Miscellaneous computations using polynomial arithmetic
(p1 <- poly_calc(1:6)) ## a monic polynomial with given zeros
720 - 1764*x + 1624*x^2 - 735*x^3 + 175*x^4 - 21*x^5 + x^6
solve(p1) ## check that the zeros are as specified
[1] 1 2 3 4 5 6
polyroot(coef(p1)) ## check using the Traub-Jenkins algorithm
[1] 1-0i 2+0i 3+0i 4-0i 5+0i 6-0i
p2 <- -change_origin(p1, 3) ## a negative shifted version of p1
plot(p1, xlim = c(-2.5, 6.5), ylim = c(-20, 20), lwd = 2, col = "steel blue", bty = "n")
lines(p2, col = "brown", lwd = 2)
abline(h = 0)
points(cbind(solve(p1), 0), pch = 1, cex = 1, col = "brown")
points(cbind(solve(p2), 0), pch = 19, cex = 0.5, col = "steel blue")
## stationary points
stat <- solve(deriv(p1))
lines(tangent(p1, stat), limits = cbind(stat-0.5, stat + 0.5),
lty = "solid", col = "black")
points(stat, p1(stat), pch = 19, cex = 0.5, col = "red")
stat <- solve(deriv(p2))
lines(tangent(p2, stat), limits = cbind(stat-0.5, stat + 0.5),
lty = "solid", col = "cadet blue")
points(stat, p2(stat), pch = 19, cex = 0.5, col = "red")
## various checks
z <- (-2):6
setNames(p1(z), paste0("z=", z))
z=-2 z=-1 z=0 z=1 z=2 z=3 z=4 z=5 z=6
20160 5040 720 0 0 0 0 0 0
setNames(p2(z), paste0("z=", z))
z=-2 z=-1 z=0 z=1 z=2 z=3 z=4 z=5 z=6
0 0 0 0 0 0 -720 -5040 -20160
setNames((p1*p2)(z), paste0("z=", z))
z=-2 z=-1 z=0 z=1 z=2 z=3 z=4 z=5 z=6
0 0 0 0 0 0 0 0 0
p3 <- (p1 - 2 * p2)^2 ## moderately complicated expression.
setNames(p3(0:4), paste0("z=", 0:4)) ## should have zeros at 1, 2, 3
z=0 z=1 z=2 z=3 z=4
518400 0 0 0 2073600
Discrete orthogonal polynomials
Find the orthogonal polynomials on \(x_0 =
(0, 1, 2, 3, 5)\) and construct R functions to evaluate them at
arbitrary \(x\) values.
x0 <- c(0:3, 5)
op <- poly_orth(x0, norm = TRUE)
plot(op, lwd = 2, legend = TRUE)
fop <- as.function(op) ## Explicit coercion needed for polylist
zap(crossprod(fop(x0))) ## Verify orthonormality
P0 P1 P2 P3 P4
P0 1 0 0 0 0
P1 0 1 0 0 0
P2 0 0 1 0 0
P3 0 0 0 1 0
P4 0 0 0 0 1
Chebyshev polynomial approximation I
The Chebyshev polynomials of the first kind, \(T_n(x)\), like all orthogonal sets, satisfy
a 2-term recurrence relation: \[
\begin{aligned}
T_0(x) &= 1,\quad
T_1(x) = x,\\
T_n(x) &= 2x T_{n-1}(x) - T_{n-2}(x), \qquad n = 2, 3, \ldots
\end{aligned}
\] The orthogonality relation is then: \[
\int_{-1}^1\frac1{\sqrt{1-x^2}}T_n(x)T_m(x)dx = c_n\delta_{nm}
\] where \(\delta_{nm}\) is the
usual Kronecker delta function, and \(c_0 =
\pi, c_n = \pi/2, n = 1, 2, \ldots\). To approximate a function,
\(f(x)\) on the range \(-1 < x < 1\), the Chebyshev
approximation of the first kind uses a polynomial of the form \[
f(x) \approx \sum_{j=0}^N \beta_jT_j(x)
\] where the coefficients are given by \[
\beta_j = \int_{-1}^1\frac1{\sqrt{1-x^2}}T_j(x)f(x)dx / c_j
\] We now explore the quality of this approximation for the
central part of a \(N(0,\sigma=\frac14)\) density using \(N=14\), (possibly a little overkill!).
