a third time of year there could be more chance of an unseasonably cold day.

`cumsum ( rnorm(50), lend="butt", lwd=12, type="h" )`

Cumulative sum of 50 draws from a normal distribution.

File this under mysteries of the Central Limit Theorem.

hi-res

## √π sqrt[pi]

π of course is the distance around a circle. √π is the area under ∫exp (−x²), and exp (−x²) is the key ingredient in the normal distribution.

$\dpi{300} \bg_white \int e^{-x^2} = \sqrt{\pi}$

That’s more or less what √π means—the area under the Bell curve.

But what does it mean mean? I mean, if π is a distance and is used to turn areas into distances — is it, like, shrinking the π even one more time? Are we talking about a half-dimension here?

edit: hmm, the end of this post seems to have been deleted by the rare weirdness of tumblr’s mass editor. I’ll see if I can’t remember how it ended. Umm, something about the moment-generating function? (i.e. going around the complex unit circle)

Which of these pictures come from a random normal distribution and which come from a mixed distribution?

` plot(rnorm,-3,3); mix <- function(x) { rnorm(x)+rnorm(x-3) } plot(mix(x), -3,3); `