Hacker News Reader

Thanks for sharing the name of that problem! I've encountered it before while optimizing batched LLM inference. The whole batch would last until all queries in a batch were done, and by changing the batch size, you'd trade off per-query-speed (better in a larger batch) with overall performance (worse with a larger batch).

Nowadays I think this is solved in an entirely different way, though.

It gets more entertaining.

It's common to wrap API calls with

retry on failure, or

spawn an identical request if taking longer than x,or

recursively spawn an identical request if taking longer than x,or

retry on failure but no more than k times.

All of these and similar patterns/decorators can be analysed using the same idea.

Oh wow, pretty cool stuff! If you have more to share, you can always dump it in my mail inbox

You should be careful with your estimation. The events should be independent to apply those properties but it is very common that one cause can influence many factors, so they are not independent and all the beauty math does not work as with independence. In the worst day, all fail, because one resource can block others and the system get strangled. There is the black swan book when one rare event make the financial market realize what is risk.

A load-balancer transparently sitting in front of the api end-point (not an uncommon scenario) usually decouples things well enough to be practically independent.

That said, silloncito's warning does need to be paid heed.

While independence is essential for the proof to go through, the relationships need not break catastrophically with break of independence, usually it is graceful degradation with degree of independence. There are however specific, often degenerate, theoretical edge cases where the degradation is rapid.

Almost every probability theorem starts with "let's take independent random variables". But in reality almost nothing is independent. Superdeterminism even claims that exactly nothing is independent.

You are right.

The degree of dependence matters though. Mutual information is one way to measure that.

Thankfully, some theorems remain valid even when independence is violated. The next stop after independence is martingale criteria. Martingale difference sequences can be quite strongly dependent yet allow some of the usual theorems to go through but with worse convergence rates.

Almost every elementary probability theorem. There are plenty of theorems conditional on bounds on how weak the dependency should be (or how fast correlations decrease with order, etc.), including CLT-like theorems, it’s just that they are difficult to state, very difficult to prove, and almost impossible to verify the applicability of. In practice you are anyway going to instead use the simpler version and check if the results make sense after the fact.

What ? I've never heard of this math. Do you mean literally those are the resulting operations in the general case or are those approximate explanation and we need to find more specific cases to make this true ?

The math is general and exact.

Max and Plus at the random variables space becomes product and convolution in their distribution function space.

    Distr(X+Y) =  DistrX ° DistrY

    Distr (X ^ Y) = DistrX * DistrY. 

    Where '^' denotes Max and '°' denotes convolution.

Note *, +, ° and ^ being commutative and associative they can be chained. One can also use their distributive properties. This really the math of groups and rings.

However, one can and one does resort to approximations to compute the desired end results.

More specifically, people are often interested not in the distribution, but some statistics. For example, mean, standard deviation, some tail percentile etc. To compute those stats from the exact distributions, approximations can be employed.

Surely this isn't quite right.

Max of variables = product of cumulative distribution functions.

Sum of variables = convolution of probability density functions.

So both of the equations you write down are correct, but only if you interpret "Distr" as meaning different things in the two cases.

[EDITED to add:] Provided the random variables in question are independent, as mentioned elsewhere in the discussion; if they aren't then none of this works.

The original post, to which I replied, is about the correspondence between summation of random variables and convolution of their distribution. Independence is sufficient for that.

I just carried through that assumption of independence in my own comment, thinking it was obvious to do that (carry over the assumptions).

But is he right about the different meanings of Distr in your equations ?

No he is not.

Both the ideas work for the cumulative distribution function that is called just distribution function in math. I think he got confused by the fact that convolution relation works also with densities (so he might have assumed that it works with densities only and not distributions)

I'm sorry, but I think you are just wrong about convolutions and cumulative distribution functions.

Let's take the simplest possible case: a "random" variable that's always equal to 0. Its cdf is a step function: 0 for negative values, 1 for positive values. (Use whatever convention you prefer for the value at 0.)

The sum of two such random variables is another with the same distribution, of course. So, what's the convolution of the cdfs?

Answer: it's not even well defined.

The convolution of functions f and g is a function such that h(x) = integral over t of f(t) g(x-t) dt. The integral is over the whole of (in this case) the real numbers.

In this case f and g are both step functions as described above, so (using the convenient Iverson bracket notation for indicator functions) this is the integral over t of [t>0] [x-t>0] dt, i.e., of [0<t<x] dt, whose value is 0 for negative x and x for positive x.

This is not the cdf of any probability distribution since it doesn't -> 1 as x -> oo. In particular, it isn't the cdf of the right probability distribution which, as mentioned above, would be the same step function as f and g.

If X,Y are independent with pdfs f,g and cdfs F,G then the cdf of X+Y is (not F conv G but) f conv G = F conv g.

