The density of the extended distribution #
The pointwise identities behind Lemmas 3 and 4 of the paper (proof of Lemma 3 in Section S1 of the supplementary material) are proved here.
- With
x_{1:T} ∼ π, the indicesbof the reference, the particle positions, and the ancestor indices generated by steps (a) and (b) of Algorithm 1 have the densityπ(x_1^{b_1}, …, x_T^{b_T}) N^{-T} fwdDensity b y a, becausex_t = x_t^{b_t}. On the set wherea_{t-1}^{b_t} = b_{t-1}for2 ≤ t ≤ T, which is the set wherebis traced back fromb_T, this density is the density (6) at(y, a, b_T). It vanishes elsewhere (target_mul_fwdDensity). - As a function of
(a, b_T), the density (6) is a function of the particle positions timesW_T^{b_T} ∏_{t,n} B_t(a_t^n|n)(extDensity_eq_mul). These products sum to one (sum_weight_mul_prod_bsprob).
The density (6) is also proved to be measurable.
If a_{t-1}^{b_t} = b_{t-1} for 2 ≤ t ≤ T, the numerator of the target density at
(x_1^{b_1}, …, x_T^{b_T}) times the density of step (b) of Algorithm 1 is
W_T^{b_T} ∏_t {∑_i g_t(x_t^i)} ∏_n m_1(x_1^n) ∏_{t ≥ 2} ∏_n W_{t-1}^{a_{t-1}^n} m_t(x_{t-1}^{a_{t-1}^n}, x_t^n), the numerator of the density (6) at (y, a, b_T).
On the set where a_{t-1}^{b_t} = b_{t-1} for 2 ≤ t ≤ T, the density of x_{1:T} ∼ π
followed by steps (a) and (b) of Algorithm 1, π(x_1^{b_1}, …, x_T^{b_T}) N^{-T} fwdDensity b y a,
is the density (6) at (y, a, b_T).
The density of x_{1:T} ∼ π followed by steps (a) and (b) of Algorithm 1 at (y, a, b) is the
density (6) at (y, a, b_T) if b is traced back from b_T, and zero otherwise.
As a function of (a, b_T), the density (6) is a function of the particle positions times
W_T^{b_T} ∏_{t,n} B_t(a_t^n|n): it is the product of
∏_t {∑_i g_t(x_t^i)} ∏_n m_1(x_1^n) ∏_{t ≥ 2} ∏_j ∑_k W_{t-1}^k m_t(x_{t-1}^k, x_t^j) / (Z N^T)
and of W_T^{b_T} ∏_{t,n} B_t(a_t^n|n).
With at least one particle, the products W_T^{b_T} ∏_{t,n} B_t(a_t^n|n) sum to one over the
ancestor indices and the final index.
The density (6) is a measurable function of the particle positions, the ancestor indices, and the final index.