PNLAlgorithm<T>

Models & Types · AiDotNet.CausalDiscovery.Functional

PNL (Post-Nonlinear Causal Model) — Y = g(f(X) + N).

Inheritance: CausalDiscoveryBase<T> → FunctionalBase<T> → PNLAlgorithm<T>

For Beginners

Sometimes the relationship between cause and effect has two layers of nonlinearity: first the cause produces an effect through some function f, then that result gets distorted by another function g (like a sensor with nonlinear response). PNL can handle this double-nonlinearity by “undoing” the outer distortion before testing for the cause-effect direction.

How It Works

PNL extends the Additive Noise Model by allowing a post-nonlinear distortion g. The model is Y = g(f(X) + N) where f is the causal mechanism, N is independent noise, and g is an invertible post-nonlinear transformation. The algorithm:

  • For each variable pair, fits the inner function f via kernel regression.
  • Estimates the post-nonlinear function g by applying a rank-based CDF transform

(probability integral transform) to approximate g^(-1).

  • Computes residuals in the “linearized” space after inverting g.
  • Tests independence of the residuals from the cause using mutual information.
  • Orients the edge in the direction where residuals are more independent.

Reference: Zhang and Hyvarinen (2009), “On the Identifiability of the Post-Nonlinear Causal Model”, UAI.

Constructors

Constructor Summary
PNLAlgorithm(CausalDiscoveryOptions options)

Properties

Property Summary
bool SupportsNonlinear { get; }
string Name { get; }