Loading [MathJax]/jax/output/CommonHTML/jax.js

Blog (mostly math)

Uniform approximations for log(x)

Take an interval of the form [δ,2Aδ]R>0. So here 0<δ<A. Consider log(x) on [δ,2Aδ].

We’ll find a sequence (Pn(x)) of polynomials of degrees deg(Pn)n such that maxx[δ,2Aδ]|log(x)Pn(x)| is O(1n).

That is, we’ll specify a seq of polys (Pn(x)) with deg(Pn)n and a constant C>0, such that |log(x)Pn(x)|Cn for all x[δ,2Aδ] and all but finitely many nZ>0.

Consider log(1+x) on (1,). Its Taylor polynomials about 0 are Sn(x)=xx22++(1)n1xnn. Writing log(1+x)=Sn(x)+Rn(x) we have Rn(x)=11+x(1x++(1)n1xn1) =1(1+(1)n1xn)1+x =(1)nxn1+x. So Rn(x)=Rn(0)+x0Rn(t)dt =(1)nx0tn1+tdt.
When x[0,1], we have |Rn(x)| =x0tn1+tdt x0tn1+0dt =xn+1n+1 1n+1. When x[1+Δ,0) for some Δ(0,1), we have |Rn(x)|=||x|0tn1+tdt| s=(t)=||x|0(s)n1s(ds)| =|x|0sn1sds |x|0sn1|x|ds =|x|n+1(1|x|)(n+1) |x|n+1Δ(n+1) 1Δ(n+1).
Combining the estimates, we have : If Δ(0,1), then maxx[1+Δ,1]|Rn(x)|1Δ(n+1).

Back to the original problem. For x[δ,2Aδ], we see log(x)=log(A)+log(1+xAA) =log(A)+Sn(xAA)+Rn(xAA).
Further maxx[δ,2Aδ]|Rn(xAA)| =maxxAA[(1δA),1δA]|Rn(xAA)| =maxu[(1δA),1δA]|Rn(u)| A/δn+1.

So maxx[δ,2Aδ]|log(x)(log(A)+Sn1(xAA))|A/δn, as needed.

In fact here the polynomials Pn(x):=log(A)+Sn1(xAA) have degrees deg(Pn)=(n1).