Part I · Chapter 6 needs review

Analytic continuation

The sum for ζ(s)\zeta(s) has a value only when the of ss is greater than 1. Analytic continuation is the fact that the function it defines there extends, in exactly one way, to the rest of the . This chapter shows what that means for a function from calculus, and then does it for ζ\zeta.

A function and a formula are different things

The geometric series 1+x+x2+x3+⋯1 + x + x^{2} + x^{3} + \cdots converges for −1<x<1-1 \lt x \lt 1, and there its sum is 1/(1−x)1/(1-x). Outside that interval the series has no sum: at x=2x = 2 its terms are 1, 2, 4, 8, …

The expression 1/(1−x)1/(1-x) has no such restriction. It is defined for every xx except 1: at x=2x = 2 it is −1, at x=−3x = -3 it is 14\tfrac14.

So there is one function, 1/(1−x)1/(1-x), and the series is a way of computing it that works only on part of its domain. The series failing at x=2x = 2 says something about the series, not about the function.

Lab 6.1 the geometric series and the function 1/(1 − x)
1/(1 − x)1 + x + … + xN − 1dashed lines: x = −1 and x = 1
Between the dashed lines, adding terms brings the orange curve onto the blue one. Outside them, adding terms makes it worse. Each curve is drawn only where it fits in the frame.

Why the extension is forced

Suppose all you knew was the geometric series, valid on (−1,1)(-1, 1). Could some other function agree with it there and differ from 1/(1−x)1/(1-x) elsewhere?

Not an analytic one. By the rigidity of chapter 5, two analytic functions that agree on an interval agree everywhere that both are defined. So a function given on a small region has at most one analytic extension to a larger one.

That extension, when it exists, is the . It is not a convention or a choice of how to assign a value to a divergent sum. It is the same function, seen on more of its domain.

Extending step by step

The Taylor series of 1/(1−x)1/(1-x) about another center cc is

11−x  =  ∑k≥0(x−c)k(1−c)k+1\frac{1}{1-x} \;=\; \sum_{k\ge 0} \frac{(x-c)^{k}}{(1-c)^{k+1}} This is valid when ∣x−c∣<∣1−c∣|x - c| \lt |1 - c|.

The series converges out to the distance from the center to the bad point x=1x = 1, and no further. Starting from the original series (center 0, valid on (−1,1)(-1, 1)), compute the function and its derivatives at c=−0.9c = -0.9. The series about that center is valid on (−2.8,1)(-2.8, 1), which reaches points the original did not. Repeating this extends the function as far to the left as desired.

On the real line this process cannot get past x=1x = 1. In the complex plane it can: a chain of overlapping discs can go around the bad point, and the function on the far side is still determined.

Lab 6.2 the Taylor series of 1/(1 − x) about a center c
1/(1 − x)N terms of the series about cdashed lines: the ends of the interval where the series converges
The interval of convergence is centered on cc and always ends at x=1x = 1. Moving the center to the left moves the other end twice as far.

Continuing zeta

The same thing happens for ζ\zeta. Chapter 2 estimated what is left after NN terms of the sum: about N1−s/(s−1)N^{1-s}/(s-1). Turn that around. Take the first NN terms and add the estimate of the rest:

ζ(s)  =  lim⁡N→∞[  1+12s+⋯+1Ns  −  N1−s1−s  ]\zeta(s) \;=\; \lim_{N\to\infty}\left[\;1+\frac{1}{2^{s}}+\dots+\frac{1}{N^{s}} \;-\; \frac{N^{1-s}}{1-s}\;\right]

For σ>1\sigma \gt 1 the extra term tends to 0 and this is the original sum. But the limit exists for every ss with σ>0\sigma \gt 0 except s=1s = 1, and it is analytic there. By the rigidity above it is the continuation of ζ\zeta to the half-plane σ>0\sigma \gt 0.

In the picture of chapter 3: for 0<σ<10 \lt \sigma \lt 1 the chain of vectors does not settle. It circles in an ever wider spiral, and the term N1−s/(1−s)N^{1-s}/(1-s) is exactly the vector from the center of that spiral to the chain's current end. Subtracting it leaves the center, and the center is ζ(s)\zeta(s).

Lab 6.3 the chain of vectors for σ ≤ 1, and the center of its spiral
the chain of vectorsthe chain with the correction subtracted● the end of the chain● ζ(s), computed independently
The blue chain widens forever when σ<1\sigma \lt 1. The orange path is the same chain with the correction removed at each step, and it closes in on the dot. (The lab also removes half of the last vector, a refinement that makes the orange path converge faster.) The dot is computed by a different method, described in chapter 7.

