Extracting orthogonal polynomial coefficients from R's poly() function?

Question

R\'s poly() function produces orthogonal polynomials for data fitting. However, I would like to use the results of the regression outside of R (say in C++), and

Answer 1

The polynomials are defined recursively using the alpha and norm2 coefficients of the poly object you've created. Let's look at an example:

z <- poly(1:10, 3)
attributes(z)$coefs
# $alpha
# [1] 5.5 5.5 5.5
# $norm2
# [1]    1.0   10.0   82.5  528.0 3088.8

For notation, let's call a_d the element in index d of alpha and let's call n_d the element in index d of norm2. F_d(x) will be the orthogonal polynomial of degree d that is generated. For some base cases we have:

F_0(x) = 1 / sqrt(n_2)
F_1(x) = (x-a_1) / sqrt(n_3)

The rest of the polynomials are recursively defined:

F_d(x) = [(x-a_d) * sqrt(n_{d+1}) * F_{d-1}(x) - n_{d+1} / sqrt(n_d) * F_{d-2}(x)] / sqrt(n_{d+2})

To confirm with x=2.1:

x <- 2.1
predict(z, newdata=x)
#               1         2         3
# [1,] -0.3743277 0.1440493 0.1890351
# ...

a <- attributes(z)$coefs$alpha
n <- attributes(z)$coefs$norm2
f0 <- 1 / sqrt(n[2])
(f1 <- (x-a[1]) / sqrt(n[3]))
# [1] -0.3743277
(f2 <- ((x-a[2]) * sqrt(n[3]) * f1 - n[3] / sqrt(n[2]) * f0) / sqrt(n[4]))
# [1] 0.1440493
(f3 <- ((x-a[3]) * sqrt(n[4]) * f2 - n[4] / sqrt(n[3]) * f1) / sqrt(n[5]))
# [1] 0.1890351

The most compact way to export your polynomials to your C++ code would probably be to export attributes(z)$coefs$alpha and attributes(z)$coefs$norm2 and then use the recursive formula in C++ to evaluate your polynomials.