You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 
NunoSempere 9ae5e09387
final formatting tweaks
2 months ago
choose add choose pkg 2 months ago
README.md final formatting tweaks 2 months ago
go.mod initialize hello world 2 months ago
makefile tweak: add some indicative printfs 2 months ago
probppl update with results 2 months ago
probppl.go final formatting tweaks 2 months ago
record.txt update with results 2 months ago

README.md

A Bayesian Nerd-Snipe

Consider the number of people you know who share your birthday. This seems an unbiased estimate of the number of people who, if they had been born the same day of the year as you, you'd know—just multiply by 365. That estimate itself is an estimate of how many people one knows at a somewhat non-superficial level of familiarity.

I asked my Twitter followers that question, and this is what they answered:

How many people do you know that were born in the same day of the year as you?

— Nuño Sempere (@NunoSempere) February 21, 2024

Now, and here comes the nerd snipe: after seeing the results of that poll, what should my posterior estimate be for the distribution of how many people my pool of followers knows enough that they'd know their birthdays if they fell on the same day as one's own?

Here is a photo of two cats for those of my readers who don't want to be spoiled and want to sketch the solution before reading on.

The formal solution.

Consider distributions over how many people someone knows. Those distributions go from the natural numbers to a probability.

One example such distribution might be

| Number of people known | Probability | | 10 | 5% | | 21 | 96% | | 1001 | 1% |

Now, consider the likelihood of getting the Twitter poll results given a given distribution. Multiply that by the prior for that distribution, normalize, and then integrate over distributions to get your final result.

The practical solution

The above is computationally intractable, so we turn to Monte Carlo approximations and other shortcuts. After tinkering for a bit, I ended considering distributions over logspace, and considering only number of people in the set: 16, 32, 64, 128, 128, 256, 512, 1024, 2046.

You can see the code here. It's written in go because I've been recently been learning its syntax, and it's only moderately slower than C in exchange for a nicer developer experience.

With that code, the posterior over the number of people my followers know stands as:

Num people known %
~16 7.2%
~32 8.6%
~64 11.0%
~128 15.7%
~256 21.9%
~512 22.6%
~1024 9.5%
~2048 3.4%

To do list

  • MVP of mappings
  • Try as cdfs: too much prob at the beginning
  • Try as pdfs: too evenly distributed
  • Try as pdfs over log space; 1, 2, 4, 8, 16, 32, ...
    • Or maybe even more coarse, 1.x^n
  • Write blogpost
  • Run model for ~half an hour
  • Post blogpost