zetesis-puremath Blueprint

6.4 Hilbert-Schmidt Independence Criterion

Definition 6.17 HSIC via product-space MMD

Lean: hsicDef

For a joint measure \(P\) on \(X \times Y\) with marginals \(P_X\), \(P_Y\),

\[ \mathrm{HSIC}(P) \; :=\; \mathrm{MMD}^2\bigl(P,\; P_X \otimes P_Y\bigr). \]

This is the modern unified formulation [ 20 , Thm 24 ] ; originally introduced as a cross-covariance-operator Hilbert-Schmidt norm [ 12 ] .

Theorem 6.18 Non-negativity of HSIC

Lean: hsicDef_nonneg

\(\mathrm{HSIC}(P) \ge 0\) [ 12 ] .

Theorem 6.19 HSIC vanishes on independent measures

Lean: hsicDef_eq_zero_of_independent

If \(P = P_X \otimes P_Y\) then \(\mathrm{HSIC}(P) = 0\) (direct consequence of \(\mathrm{MMD}^2(P, P) = 0\)) [ 12 ] .

Theorem 6.20 HSIC zero iff independent

Lean: hsicDef_zero_iff_independent

Under IsCharacteristic \(\mathcal{H}\) on \(X \times Y\), \(\mathrm{HSIC}(P) = 0 \iff P = P_X \otimes P_Y\) [ 12 , Thm 4 ] , [ 20 , § 4.2 ] .