Oops, one thing in the above is completely wrong (I wrote it before thinking things through carefully, and then forgot to delete it).

It is not at all true that "it's not even well defined", and indeed the following couple of paragraphs determine exactly what the thing in question is. It's not an actual cdf, but the problem isn't that it's ill-defined but that it's well-defined but has the wrong shape to be a cdf.

$Let M=Max(X,Y)$. If $X$ and $Y$ are independent then: $F_M(k) = P(M \leq K) = P((X \leq K) and (Y \leq K))$, so that

  $P(X \leq K) x P(Y \leq K) = F_X(K) x F_Y(K)$.

 So $F_M = F_X \times F_Y$

Thanks for making the assumption of independence explicit and welcome to HN.

Thank you for your welcome, I must have been lurking here for around 30 years or more (always changing accounts). Anyway in this specific case, since M = Max(X,X) = X you can't have F(M) = F(X)*F(X) = F(X) except when F(X) in {0,1}, so the independence property is essential. Welcome fellow Lisper (for the txr and related submission) and math inspired (this one and another related to statistical estimation) with OS related interest (your HN account), OS are not my cup of tea but awk is not bad).

  In another post there are some comments between topology and deep learning. I wonder if there is a definition similar to dimension in topology which would allow you to estimate the minimal size (number of parameters) in a neural network so that is able to achieve a certain state (for example obtaining the capacity to one shot learning with high probability).

Yes independence is absolutely an assumption that I (implicitly) made. It's essential for the convolution identity to hold as well, I just carried through that assumption.

We share interest in AWK (*) then :) I don't know OS at all. Did you imply I know lisp ? I enjoy scheme, but used it in anger never. Big fan of the little schemer series of books.

(*) Have to find that Weinberger face Google-NY t-shirt. Little treasures.

Regarding your dimensions comment, this is well understood for a single layer, that is, for logistic regression. Lehmann's book will have the necessary material. With multiple layers it gets complicated real fast.

The best performance estimates, as in, within realms of being practically useful, largely come from two approaches, one from PAC-Bayesian bounds, the other from Statistical Physics (but these bounds are data distribution dependent). The intrinsic dimension of the data plays a fundamental role there.

The recommended place to dig around is JMLR (journal of machine learning research).

Perhaps your txr submission suggests a lisp flavor. The intrinsic dimension concept looks interesting, also the V.C. dimension, but both concepts are very general. Perhaps Lehmann's book is: Elements of large sample theory.

Txr is super interesting.

I meant Lehmann's Theory of Point Estimation, but large sample theory is a good book too. The newer editions of TPE are a tad hefty in number of pages. The earlier versions would serve you fine.

The generic idea is that smaller these dimensions, easier the prediction problem. Intrinsic dimension is one that comes closest to topology. VC is very combinatorial and gives the worst of worst case bounds. For a typical sized dataset one ends up with an error probability estimate of less than 420. With PAC-Bayes the bounds are atleast less than 1.0.

Check out tropical algebras

Fascinating thank you for reminding me about that

That doesn't sound right. If P(X) is the vector {0.5,0,0.5} and P(Y) is {0.5,0.5,0}, P(X)P(Y) is {0.25,0,0} and that's both not normalized and clearly not the distribution for max(X,Y). Did you get that from an LLM?

You are using PMFs. I meant and wrote distribution function aka cumulative distribution function. They are closed under products.

> Did you get it from LLM

LOL. There must be a fun and guilty story lurking inside the accusation.

On a more serious note, I would love it if LLMs could do such simplifications and estimations on their own.

Distributions can be either PDFs or CDFs. To be honest I'd never heard of assuming that a distribution was a CDF unless otherwise specified.

May I raise you a

https://en.m.wikipedia.org/wiki/Distribution_function

It's right in the title.

In probability theory, integration theory, as well as electrical engineering, "distribution function", unless further clarified, means that cumulative thing.

In math, nomenclature overloading can be a problem. So context matters. In the context of dirac delta, distribution means something else entirely -- generalized functions.

Oh! So sad you deleted your point about densities. One can only laugh at and enjoy these idiosyncrasies of nomenclature.

In Electrical Engineering one uses j for imaginary numbers because i is taken (by current).

This is a natural point of confusion. The true (IMO) primitive concept here is the probability measure. Probability measures on the real line are in canonical bijection with CDFs, the latter being axiomatizable as càdlàg functions (see https://en.wikipedia.org/wiki/Càdlàg) asymptotic to 0 (resp. 1) at minus infinity (resp. infinity). On the other hand, not every probability measure has a density function. (If you want the formalism of densities to capture all probability measures, you need to admit more exotic generalized functions à la Dirac.)