zetesis-puremath Blueprint

6.3 Maximum Mean Discrepancy

Definition 6.10 Squared MMD

Lean: mmdSq

\(\mathrm{MMD}^2(P, Q) \; :=\; \| \mu _P - \mu _Q \| ^2\) [ 11 , Defn. 2 ] .

Theorem 6.11 Non-negativity of \(\mathrm{MMD}^2\)

Lean: mmdSq_nonneg

\(\mathrm{MMD}^2(P, Q) \ge 0\) [ 11 , Lemma 5 ] .

Theorem 6.12 \(\mathrm{MMD}^2\) on identical measures

Lean: mmdSq_self

\(\mathrm{MMD}^2(P, P) = 0\) [ 11 , Lemma 5 ] .

Theorem 6.13 Symmetry of \(\mathrm{MMD}^2\)

Lean: mmdSq_comm

\(\mathrm{MMD}^2(P, Q) = \mathrm{MMD}^2(Q, P)\) [ 11 , Lemma 5 ] .

Definition 6.14 Characteristic RKHS

Lean: IsCharacteristic

\(\mathcal{H}\) is characteristic on \(X\) iff the kernel mean embedding \(P \mapsto \mu _P\) is injective. This is the modern definition [ 20 , Defn. 6 ] .

Theorem 6.15 Zero-iff-equal under characteristic kernel

Lean: mmdSq_zero_iff

Under IsCharacteristic \(\mathcal{H}\), \(\mathrm{MMD}^2(P, Q) = 0 \iff P = Q\) [ 20 , Thm 8 ] , [ 11 , Thm 5 ] .

Theorem 6.16 Forward direction: equal measures \(\Rightarrow \) zero MMD

Lean: mmdSq_eq_zero

\(P = Q \Rightarrow \mathrm{MMD}^2(P, Q) = 0\) (direct consequence of \(\mu _P = \mu _Q\)) [ 11 , Lemma 5 ] .