mastodon.world is one of the many independent Mastodon servers you can use to participate in the fediverse.
Generic Mastodon server for anyone to use.

Server stats:

8.1K
active users

#probability

0 posts0 participants0 posts today

My treatise on dice probability in TTRPGs is really coming along! Working out the probabilities for contested rolls was…way too difficult (thank you Interlock) but it’s mostly there. No idea how I’m going to fit in bonus dice with modifiers but that’s a problem for another day.

My previous attempt to calculate how fast sums of the Pareto and Student's-t for parameter = 2 converge to a normal distribution failed due to both a bad fitting formula and accuracy issues. Now I am simulating the summation of up to 1e4930 random variables before something goes awry. Convergence is sloooooooow.

#probability #statistics #GCLT

skewray.com/articles/the-gener

www.skewray.comThe Generalized Central Limit Theorem for α=2: Pareto and Student's-t | Skewray ResearchWhat happens when the central limit theorem convergence at the exact parameter where the variance becomes infinite? Extremely slow convergence.

Dear colleagues working with Markov-chain Monte Carlo: could you share any works that explore Markov-chain "convergence", and precision of mean estimates, with methods that use *quantiles* (or interquartile range, or median absolute deviation, or similar), rather than standard deviation and similar quantities?

Just to be clear, I don't mean estimation of quantiles, but estimation *by means of* quantiles.

Thank you!

Continued thread

This is an example of amortized complexity and Strassen's "asymptotic spectra" (nice monograph by Zuiddam & Wigderson: math.ias.edu/~avi/PUBLICATIONS)

Strassen developed this to understand the #complexity of matrix multiplication and #tensors, but it turns out to also show up in a bunch of places:
- #Entropy
- #Quantum information
- Shannon capacity of graphs
- Communication complexity en.wikipedia.org/wiki/Communic
- Circuit complexity (Robere & Zuiddam eccc.weizmann.ac.il/report/202)

Real coin flips are ~49-51 not 50-50 scientificamerican.com/article

But you can guarantee equal probability with a simple trick! Flip 2x in a row starting with the same side up.

HT->call it H
TH->call it T
HH,TT->try again

(due to von Neumann en.wikipedia.org/wiki/Randomne)

This leads to randomness extractors: from a given random process, what's the biggest uniform distribution you can get efficiently?

Randomness extractors give another interpretation of #entropy:

avg # bits needed to *describe* the outcome
=
# uniformly random bits you can *extract* from the outcome

Illustration of a hand flipping a coin.
Scientific American · Scientists Destroy Illusion That Coin Toss Flips Are 50–50By Shi En Kim

In grad school I noticed a printed message on the wall of our research lab. It was something like "distill information from the hint of implication" (bad memory).

I innocently said to an older grad student, "Yeah, that does sound like a way to commit Type-I errors" or something equivalent.

Well, that's not what the older grad student took from the sign and I got a very cold, slightly huffy response.

Replied in thread

@androcat
It applies to all branches of science. For example statistics: there's no such thing as "letting the data speak for themselves"; they're always interpreted against a background of assumptions and prior knowledge.

I spent two weeks trying to numerically demonstrate convergence of the generalized central limit theorem for the critical case where the variance just becomes infinite. I failed! I suspect my choice of norm was a poor one.

Modeling the sum of 10^30 random variables is challenging.

#probability #statistics

skewray.com/articles/the-gener

www.skewray.comThe Generalized Central Limit Theorem for α=2: Pareto and Student's-t | Skewray ResearchWhat happens when the central limit theorem convergence at the exact parameter where the variance becomes infinite? Extremely slow convergence.