Independent coordinates: conditioning, factorization, moments #
Tools for Distribution.independent p, the law of a random point x : ι → α whose coordinates
x i ~ p i are independent (the basic factorization independent_expect_prod and the marginal
independent_expect_eval are in Dynamics.Distribution).
- Conditioning on one coordinate. Resampling the coordinate
idoes not change the law (independent_expect_update), equivalently the expectation is thep i-average over the valuebofx iof the expectation withx iset tob(independent_expect_eq_sum_update); for i.i.d. products overFin (m + 1), conditioning on the first coordinate (independent_expect_fin_succ,independent_prob_fin_succ). - Factorization.
𝔼[g(x i) G(x)] = 𝔼[g] 𝔼[G]whenGignoresx i(independent_expect_mul_of_update); two distinct coordinates are independent (independent_expect_pair); functions of the coordinates inSand outsideSare uncorrelated (expect_mul_of_dep). - Moments of sums. For centred coordinate functions the second and third moments of
∑ i, W i (x i)are the sums of the coordinate moments (independent_expect_sum_sq,independent_expect_sum_cube). - Coordinatewise combination. If
x ~ independent Pandy ~ independent Qare independent,fun i => φ i (x i) (y i)has lawindependent R, whereR iis the law ofφ i a b(expect_expect_coord).
The uniform analogues for avg over dependent products ∀ i, β i (avg_pi_prod, avg_pi_eval,
avg_pi_split) are in Dynamics.Uniform.
Conditioning on one coordinate #
Moving the weight of one coordinate: w(x) p_i(a) = p_i(x i) w(x[i ↦ a]).
Resampling one coordinate: replacing the coordinate i of an independent point by an
independent copy does not change the law.
Conditioning on one coordinate: the expectation is the average, over the value b of
coordinate i, of the expectation with that coordinate set to b.
Conditioning on the first coordinate: an i.i.d. product over Fin (m + 1) is the
average, over the first coordinate b, of the product over Fin m with b prepended.
Probabilities under an i.i.d. product over Fin (m + 1), conditioned on the first
coordinate.
Factorization #
Independence of one coordinate: if G does not depend on the coordinate i, then
𝔼[g(x i) G(x)] = 𝔼[g] 𝔼[G].
Two distinct coordinates of an independent product are independent.
Moments of sums of centred coordinates #
Two distinct coordinates are independent; for centred functions the mixed moment vanishes.
Second moment of a sum of independent centred coordinates.
The mixed third moments of independent centred coordinates vanish unless the three indices agree.
Third moment of a sum of independent centred coordinates.
Coordinatewise combination of independent families #
Coordinatewise combination: if x ~ independent P and y ~ independent Q are
independent, and R i is the law of φ i a b for a ~ P i, b ~ Q i, then
fun i => φ i (x i) (y i) has law independent R.
Independence of disjoint sets of coordinates: under an independent product, a function
of the coordinates in S and a function of the coordinates outside S are uncorrelated.