The whole plane

gave two further formulas that extend ζ\zeta to every except s=1s = 1. At s=1s = 1 the function has : it behaves like 1/(s−1)1/(s-1), as in Lab 2.2. From here on, ζ(s)\zeta(s) means this extended function.

The extended function takes the values ζ(0)=−12\zeta(0) = -\tfrac12 and ζ(−1)=−1/12\zeta(-1) = -1/12. The original sum at s=−1s = -1 would be 1+2+3+4+⋯1 + 2 + 3 + 4 + \cdots, which has no sum. The statement "1+2+3+⋯=−1/121 + 2 + 3 + \cdots = -1/12" is shorthand for ζ(−1)=−1/12\zeta(-1) = -1/12, in the same way that "1+2+4+8+⋯=−11 + 2 + 4 + 8 + \cdots = -1" is shorthand for the value of 1/(1−x)1/(1-x) at x=2x = 2.

Things to try

  • In Lab 6.1, read values at x=−1x = -1. The partial sums alternate between 1 and 0 as N changes, while the function is 12\tfrac12.
  • In Lab 6.2, slide the center from 0 to −3. The series now reproduces the function at x=−5x = -5, far outside the original interval.
  • In Lab 6.3, set σ=0.5\sigma = 0.5 and t=14.13t = 14.13. The center of the spiral is at the origin: ζ(s)=0\zeta(s) = 0 there.
  • Raise σ\sigma past 1. The blue spiral stops widening and winds inward, and both paths end at the same point.

Exercises

  1. Derive the Taylor series of 1/(1−x)1/(1-x) about the center cc.
    Solution

    Write 1−x=(1−c)−(x−c)=(1−c)(1−u)1 - x = (1-c) - (x-c) = (1-c)(1-u) with u=x−c1−cu = \dfrac{x-c}{1-c}. Then 11−x  =  11−c(1+u+u2+⋯ )\frac{1}{1-x} \;=\; \frac{1}{1-c}\left(1+u+u^{2}+\cdots\right) valid for ∣u∣<1|u| \lt 1, that is, ∣x−c∣<∣1−c∣|x - c| \lt |1 - c|.

  2. Use the formula for ζ\zeta with N=1N = 1, and half of the last term removed, to estimate ζ(12)\zeta(\tfrac12).
    Solution

    1−12−11/21−12=1−12−2=−1.51 - \tfrac12 - \dfrac{1^{1/2}}{1 - \tfrac12} = 1 - \tfrac12 - 2 = -1.5. The true value is −1.4604. With N=10N = 10 the same recipe gives −1.4617.

  3. In what sense is "1+2+4+8+⋯=−11 + 2 + 4 + 8 + \cdots = -1" true, and in what sense false?
    Solution

    As a statement about adding numbers it is false: the partial sums 1, 3, 7, 15, … grow without bound. It is true as shorthand for "the function 1/(1−x)1/(1-x), whose series is 1+x+x2+⋯1 + x + x^{2} + \cdots where that converges, has the value −1 at x=2x = 2". The statement ζ(−1)=−112\zeta(-1) = -\tfrac{1}{12} is of the same kind.

Check your understanding

  1. The geometric series diverges at x=2x = 2, and 1/(1−x)1/(1-x) equals −1 there. Which statement is right?

    • The series secretly adds up to −1
    • The function has the value −1; the series is a way of computing the function that works only for ∣x∣<1|x| \lt 1
    • The function is undefined at 2

    The sum 1+2+4+⋯1 + 2 + 4 + \cdots has no value. The function it came from does, and analytic continuation is about the function.

  2. Why is there only one analytic continuation?

    • Two analytic functions that agree on any interval agree wherever both are defined
    • Mathematicians agreed on a convention
    • Because the series converges everywhere

    Rigidity: the values near one point fix every Taylor coefficient, and so the whole function. This is what makes it legitimate to speak of ζ(s)\zeta(s) where the sum diverges.

  3. In Lab 6.3 with σ=0.5\sigma = 0.5 the chain of vectors never settles. What is ζ(s)\zeta(s) in that picture?

    • The point where the chain ends
    • It does not exist
    • The center that the chain spirals around

    The chain's end is the center plus the vector N1−s/(1−s)N^{1-s}/(1-s), whose length grows like N1−σN^{1-\sigma}. Removing that vector leaves the center, which is the continued function.