Note: blender python vectors are inaccurate

Coding Python Support
1

try this in blender python console:

lst = [1, 0.7]
0.7 in lst
output: True

vec = Vector ((1, 0.7))
0.7 in vec
output: False

vec
Vector((1.0, 0.699999988079071))

2

Off the top of my head, Python uses 64-bit floating point numbers out of the box. Vector however uses 32-bit float point numbers as far as I remember.

1 Like
3

You may experience this issue elsewhere as well if you use in with floats because of binary precision:

>>> (0.1 + 0.2) in [0.3]
False

You should not use ‘in’ with floats in general in Python. Use some other method for searching taking the precision into account.

1 Like
4

Using any sort of equality (==) test with floating point numbers is a poor (not robust) software design choice.

For more dependable results, it is better to determine an appropriate “epsilon” (Δ)(acceptable error range), and do a range test.

poor test:
x == y

better test:
Δ = 1e-6
x > (y - Δ) and x < (y + Δ)

see Geometry Nodes’ “Compare” node in “Equal” mode for this in operation:
Screenshot at 2024-05-04 02-25-45

64-bit “double” floating point numbers (as in Python) are more forgiving, but blender is hard-coded to use 32-bit “single” floating point numbers – since available memory is often at a premium with large models. A conscious design choice of bigger models instead of more precise models. Blender was designed for Art, not for numerically accurate scientific results.

Also, the examples (7/10 and 3/10) are both values that are inexpressible in binary. No matter how many bits of mantissa your floating point numbers contained, they could Never EQUAL the quantities 7/10 and 3/10.

Such is life in the Digital Age. :face_with_diagonal_mouth:

4 Likes
5

Thanks.
It is good to know about the epsilon method.
I just suspect it is somehow slower.
,
or is it worth, to use numpy vectors instead ?
it depends 
 ?

1 Like
6

it also appeared to me, that (physics) collision simulation (bullet) is not very accurate in blender,
it may be related or not,
but these FP errors can add over many calculations.

7

Start here

and work your way outward.

Part of the “Science” in “Computer Science” to identify problems like you are noticing, and to design / implement / validate robust algorithms to deal with them.

Part of the “Engineering” in “Software Engineering” is to trade-off accuracy / reliability / speed / memory-requirements to provide cost-effective software to solve a given problem.

“Risk / Benefit Analysis” can assist to allocate time & resources in solving any given problem.

8

32-bit or 64-bit, floating point error still there (lower but still there) and “==” operator does not works.

1 Like