Incorrect convex hull for nearly collinear points about geo HOT 4 CLOSED

Yes, generally, floating-point numbers has some caveats when converting between display and internal representation.

For instance, 0.1_f64 + 0.2 == 0.3 evaluates to false. This is because, 0.1 is not representible in f64 (or f32), and it's approximated. Now, the degree of this approximation depends on the float type, and in your input I believe the extra point is approximated to fall outside the "true conv hull" when using f32.

If you want to use fixed point/precision, please scale and convert to integers and run the hull (the crate supports hull on i32 or i64 points).

from geo.

Can you try with f64 precision? The points may be too large for f32 to correctly work. You can also try subtracting out the first point, before calling the algo to see if that fixes it.

from geo.

Thanks for the suggestion, f64 works correctly. Interestingly, 32-bit precision still fails even after subtracting the points so that the leftmost point is at the origin. Not the first time I've encountered floating point issues. Hopefully I won't get an edge case where 64-bit fails too, otherwise I could consider a rational/fraction type instead of floats

from geo.

Closing; the output is correct for the given input (and data type).

from geo.