Why Mathematica does not simplify sinh(arccosh(x))
johndcook.com105 points by ibobev 4 days ago
105 points by ibobev 4 days ago
This sentence confused me: "For example, Sinh[ArcCosh[-2 + 0.001 I]] returns 11.214 + 2.89845 I but Sinh[ArcCosh[-2 + 0.001 I]] returns 11.214 - 2.89845 I," not the least of which because the two input expressions are the same, but also because we started out by saying Sinh[ArcCosh[-2]] = -Sqrt[3], which is not at all near 11.214 +/- 2.89845 I.
I think the author meant to say, "ArcCosh[-2 + 0.001 I] returns 1.31696 + 3.14102 I but ArcCosh[-2 - 0.001 I] returns 1.31696 - 3.14102 I," because we are talking about defining ArcCosh[] on the branch cut discontinuity, so there is no need to bring Sinh[] into it (and if we do, we find the limits are the same: the imaginary component goes to zero and Sinh[ArcCosh[-2 +/- t*I]] approaches -Sqrt[3] as t goes to zero from above or below). I am not sure what went wrong to get what they wrote.
Thanks. That was a mess. Don't know what happened, but I fixed it this morning.
This is a general pattern in CAS. For a more basic case, it’s not obvious sqrt(square(x)) will simplify to x without any further assumptions on x.
I think you would get sqrt(x^2) = x, if x belonged to the natural domain of sqrt, which is a Riemann surface, that may also be defined using the language of "sheaves". I don't know how to connect this to the article or Mathematica.
That's not what it simplifies to using a real or complex number domains for x, it's abs(x). CAS need type inference assumptions and/or type qualifiers to be more powerful.
Edit: Fixed stuff.
For x = -i, square(x) = -1, sqrt(square(x)) = i. Meanwhile, abs(x) = 1. You're right that it simplifies to abs(x) for real x, but that no longer holds for arbitrary complex values.
for arbitrary complex values sqrt() gives 2 answers with +- signs
so sqrt(square(-i)) = +-i, one of which is x
I've never seen a CAS that gives two answers for sqrt. Mathematica doesn't, sympy doesn't, and IIRC Maxima also doesn't.
Right, that's why you need further assumptions on x in order for that simplification to hold.
It's not a simplification, it's wrong. Sqrt(square(x)) equals abs(x).
It also equals x with appropriate assumptions (x > 0).
so there's an unconditionally correct answer (it's also equal to abs(x) for x>0), and then there is an answer that is only correct for half the domain, which requires an additional assumption.
I really wish Mathematica would open-source the heuristics behind these core functions (including common mathematical functions, Simplify, Integrate, etc.). The documentation is good, but it still lags behind the actual implementation. It would be much easier if we could peek inside the black box.
That blackbox being their entire moat, I would assume they'd never want to open-source any function. Mathematica as a front-end has innumerable frustrating bugs, but its CAS is top-notch. Especially combined with something like Rubi for integration, for me nothing comes close to Mathematica for algebraic computations.
Many functions source ate viewable. Use https://resources.wolframcloud.com/FunctionRepository/resour...
Many built-in functions are open source too. Use the "PrintDefinitions" ResourceFunction to see the code of functions that are implemented in Wolfram Language itself.
Source available? The license is still proprietary, right?
Yes, it is all proprietary, but there are still ways to inspect most of the WL-implemented functions since the system does not go to extreme pains to keep them hidden from introspection. It is not unlike Maple in that sense.
For Simplify, I expect its a black, or at least gray box to Mathematica maintainers, too.
It will have simple rules such as constant folding, “replace x - x by zero”, “replace zero times something with the conditions under which ‘something’ has a value”, etc, lots of more complex but still easy to understand rules with conditionals such as “√x² = |x| if x is real”, and some weird logic that decides the order in which to try simplification rules.
There’s an analogy with compilers. In LLVM, most optimization passes are easy to understand, but if you look at the set of default optimization passes, there’s no clear reason for why it looks like it looks, other than “in our tests, that’s what performed best in a reasonable time frame”.
