I run a model (MTK) where one term of an equation does log(x^0.5). This seems to work sometimes, but also crashes for some parameter values when I do Monte Carlo simulation. If I change the term to log(sqrt(x)), the simulation seems to always work. Question: Is there a reason for the problem with x…

any reason you aren’t writing this as 0.5*log(x)? I’d expect that to be notably faster

I could, but the logarithm contains other multiplicative terms, too. But I could write it as a sum of logarithms.

in general, that will be a lot less likely to suffer from overflow and underflow (and depending on how many times you need to call log, it will likely be faster as well).

OK - I’ll try to do that. It still puzzles me that the simulation crashes when I do log(x^0.5), but not when I do log(sqrt(x)). Will do it tomorrow… I left my job PC with the VSCode file open, and it messes up the file slightly if I open it form other computers while still open at work.

[image] BLI: I run a model (MTK) where one term of an equation does log(x^0.5). This seems to work sometimes, but also crashes for some parameter values when I do Monte Carlo simulation. What do you mean by “crash”? Do you mean it throws an exception? What is the exact error message? (MTK …

The solver seems to try to take the square root of a negative number… I do run a Monte Carlo simulation where I let some dynamic model parameters change \pm 10% from a uniform distribution. I’m pretty sure these variations are too small to make the system unstable. And the “instability”/square root …

[image] BLI: Regarding efficiency, will the first expression (\log(x^a\cdot y^b)log(xa⋅yb)\log(x^a\cdot y^b)) always be slower than the third expression (a\cdot \log(x) + b\cdot \log(y)a⋅log(x)+b⋅log(y)a\cdot \log(x) + b\cdot \log(y))? Generally speaking, yes. Summation and multiplication are…

In general x^y is computed as exp(y*log(x)) (except you actually need to use a slower version of the logarithm that is extra accurate to prevent roundoff error. sqrt is fairly fast (about as fast as division so about 10x worse than multiplication/addition), but currently the compiler isn’t able to o…

MTK - equation with sqrt(x) vs. x^0.5?

Specific Domains Modelling & Simulations

Topic		Replies	Views
Fastest `sqrt` and `log` with negative check Performance question , ccall , llvm , math	17	1225	January 14, 2024
Taking power with float exponent x^y is slower than exp(y*log(x))? Performance	8	423	September 11, 2025
Most accurate way to compute x/sqrt(y) in Julia? Performance precision	11	1045	December 1, 2020
Sum of logs Performance	19	1171	May 10, 2022
Loglogistic function robustness in Turing.jl General Usage turing , math	4	597	February 10, 2021

MTK - equation with sqrt(x) vs. x^0.5?

Related topics