squiggle.c is a [grug-brained](https://grugbrain.dev/) self-contained C99 library that provides functions for simple Monte Carlo estimation, inspired by [Squiggle](https://www.squiggle-language.com/).
- or more simply, by inlining the sampling functions (adding an `inline` directive before their function declaration)
- **Because there are few abstractions between it and machine code** (C => assembly => machine code with gcc, or C => machine code, with tcc), leading to fewer errors beyond the programmer's control.
You can follow some example usage in the [examples/](examples]) folder. In [examples/core](examples/core/), we build up some functionality, starting from drawing one sample. In [examples/more](examples/more), we present a few more complicated examples, like finding confidence intervals, a model of nuclear war, an estimate of how much exercise to do to lose 10kg, or an example using parallelism.
GCC has an extension which allows a program to define a function inside another function. This makes squiggle.c code more linear and nicer to read, at the cost of becoming dependent on GCC and hence sacrificing portability and increasing compilation times. Conversely, compiling with tcc (tiny c compiler) is almost instantaneous, but leads to longer execution times and doesn't allow for nested functions.
~~My recommendation would be to use tcc while drawing a small number of samples for fast iteration, and then using gcc for the final version with lots of samples, and possibly with nested functions for ease of reading by others.~~
My previous recommendation was to use tcc for marginally faster iteration, but nested functions are just really nice. So my current recommendation is to use gcc throughout, though keep in mind that modifying code to not use nested functions is relatively easy, so keep in mind that you can do that if you run in other environments.
- I offer no guarantees about stability, correctness, performance, etc. I might, for instance, abandon the version in C and rewrite it in Zig, Nim or Rust.
- This project mostly exists for my own usage & for my own amusement.
- Caution! Think carefully before using this project for anything important.
- If you wanted to pay me to provide some stability or correctness, guarantees, or to tweak this library for your own usage, or to teach you how to use it, you could do so [here](https://nunosempere.com/consulting).
- I am conflicted about parallelism. It *does* add more complexity, complexity that you can be bitten by if you are not careful and don't understand it. And this conflicts with the initial grug-brain motivation. At the same time, it is clever, and it is nice, and I like it a lot.
The bailey:
- I've been hacking at this project for a while now, and I think I have a good grasp of its correctness and limitations. I've tried Nim and Zig, and I prefer C so far.
- I think the core interface is not likely to change much stable, though I've been changing the interface for parallelism and for getting confidence intervals.
- I am using this code for a few important consulting projects, and I trust myself to operate it correctly.
- Nonetheless, the user should understand the limitations of sampling-based methods. See the section on [Tests and the long tail of the lognormal](https://git.nunosempere.com/personal/squiggle.c#tests-and-the-long-tail-of-the-lognormal) for a discussion of how sampling is bad at capturing some aspects of distributions with long tails.
- It should be fast. But when speed conflicts with simplicity, choose simplicity. For example, there might be several possible algorithms to sample a distribution, each of which is faster over part of the domain. In that case, it's conceptually simpler to just pick one algorithm, and pay the—normally small—performance penalty. In any case, though, the code should still be *way faster* than, say, Python.
Note that being terse, or avoiding verbosity, is a non-goal. This is in part because of the constraints that C imposes. But it also aids with clarity and conceptual simplicity, as the issue of correlated samples illustrates in the next section.
Should `c` be equal to `2`? or should it be equal to 2 times the expected distribution of the ratio of two independent draws from a (`2 * a/a`, as it were)?
Exercise for the reader: What possible meanings could the following represent in [squiggle](https://www.squiggle-language.com/playground?v=0.8.6#code=eNqrVkpJTUsszSlxzk9JVbJSys3M08jLL8pNzNEw0FEw1NRUUKoFAOYsC1c%3D)? How would you implement each of those meanings in squiggle.c?
Right now, I am drawing samples from a random number generator. It requires some finesse, particularly when using parallelism. But it works fine.
But..., what if we could do something more elegant, more ingenious. In particular, what if instead of drawing samples, we had a mesh of equally spaced points in the range of floats. Then we could, for a given number of samples, better estimate the, say, mean of the distribution we are trying to model...
The problem with that is that if we have some code like:
```C
double model(...){
double a = sample_to(1, 10, i_mesh++);
double b = sample_to(1, 2, i_mesh);
return a * b;
}
```
Then this doesn't work, because the values of a and b will be correlated: when a is high, b will also be high. What might work would be something like this:
But that requires us to encode the shape of the model into the sampling function. It leads to an ugly nesting of for loops. It is a more complex approach. It is not [grug-brained](https://grugbrain.dev/). So every now and then I have to remind myself that this is not the way.
`make verify` is an alias that runs all the tests and just displays the ones that are failing.
These tests are somewhat rudimentary: they get between 1M and 10M samples from a given sampling function, and check that their mean and standard deviations correspond to what they should theoretically should be.
