Sigmoid Functions

Definition

Sigmoid functions are bounded S-shaped real functions, often monotone, with limiting behavior at both extremes. The canonical logistic sigmoid is

which maps real inputs to (0,1).

Why It Matters

Sigmoids are a shared mathematical shape across neural activations, probability response curves, cumulative distribution functions, and bounded growth models. They are useful whenever a system transitions smoothly between regimes rather than growing or changing without limit.

Formalism / Key Objects

  • Logistic function: \sigma(x)=1/(1+e^{-x}).
  • Derivative: \sigma'(x)=\sigma(x)(1-\sigma(x)), a bell-shaped curve.
  • Focused annex: Logistic Sigmoid and Saturating Growth stores the derivative and growth-model equations.
  • Inverse logit: \operatorname{logit}(p)=\log(p/(1-p)) for p \in (0,1).
  • Common variants: hyperbolic tangent, error-function sigmoid, arctangent sigmoid, Gompertz-style curves, and CDFs of common probability distributions.

Connections

Common Confusions

  • “Sigmoid” can mean a broad S-shaped function class or specifically the logistic function, depending on context.
  • A sigmoid curve is not the same as an exponential curve; it may look exponential locally while remaining bounded globally.
  • Using a sigmoid activation does not by itself guarantee good training or generalization.

Key Sources