Why the partial sums nest toward a limit
Assume b_1 \ge b_2 \ge b_3 \ge \cdots \ge 0 with
b_n \to 0, and write
s_n for the partial sum. The whole proof is watching the even and
odd partial sums close in on each other like a vice.
Step 1 — the even partial sums increase. Group the new pair of terms added
going from s_{2n} to s_{2n+2}:
s_{2n+2} - s_{2n} = b_{2n+1} - b_{2n+2} \;\ge\; 0,
which is non-negative because b_{2n+1} \ge b_{2n+2}. So
s_2 \le s_4 \le s_6 \le \cdots — the even sums climb.
Step 2 — the odd partial sums decrease. Likewise, going from
s_{2n-1} to s_{2n+1}:
s_{2n+1} - s_{2n-1} = -\,b_{2n} + b_{2n+1} = -(b_{2n} - b_{2n+1}) \;\le\; 0,
so s_1 \ge s_3 \ge s_5 \ge \cdots — the odd sums fall.
Step 3 — every even sum lies below every odd sum. The two run toward each
other because a single term separates consecutive partial sums:
s_{2n+1} - s_{2n} = b_{2n+1} \;\ge\; 0 \quad\Longrightarrow\quad s_{2n} \le s_{2n+1}.
So the rising even sums are all bounded above (by s_1) and the
falling odd sums are all bounded below (by s_2) — two nested,
monotone, bounded sequences. Each converges; call the limits
L_{\text{even}} and L_{\text{odd}}.
Step 4 — the two limits coincide. Their gap is a single term, which vanishes:
L_{\text{odd}} - L_{\text{even}} = \lim_{n\to\infty} (s_{2n+1} - s_{2n}) = \lim_{n\to\infty} b_{2n+1} = 0.
So L_{\text{even}} = L_{\text{odd}} = s; the even and odd partial
sums squeeze onto one common limit, and the series converges:
\sum_{n=1}^{\infty} (-1)^{n+1} b_n = s, \qquad s_2 \le s_4 \le \cdots \le s \le \cdots \le s_3 \le s_1.
Step 5 — read off the error bound. Because the limit
s is always boxed between consecutive partial sums
s_n and s_{n+1}, and those two differ by
exactly b_{n+1}, the distance from any partial sum to the true value
is at most one term:
|s - s_n| \;\le\; |s_{n+1} - s_n| = b_{n+1}.
Worked example: the alternating harmonic series
The plain harmonic series diverges, but flip the signs and it springs to life. With
b_n = 1/n — positive, decreasing, tending to zero — the test
applies, and the value is a celebrated one:
\sum_{n=1}^{\infty} \frac{(-1)^{n+1}}{n} = 1 - \frac12 + \frac13 - \frac14 + \cdots = \ln 2.
The error bound says that summing to the nth term lands within
\tfrac{1}{n+1} of \ln 2 — convergence
slow enough to be honest, but guaranteed.
Let b_n > 0 with
b_{n+1} \le b_n for all n and
\displaystyle\lim_{n\to\infty} b_n = 0. Then:
-
Convergence. \displaystyle\sum_{n=1}^{\infty} (-1)^{n+1} b_n
converges to some sum s.
-
Nesting. The limit is bracketed by consecutive partial sums:
s_{2} \le s_{4} \le \cdots \le s \le \cdots \le s_{3} \le s_{1}.
-
Error bound. The truncation error is at most the first omitted term:
|s - s_n| \le b_{n+1}.
Most convergence tests tell you only that a series converges; the alternating test
hands you a free, fully explicit error bar. Because s always lives
in the interval [s_n, s_{n+1}] (or
[s_{n+1}, s_n]), the moment you stop after term
n your worst-case error is exactly the size of the
next term:
|s - s_n| \le b_{n+1}.
Want \ln 2 to three decimals? You need
b_{n+1} = \tfrac{1}{n+1} < 5 \times 10^{-4}, i.e. about
n = 2000 terms — no remainder integral, no tail estimate, just
"look at the next term". And the sign of that term even tells you which side of
s you are on. It is the most practical error bound in elementary
analysis.