If you run `make run` (or `make verify`), you will see errors such as these:
```
[-] Mean test for normal(47211.047473, 682197.019012) NOT passed.
Mean of normal(47211.047473, 682197.019012): 46933.673278, vs expected mean: 47211.047473
delta: -277.374195, relative delta: -0.005910
[-] Std test for lognormal(4.584666, 2.180816) NOT passed.
Std of lognormal(4.584666, 2.180816): 11443.588861, vs expected std: 11342.434900
delta: 101.153961, relative delta: 0.008839
[-] Std test for to(13839.861856, 897828.354318) NOT passed.
Std of to(13839.861856, 897828.354318): 495123.630575, vs expected std: 498075.002499
delta: -2951.371925, relative delta: -0.005961
```
These tests I wouldn't worry about. Due to luck of the draw, their relative error is a bit over 0.005, or 0.5%, and so the test fails. But it would surprise me if that had some meaningful practical implication.
The errors that should raise some worry are:
```
[-] Mean test for lognormal(1.210013, 4.766882) NOT passed.
Mean of lognormal(1.210013, 4.766882): 342337.257677, vs expected mean: 288253.061628
delta: 54084.196049, relative delta: 0.157985
[-] Std test for lognormal(1.210013, 4.766882) NOT passed.
Std of lognormal(1.210013, 4.766882): 208107782.972184, vs expected std: 24776840217.604111
What is happening in this case is that you are taking a normal, like `normal(-0.195240, 4.883106)`, and you are exponentiating it to arrive at a lognormal. But `normal(-0.195240, 4.883106)` is going to have some non-insignificant weight on, say, 18. But `exp(18) = 39976300`, and points like it are going to end up a nontrivial amount to the analytical mean and standard deviation, even though they have little probability mass.
The reader can also check that for more plausible real-world values, like those fitting a lognormal to a really wide 90% confidence interval from 10 to 10k, errors aren't egregious:
Overall, I would caution that if you really care about the very far tails of distributions, you might want to instead use tools which can do some of the analytical manipulations for you, like the original Squiggle, Simple Squiggle (both linked below), or even doing lognormal multiplication by hand, relying on the fact that two lognormals multiplied together result in another lognormal with known shape.
In fact, squiggle.c does have a few functions for algebraic manipulations of simple distributions at the end of squiggle.c. But these are pretty rudimentary, and I don't know whether I'll end up expanding or deleting them.
This library differentiates between core functions, which are pretty tightly scoped, and expansions and convenience functions, which are more meandering. Expansions are in `squiggle_more.c` and `squiggle_more.h`. To use them, take care to link them:
`squiggle_more.c` has some helper functions to get confidence intervals. They are in `squiggle_more.c` because I'm still mulling over what their shape should be, and because until recently they were pretty limited and sub-optimal. But recently, I've put a bunch of effort into being able to get the confidence interval of an array of samples in O(number of samples), and into making confidence interval functions nicer and more general. So I might promote them to the main `squiggle.c` file.
I provide some functions to draw samples in parallel. For "normal" squiggle.c models, where you define one model and then draw samples from it once at the end, they should be fine.
But for more complicated use cases, my recommendation would be to not use parallelism unless you know what you are doing, because of intricacies around setting seeds. Some gotchas and exercises for the reader:
- If you run the `sampler_parallel` function twice, you will get the same result. Why?
- If you run the `sampler_parallel` function on two different inputs, their outputs will be correlated. E.g., if you run two lognormals, indices which have higher samples in one will tend to have higher samples in the other one. Why?
- For a small amount of samples, if you run the `sampler_parallel` function, you will get better spread out random numbers than if you run things serially. Why?
I provide some auxiliary functions that take a cdf, and return a sample from the distribution produced by that cdf. This might make it easier to program models, at the cost of a 20x to 60x slowdown in the parts of the code that use it.
The process of taking a cdf and returning a sample might fail, e.g., it's a Newton method which might fail to converge because of cdf artifacts. The cdf itself might also fail, e.g., if a distribution only accepts a range of parameters, but is fed parameters outside that range.
The first approach produces terser programs but might not scale. The second approach seems like it could lead to more robust programs, but is more verbose.
Behaviour on error can be toggled by the `EXIT_ON_ERROR` variable. This library also provides a convenient macro, `PROCESS_ERROR`, to make error handling in either case much terser—see the usage in example 4 in the examples/ folder.
Overall, I'd describe the error handling capabilities of this library as pretty rudimentary. For example, this program might fail in surprising ways if you ask for a lognormal with negative standard deviation, because I haven't added error checking for that case yet.
- [ ] Point out that, even though the C standard is ambiguous about this, this code assumes that doubles are 64 bit precision (otherwise the xorshift should be different).
- [ ]~~Think about whether to write a simple version of this for [uxn](https://100r.co/site/uxn.html), a minimalist portable programming stack which, sadly, doesn't have doubles (64 bit floats)~~