x <- polynomial()
Tr <- polylist(1, x)
for(j in 3:15) {
Tr[[j]] <- 2*x*Tr[[j-1]] - Tr[[j-2]]
}
Tr <- setNames(Tr, paste0("T", sub(" ", "_", format(seq_along(Tr)-1))))
The necessary inner product function has been supplied as part of the
package, slightly indirectly, but are easy to specify anyway
ChebyT <- function(p, q = p) {
integrate(function(x) 1/sqrt(1-x^2)*p(x)*q(x), lower = -1, upper = 1,
subdivisions = 500, rel.tol = .Machine$double.eps^0.5)$value
}
zap(outer(Tr, Tr, Vectorize(ChebyT))*2/pi) ## check of orthogonality
T_0 T_1 T_2 T_3 T_4 T_5 T_6 T_7 T_8 T_9 T10 T11 T12 T13 T14
T_0 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0
T_1 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0
T_2 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0
T_3 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0
T_4 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0
T_5 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0
T_6 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
T_7 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0
T_8 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0
T_9 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0
T10 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0
T11 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0
T12 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0
T13 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0
T14 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1
The approximating polynomial for the normal density can now be found
and tested.
fx <- function(x) dnorm(x, sd = 0.25)
b <- sapply(Tr, ChebyT, q = fx)/sapply(Tr, ChebyT)
fx_approx <- sum(b * Tr)
curve(fx, xlim = c(-1,1))
lines(fx_approx, col = "red", limits = c(-1, 1))
curve(1e4*(fx(x) - fx_approx(x)), xlim = c(-1,1)) ## error pattern x 10000
The approximation may be extended to the cumulative distribution
function:
Fx <- function(x) pnorm(x, sd = 0.25)
Fx_approx <- integral(fx_approx) + 0.5
curve(Fx, xlim = c(-1, 1))
lines(Fx_approx, col = "red", limits = c(-1,1))
curve(1e4*(Fx(x) - Fx_approx(x)), xlim = c(-1,1)) ## error pattern x 10000
Chebyshev polynomial approximation II
Functions that “bend the other way”, Chebyshev polynomials of the
second kind can sometimes be more useful. These have the same recurrence
relation as those of the first kind, but a different starting set and a
different, but similar looking orthogonality relationship. \[
\begin{aligned}
U_0(x) &= 1,\quad
U_1(x) = 2x,\\
U_n(x) &= 2x U_{n-1}(x) - U_{n-2}(x), \qquad n = 2, 3, \ldots\\
\int_{-1}^1&\sqrt{1-x^2}U_n(x)U_m(x)dx = \frac{\pi}2\delta_{nm}
\end{aligned}
\] We use these polynomials to build a polynomial approximation
to the arc sine function, \(f(x) =
\sin^{-1}x\). The process is entirely analogous to the one
outlined previously.
x <- polynomial()
Ur <- polylist(1, 2*x)
for(j in 3:15) {
Ur[[j]] <- 2*x*Ur[[j-1]] - Ur[[j-2]]
}
Ur <- setNames(Ur, paste0("U", sub(" ", "_", format(seq_along(Ur)-1))))
plot(Ur, ylim = c(-2,2))
ChebyU <- function(p, q = p) {
integrate(function(x) sqrt(1-x^2)*p(x)*q(x), lower = -1, upper = 1,
subdivisions = 500, rel.tol = .Machine$double.eps^0.5)$value
}
b <- sapply(Ur, ChebyU, q = asin)/(pi/2)
asin_approx <- sum(b * Ur)
curve(asin, xlim = c(-1,1))
lines(asin_approx, col = "red", limits = c(-1,1))
curve(1e4*(asin(x) - asin_approx(x)), xlim = c(-1,1), ylim = c(-50,50)) ## errors by 10000
Overall the approximation is not especially good, and degrades
considerably close to the ends of the range where the derivative of the
function itself becomes unbounded.
