Hints

MATH-C08-C-001

Too large means overflow.

Subtract the maximum logit.

m is the maximum value.

Multiply one pair first: 0.1 * 0.1 = 0.01.

The denominator is 2h.

Floating point stores finite approximations.

Epsilon is a guardrail, but it is still part of the expression.

Compare how much the output changes from the same kind of input error.

Repeated multiplication by numbers below 1 shrinks scale.

Think about how many parameters a large network has.

Floating-point arithmetic rounds intermediate values.

Stable softmax subtracts the same constant from every logit, so differences are preserved.

Large h is less local. Tiny h can make subtractive rounding error dominate.

Epsilon changes the expression. Use it when there is a specific danger.