Algorithm 4: Elim -Cons PE
	input: A belief network ℬ = {P1, …, Pn} where Pi's are assume to have a sparse representation; A constraint expression over k variables, ℛ = {RQ1, …, RQt} an ordering d = {X1, …, Xn}
	output: The belief P(ℛ).
1	Place buckets with observed variables last in d (to be processed first)	// Initialize
	Partition ℬ and ℛ into bucket1, …, bucketn, where bucketi contains all CPTs and constraints whose highest variable is Xi
	Let S1, …, Sj be the scopes of the CPTs, and Q1, …Qt be the scopes of the constraints.
	We denote probabilistic functions as λ s and constraints by Rs
2	for p ← n down to 1 do	// Backward
	Let λ1, …, λj be the functions and R1, …, Rr be the constraints in bucketp
	Process-bucket-RELp(Σ, (λ1, …, λj),(R1, …, Rr))
3	return P(ℛ) as the result of processing bucket1.