A simple branching process
If an organism can have \(0, 1, 2,
...\) offspring with probabilities \(p_0 , p_1 , p_2 , ...\) the generating
function for this discrete distribution is defined as \(P ( s ) = p_0 + p_1 s + p_2 s^2 + \cdots\).
If only a (known) finite number of offspring is possible, the generating
function is a polynomial. In any case, if all offspring themselves
reproduce independently according to the same offspring distribution,
then the generating function for the size of the second generation can
be shown to be \(P ( P ( s ))\), and so
on. There is a nice collection of results connected with such simple
branching processes: in particular the chance of ultimate extinction is
the (unique) root between 0 and 1 of the equation \(P ( s )-s = 0\). Such an equation clearly
has one root at \(s = 1\), which, if
\(P ( s )\) is a finite degree
polynomial, may be “divided out”. Also the expected number of offspring
for an organism is clearly the slope at \(s=1\), that is, \(P'(1)\).
Consider a simple branching process where each organism has at most
three offspring with probabilities \[\textstyle p_0 = {\textstyle
\frac1{10}},\quad p_1 = {\textstyle \frac5{10}},\quad p_2 =
{\textstyle \frac3{10}},\quad p_3={\textstyle \frac1{10}},\quad p_4 =
p_5 = \cdots = 0.\] The following will explore some of its
properties without further discussion.
P <- polynomial(c(1, 5, 3, 1))/10
s <- polynomial(c(0, 1))
(mean_offspring <- deriv(P)(1))
[1] 1.4
pretty_poly <- bquote(italic(P)(italic(s)) == ' '*
.(parse(text = gsub("x", "italic(' '*s)", as.character(P)))[[1]]))
plot(s, xlim = c(0,1), ylim = c(0,1), bty = "n", type = "n", main = pretty_poly,
xlab = expression(italic(s)), ylab = expression(italic(P)(italic(s))))
x <- c(0,1,1,0)
y <- c(0,0,1,1)
segments(x, y, 1-y, x, lty = "solid", lwd = 0.2)
lines(s, limits = 0:1, col = "grey")
lines(P, limits = 0:1)
lines(tangent(P, 1), col = "red", limits = c(0.5, 1.0), lwd = 1.5)
(ep <- solve((P - s)/(1 - s))) ## extinction; factor our the known zero at s = 1
[1] -4.236068 0.236068
ex <- ep[2] ## extract the appropriate value (may be complex, in general)
plot(s, xlim = c(0,1), ylim = c(0,1), type = "n", bty = "n", main = pretty_poly,
xlab = expression(italic(s)), ylab = expression(italic(P)(italic(s))))
segments(x, y, 1-y, x, lty = "solid", lwd = 0.2)
lines(s, col = "grey", limits = 0:1) ## higher generations
lines(P, col = 1, limits = 0:1)
lines(P(P), col = 2, limits = 0:1)
lines(P(P(P)), col = 3, limits = 0:1)
lines(P(P(P(P))), col = 4, limits = 0:1)
arrows(ex, P(ex), ex, par("usr")[3], angle = 15, length = 0.125)
Polynomials can be numerically fragile
This can easily lead to surprising numerical problems.
x0 <- 80:89
y0 <- c(487, 370, 361, 313, 246, 234, 173, 128, 88, 83)
p <- poly_calc(x0, y0) ## leads to catastropic numerical failure!
range(p(x0) - y0) ## these should be "close to zero"!
[1] -2.5 96.5
p1 <- poly_calc(x0 - 84, y0) ## changing origin fixes the problem
range(p1(x0 - 84) - y0) ## these are 'close to zero'.
[1] -2.199840e-11 7.389644e-12
plot(p1, xlim = c(80, 89) - 84, xlab = "x0 - 84")
points(x0 - 84, y0, col = "red")
## Can we now write the polynomial in "raw" form?
p0 <- change_origin(p1, -84) ## attempting to change the origin back to zero
## leads to severe numerical problems again
plot(p0, xlim = c(80, 89))
points(x0, y0, col = "red") ## major errors due to finite precision
This inner product function, along with some others for the classical
orthogonal polynomials of mathematical physics are included, although if
formulae for the recurrence relation are known, using them directly
would be more efficient.