Thus g(-1) is 1 + 2 + 3 + ... while g(z) is otherwise 0 + 0 + 0 + ... = 0. And then the zero function is the unique analytic continuation of g.
The reason people care about the Riemann zeta function is because of its deep connections to analysis, number theory, and physics.
(a <*> b)(n) = \sum_{k | n} a(k) * b(n/k)
It is easy to check that this is a ring, and it has wonderful properties which make it very easy to solve many equations of interest in this ring. This is usually called the ring of "arithmetic functions".However, it is difficult to get asymptotic estimates for the coefficients of a series by purely algebraic means. This is the first and last time that complex valued functions enter the picture, but it's a very neat trick. Let's consider two series a, b and define the functions
A(s) = \sum_{n >= 1} a(n) * n^-s
B(s) = \sum_{n >= 1} b(n) * n^-s
then we have A(s) * B(s) = \sum_{n >= 1} (\sum_{k | n} a(k) * b(n/k)) * n^-s = \sum_{n >= 1} (a <*> b)(n) n^-s
So this mapping, from the ring of arithmetic functions to the ring of (partial) complex functions with pointwise addition and multiplication. Glossing over some details for now, this allows you to analyze the function belonging to a sequence to gain information about the sequence itself. In particular, you can use the theory of complex integration and Cauchy's residue theorem to gain information about (partial sums of) coefficients.Unfortunately, the world is not quite this simple. The functions we are mapping into typically aren't very well behaved and usually aren't defined on large parts of the complex plane (consider a(n) = n^n). This means that all of our nice tools from complex analysis actually won't work very well!
The whole idea behind "analytic continuations" is that we aren't actually using this mapping! We are constructing a different (partial, injective) ring homomorphism from arithmetic functions to meromorphic complex functions.
The idea behind this is that meromorphic functions are rather restricted in what they can do. In particular, there is at most one meromorphic function A with A(s) = \sum_{n >= 1} a(n) n^-s for s with Re(s) > k, for some k. We define our mapping from sequences to functions by mapping the sequence a(n) to the meromorphic function A(s) with A(s) = \sum_{n >= 1} a(n) n^-s for Re(s) > k for some k, if this function exists.
By the same argument as above, this is a ring homomorphism, and since it is injective we can still use information about the functions to gain information about the underlying sequences.
For example, the Riemann zeta function is not really defined by the equation Zeta(s) = \sum_{n >= 1} n^-s. It is defined to be the unique meromorphic function with Zeta(s) = \sum_{n >= 1} n^-s for all s with R(s) > 1. In particular, Zeta(-1) has nothing to do with \sum_{n >= 1} n. The latter expression doesn't define a complex number at all, but for the former it is not too difficult to show that Zeta(-1) = -1/12.
The main advantage, though, is that meromorphic functions are very well behaved. This allows us to use Cauchy's residue theorem and Mellin transforms to get very deep results about the underlying sequences. If you play this game with the "von Mangoldt" sequence you can, for instance, derive an asymptotic bound on the density of the prime numbers. This is a surprisingly simple derivation, given that this problem had the worlds greatest mathematicians stumped for a hundred years!
Summing up, the mapping or "continuation" you use is choosen (!) so that multiplication of functions corresponds to your chosen multiplication in the ring of sequences and so that you get functions which are as well-behaved as possible. There is a large design space here, and you can find different "analytic continuations" for a given sequence.
It makes one wonder if there can be a better theory then that doesn't require magic. There is at least one book I know of (before I became a plasma physicist) which apparently avoids the regularization altogether, but people still do regularization like this in particle physics.
A bunch of people had issues with the Numberphile episode where these restrictions where elided (due to time, I'd imagine), but they had another episode featuring the zeta function and a math professor (rather than physics professors, who were featured on the one that went viral) explaining the summation with more context: https://www.youtube.com/watch?v=0Oazb7IWzbA.
lim \epsilon -> 0+ ( \sum_{n=1}^\infty n e^{- \epsilon n} + ...),
that is the series was multiplied by a decaying exponential function with a rate of decay that goes to zero. This sum can easily be evaluated for small epsilon takes the form
sum = 1/epsilon - 1/12 + ...