A lot of problems look like this. A while ago I was working on a calendar event optimization (think optimizing “every Monday from Jan 1, 2026 to March 10, 2026” + “every Monday from March 15, 2026 to March 31, 2026” to simply “every Monday from Jan 1, 2026 to March 31, 2026”). I wrote a number of intuitive and simple optimization passes as well as some unit tests. To my horror, some passes need to be repeated twice in different parts of the pipeline to get the tests to pass.
As a term-rewriting system the rule x-x=0 presumably won’t be in Simplify, it’ll be inside - (or Plus, actually). Instead I’d expect there to be strategies. Pick a strategy using a heuristic, push evaluation as far as it’ll go, pick a strategy, etc. But a lot of the work will be normal evaluation, not Simplify-specific.
More generally it's not at all clear what 'simplify' means.
Is x*x simpler than x^2? Probably? Is sqrt(5)^3 simpler than 5^(3/2)? I don't know.
It entirely depends on what you're going to be doing with the expression later.
In this case, a heuristic like "less parameters, less operators and less function calls" covers all the cases.
I think "simplify" is pretty clear here. For trigonometric functions you would expect a trig function and an inverse trig function to be simplified. We all know what we'd expect if we saw sin(arcsin(x)) (ie x). If we saw cos(arcsin(x)) I'll spoil it for you: it simplifies to sqrt(1-x^2).
Hyperbolic functions aren't used as much but the same principle applies. Here the core identity is cosh^2(x) = sinh^2(x) = 1 so:
sinh(arccosh(x))
= sqrt(1 + cosh^2(arccosh(x))
= sqrt(1 + x^2)
How is going from two functions with one variable to three functions with a variable and a constant a simplification?
If you can't recognize how much simpler the simplified version is, I'm not sure exactly what to tell you. But let's think about it in terms of assembly steps:
1. Multiply the input by itself
2. Add 1
3. Take the square root. There is often a fast square root function available.
The above is a fairly simply sequence of SIMD instructions. You can even do it without SIMD if you want.
Compare this to sinh being (e^x - e^-x) / 2 (you can reduce this to one exponentiation in terms of e^2x but I digress) and arccosh being ln(x + sqrt(s^2 - 1)) and you have an exponentiation, subtraction, division, logarithm, addition, square root and a subtraction. Computers generally implement e^2 and logarithm using numerical method approximations (eg of a Taylor's series expansion).
If simplify means make it fast for a computer to run we might as well make division illegal.
I've only an A-Level in Further Maths from 1997, but understand complex numbers and have come across complex inverse trig functions before.
My takeaway for other people like me from this is "computer is correct" because the proof shows that we can't define arccosh using a single proof across the entire complex plane (specifically imaginary, including infinity).
The representation of this means we have both complex functions that are defined as having coverage of infinity, and arccosh, that a proof exists in only one direction at a time during evaluation.
This distinction is a quirk in mathematics but means that the equation won't be simplified because although it looks like it can, the underlying proof is "one sided" (-ve or +ve) which means the variables are fundamentally not the same at evaluation time unless 2 approaches to the range definition are combined.
The QED is that this distinction won't be shown in the result's representation, leading to the confusion that it should have been simplified.
Simple rule to keep in mind that even math savvy people seem to forget about is that: sqrt(x²) = |x| with bars for absolute value.
For a programmer, it's clear that we have lost the sign information but not the magnitude.
Simple. Makes most sign and solution reasoning explicit instead of implicit when solving quadratics or otherwise working with square roots.
And yet it incorrectly simplifies f(x) = x/x with f(x) = 1
Does anyone else think that the latest LLMs - some of which can be used locally for free - combined with proof-verifying software like Coq or Lean for mistake-detection, might make many uses of Computer Algebra Systems like Mathematica obsolete?
Certainly, people don't need Wolfram Alpha as much.
On another point, it sucks to know what this means for Algebraic Geometry (the computational variant), which you could partly motivate, until now, for its use in constructing CASes.
For me Mathematica is much more akin to numpy+sympy+matplotlib+... with absolutely crazy amount of batteries included in a single coherent package with IDE and fantastic documentation. In a way numpy ecosystem already "won" industry users over, yet Wolfram stack is still appealing to me personally for small experiments.
Coq/Lean target very different use cases.