A Julia package for solving three-term recurrence relationships
This package implements simple algorithms for computing solutions to recurrence relationships, including forward recurrence for initial value-problems, Olver's algorithm for two-point boundary value problems, and Clenshaw's algorithm for computing dot products with a given vector as needed for evaluating expansions in orthogonal polynomials.
As an example, consider computing the first million Chebyshev U polynomials:
julia> using RecurrenceRelationships
julia> n = 1_000_000; x = 0.1;
julia> @time forwardrecurrence(fill(2, n), zeros(n), ones(n), x)
0.006259 seconds (8 allocations: 30.518 MiB, 19.45% gc time)
1000000-element Vector{Float64}:
1.0
0.2
-0.96
-0.392
0.8815999999999999
0.56832
-0.767936
-0.7219072000000001
⋮
0.2777381949694639
-0.9332843889921485
-0.4643950727678936
0.8404053744385698
0.6324761476556076
-0.7139101449074483
-0.7752581766370972
Note this is faster than the explicit formula:
julia> θ = acos(x); @time sin.((1:n) .* θ) ./ sin(θ)
0.010828 seconds (9 allocations: 7.630 MiB)
1000000-element Vector{Float64}:
1.0
0.2
-0.96
-0.392
0.8816
0.5683199999999999
-0.7679359999999998
-0.7219072000000001
⋮
0.27773819504129643
-0.9332843889430753
-0.46439507272938874
0.840405374430785
0.6324761477118992
-0.713910144815891
-0.7752581766080119
Note forward recurrence is actually more accurate than the explicit formula which can be seen by comparison with a high precision calculation (though the accuracy is worse near ±1):
julia> norm(forwardrecurrence(fill(2, n), zeros(n), ones(n), x) - forwardrecurrence(fill(2, n), zeros(n), ones(n), big(x)))
5.93151258729473921191879934738972139533978757730288453476821870826190721098765e-10
julia> norm(sin.((1:n) .* θ) ./ sin(θ) - forwardrecurrence(fill(2, n), zeros(n), ones(n), big(x)))
4.538171757754684777956652395339636096999624380286573911589424226541646390097931e-08
We can make it even faster using FillArrays.jl:
julia> using FillArrays
julia> @time forwardrecurrence(Fill(2, n), Zeros(n), Ones(n), x)
0.003387 seconds (5 allocations: 7.630 MiB)
1000000-element Vector{Float64}:
1.0
0.2
-0.96
⋮
0.6324761476556076
-0.7139101449074483
-0.7752581766370972
We can also use LazyArrays.jl to represent Chebyshev T recurrence lazily:
julia> using LazyArrays
julia> @time forwardrecurrence(Vcat(1, Fill(2, n-1)), Zeros(n), Ones(n), x)
0.002740 seconds (103 allocations: 7.634 MiB)
1000000-element Vector{Float64}:
1.0
0.1
-0.98
-0.296
0.9208
0.48016
-0.824768
-0.6451136
0.6957452799999999
⋮
0.9968292069233
0.17885499217086823
-0.9610582084891264
-0.3710666338686935
0.8868448817153877
0.548435610211771
-0.7771577596730335
-0.7038671621463777
And this matches the explicit formula:
julia> @time cos.((0:n-1) .* θ)
0.042121 seconds (6 allocations: 7.630 MiB, 75.68% gc time)
1000000-element Vector{Float64}:
1.0
0.1
-0.98
-0.29600000000000004
0.9208
0.48016000000000003
-0.824768
-0.6451136000000001
0.6957452799999999
⋮
0.9968292069266631
0.178854992187329
-0.9610582084686914
-0.37106663377504573
0.8868448817354809
0.5484356102238197
-0.7771577596278382
-0.7038671620731024
Olver's algorithm is an approach to computing minimal solutions to recurrence relationships as well as solve inhomogenuous equations. A simple example is the Bessel equation following the DLMF:
julia> N = 1000; x = 10.0;
julia> a,b,c = ones(N-1), -range(2; step=2, length=N)/x, ones(N-1);
julia> u = olver([1; zeros(N-1)], a,b,c);
julia> J0 = -1/(-1 + 2sum(u[2:2:end])); # Use normalization condition
julia> [J0; -u*J0]
37-element Vector{Float64}:
-0.24593576445134843
0.043472746168861556
0.25463031368512073
0.058379379305186836
-0.21960268610200867
-0.23406152818679374
-0.014458842084785059
0.21671091768505168
0.31785412684385733
0.29185568526512007
0.20748610663335887
0.12311652800159764
⋮
1.4405452917644452e-9
2.762005267054569e-10
5.0937552445022125e-11
9.049766986667002e-12
1.5510960776464958e-12
2.568094792119719e-13
4.1122693467730146e-14
6.375758981501052e-15
9.573158101767982e-16
1.339885277011767e-16
-1.9396116268561495e-17
This matches the special functions definition:
julia> [besselj(k,x) for k=0:36]
37-element Vector{Float64}:
-0.2459357644513483
0.04347274616886144
0.25463031368512057
0.058379379305186795
-0.2196026861020085
-0.23406152818679363
-0.014458842084785123
0.2167109176850515
0.31785412684385717
0.29185568526512007
0.20748610663335898
0.1231165280015976
⋮
1.4405452917644441e-9
2.7620052670546077e-10
5.093755244504228e-11
9.049766986775815e-12
1.5510960782574664e-12
2.568094827689877e-13
4.112271491025767e-14
6.3758926566612455e-15
9.581766237065793e-16
1.3970838454349004e-16
1.9782068097851055e-17
Clenshaw's algorithm is an efficient way to compute expansions in orthogonal polynomials. Here we compute
$$
∑_{k=0}^{n-1} {U_k(x) \over k+1}
$$
for
julia> @time clenshaw(inv.(1:n), fill(2, n), zeros(n), ones(n+1), x)
0.006446 seconds (12 allocations: 30.518 MiB)
0.8396901361362448
This matches the explicit expression:
julia> @time sum(sin((k+1)*θ)/(k+1) for k=0:n-1)/sin(θ)
0.161225 seconds (8.03 M allocations: 124.408 MiB, 2.74% gc time, 16.44% compilation time)
0.8396901361362544
Again, using FillArrays.jl is faster. And we can use InfiniteArrays.jl to allow any length of coefficients:
julia> using InfiniteArrays
julia> @time clenshaw(inv.(1:n), Fill(2, ∞), Zeros(∞), Ones(∞), x)
0.004574 seconds (5 allocations: 7.630 MiB)
0.8396901361362448