Crucially, there was another term in the calculation that naturally appeared that canceled the 1/epsilon. Without that other term, the sum would of course be infinite when epsilon -> 0.
This is much simpler than analytic continuation through the complex plane, and again, this is how the physics calculation appears in QFT courses. There is no need to appeal to complex analysis here, which leads to all of this mysticism and confusion.
So how exactly is it that by multiplying a regular sum of positive numbers by a decaying exponent do you get a negative number when you take the limit?
1) You introduce a family of series, parametrized by epsilon, whose terms, for small epsilon, closely approximate your original series. Sort of; obviously for large n they don't. But the idea is that if you pick any N and delta you can pick epsilon such that for n < N the approximation is within delta of the actual terms.
2) You show that the series in this family can all be summed and the sum of each one is 1/epsilon - 1/12 + O(epsilon).
Now of course what this means is that the sums blow up as you approximate your original series better and better, since 1/epsilon gets large. That's good, because your original series totally diverges off to infinity. ;)
The part after this point I'm less clear on, but it sounds like in the computation involved what you actually have is your (divergent) series 1+2+3+... plus some _other_ (also divergent, going off to negative infinity) stuff. And that you might be able to arrange things such that the other divergent stuff looks like -1/epsilon, cancels out the 1/epsilon from your approximation, and you come out with the sum of the two things being -1/12.
The obvious issue here is that once you start adding up divergent things by rearranging terms and telescoping you can come up with whatever answer you want: see <https://en.wikipedia.org/wiki/Riemann_series_theorem>. So this procedure all only makes sense if there are some sort of fundamental reasons to think that this particular rearrangement is the "right" one in some sense.
The issue is, of course, the illegitimate manipulation of a diverging series, which was the exact issue that prompted the original article (due to Numberphile doing it) in the first place.
Now e^(-x) is a totally bogus "cutoff function" per the definition in Terry's blog post, since it is not compactly supported, but it _is_ bounded, _does_ equal 1 at 0, and drops off fast enough that for practical purposes it can be used to do smoothed sums. In particular the smoothed sums will converge for most cases (e.g. anything where the sequence we're "summing" has at most polynomial growth will do so), which means you can at least try to do the rest of the analysis. I suspect, but have not checked, that the other places where compact support is used in his presentation also work out for the sorts of sequences we're talking about.
Either way, the upshot is that in some sense you have some sequence of approximations to your "actual" sum, indexed by N, and you show that for large N they all look like "power series" in 1/N which allows some finite number of negative exponents and all the approximations have matching coefficients for the negative exponents and the same constant term. And then you compute that constant term. Calling that the sum of the series is nonsense, of course, but it can still give you interesting information about something, maybe.
Funny, I thought the entire point of this 'simplification' was not to be mystical and hand wavey.
You can define a function f from a subset of the complex numbers to the complex numbers where f(z) = \sum_{n=1}^\infty 1/n^z. Be careful with the domain of this function: the series does not converge for all z. You can plug in -1 and see that symbolically, f(-1) = 1 + 2 + 3 + .... But the series does not converge for z = -1, and it is simply not true that \sum_{n=1}^\infty n = -1/12; the series does not converge; equating it to something is a nonsensical thing to do.
What is going then? Even though f is not defined for all complex numbers, there exist functions from the complex numbers to the complex numbers that -- restricted to the domain of f -- are equal to f. They "continue" f to all of the complex numbers. And if one imposes a restriction on these continuations (namely that they are analytic), then it turns out that there is a unique analytic continuation of f: the Riemann zeta function. And zeta(-1) = -1/12.
Don’t confuse the definitions here. The Riemann zeta function can be defined as the analytic continuation of the series, the series is not defined in terms of the Riemann zeta function!