Lesson 5a: The Real-Analysis Bridge
Somewhere in the middle of Lesson 5, the proofs quietly changed character. Up through Lesson 4, most arguments were computational: expand this, parametrize that, watch the terms cancel. Lesson 5's arguments were estimates: bound the integrand, let the radius grow, squeeze the derivative to zero. Following an estimate argument and constructing one yourself turn out to be two different skills, and I could feel the difference.
The structure of this project (session logs, spaced review, a running record of which kinds of arguments land quickly and which take longer; see How I built this) made that pattern visible instead of letting it slide by: bounding-style proofs were taking me measurably longer to internalize than computational ones. Not because any step was mysterious, but because the machinery underneath them had entered my hands piecemeal, exactly when each theorem happened to need it. And next on the syllabus was Taylor and Laurent series, where estimates stop being an occasional trick and become the whole game. So rather than push forward, the better move was to pause and lay the foundation properly: no new complex analysis, just the real analysis everything after this point stands on, ending exactly at Lesson 6's doorstep.
Where I was starting from, for calibration: multivariable and vector calculus (through Stokes' theorem) plus proof experience from discrete math — enough to carry me through five lessons. But if you're following a similar path, this is the material I'd want in hand before the series-heavy half of the subject begins.
Bounds, suprema, and infima
A number $b$ is an upper bound of a set $A$ if $a\le b$ for every $a\in A$. The definitional trap is the word "the": a bounded set never has just one upper bound; it has an entire ray of them. For $A=\left\lbrace 1-\frac1n\right\rbrace = \lbrace 0, \tfrac12, \tfrac23, \dots\rbrace$, every number in $[1,\infty)$ is an upper bound; the supremum $\sup A$ is the least of them: $\sup A=1$. Note $1\notin A$ (the sup needn't belong to the set) while $\inf A=0$ (greatest lower bound) is attained at $n=1$.
The working characterization: $s=\sup A$ iff $s$ is an upper bound and for every $\varepsilon>0$ some $a\in A$ has $a>s-\varepsilon$: nothing smaller than $s$ still works.
(Practice: $B=\left\lbrace (-1)^n\left(1+\frac1n\right)\right\rbrace$ splits into two interleaved strands with limits $\pm1$, but sup and inf care about extremes, not limits: $\sup B=\tfrac32$, $\inf B=-2$, both attained.)
Underneath it all sits the completeness axiom: every nonempty subset of $\mathbb{R}$ bounded above has a supremum in $\mathbb{R}$. This separates $\mathbb{R}$ from $\mathbb{Q}$ (the rationals below $\sqrt2$ have no rational least upper bound), and it is, quietly, the reason limits exist at all.
Key takeaway: $\sup A$ = least upper bound, $\inf A$ = greatest lower bound; neither needs to belong to $A$. Completeness: every nonempty set of reals bounded above has a supremum in $\mathbb{R}$.
What a limit actually is: $\varepsilon$–$N$
Every "$R\to\infty$" and "$\varepsilon\to0$" move in earlier lessons was leaning on a definition I'd never written down:
$$a_n\to L \quad\text{means}\quad \forall\varepsilon>0\ \exists N\ \forall n\ge N:\ |a_n-L|<\varepsilon.$$
$\varepsilon$ is a challenge ("get within this tolerance"); $N$ is the response — one fixed cutoff, depending on $\varepsilon$, past which every term complies. The index $n$ walks; the post $N$ stands still.
This definition can be played with directly: the partial sums of $\sum z^n$ trace a path in the plane, with an adjustable $\varepsilon$-disk around the limit. Scroll through the demo below — each step re-challenges the same fixed point $z$ with a smaller $\varepsilon$, and the required $N$ has to grow to keep up.
Fix $z=0.62+0.40i$, inside the unit circle, so the series converges to $1/(1-z)$. Challenge with $\varepsilon=0.35$: past $N=10$ every partial sum sits inside the disk (green), so $N=10$ is a valid response. Not the smallest possible response — just a response, which is all the definition asks for.
Tighten the challenge to $\varepsilon=0.15$. The disk shrinks, and the smallest $N$ that works climbs to $N(\varepsilon)=8$, the readout below the plane tracks it. $N=10$ still (barely) covers it.
Down to $\varepsilon=0.05$. Now $N(\varepsilon)=11$ (one more than the $N=10$ we'd fixed), and the tail of the path that used to read "inside for good" turns back to blue. The old response has stopped working: this is exactly why the definition says "$\exists N$" after "$\forall\varepsilon$," not before.
Raise $N$ to $14$ and the tail turns green again. For every $\varepsilon>0$ there is always some $N$ that works: the required $N$ just has to grow as $\varepsilon$ shrinks, exactly the pattern the definition demands.
Worked: $\frac1n\to0$: given $\varepsilon$, $N=\frac1\varepsilon$ works. This is exactly the shape of the Liouville argument: $|f'(z_0)|\le M/R$ with $R$ growing is $\frac1n\to0$ wearing a costume. Similarly $\frac{n}{n+1}\to1$: simplify the error to a single fraction first, $\left|\frac{n}{n+1}-1\right|=\frac{1}{n+1}$, giving $N=\frac1\varepsilon-1$.
Why the definition is shaped this way. Test "gets closer and closer" against $a_n=(-1)^n$: challenge with $\varepsilon=\tfrac12$ and no cutoff $N$ ever works for the candidate limit $1$: past any $N$ there are still terms at distance $2$. Same for $-1$, or any candidate. The definition correctly refuses to let an oscillating sequence converge, and the load-bearing phrase is "for every $\varepsilon$."
A harder one, with a technique worth keeping: $\sqrt{n+1}-\sqrt n\to0$. Multiply by the conjugate to get $\frac{1}{\sqrt{n+1}+\sqrt n}$, then loosen: $\sqrt{n+1}+\sqrt n>2\sqrt n$, so the term is under $\frac{1}{2\sqrt n}$, giving $N=\frac{1}{4\varepsilon^2}$. Loosening a bound to something solvable is the single most reused move in this lesson — a bound doesn't need to be tight to do its job.
The continuous version swaps $\delta$ into $N$'s role: $f$ is continuous at $c$ iff $\forall\varepsilon>0\ \exists\delta>0:|x-c|<\delta\Rightarrow|f(x)-f(c)|<\varepsilon$. (For $f(x)=3x$ at $c=2$: $\delta=\varepsilon/3$.)
And the fact tying this section to the last: a bounded monotone sequence converges — increasing and bounded above means it converges to the supremum of its terms. Completeness manufactures the limit; this is the engine that makes series converge below.
Key takeaway: $a_n\to L$ iff $\forall\varepsilon>0\ \exists N\ \forall n\ge N: |a_n-L|<\varepsilon$, "for every $\varepsilon$" rules out oscillation. Bounded monotone sequences converge: completeness is what makes limits exist.
The triangle inequality, three ways
The two-term version dates to Lesson 1: $|z+w|\le|z|+|w|$.
Chained. No new idea, just grouping: $|z_1+z_2+z_3|\le|z_1+z_2|+|z_3|\le|z_1|+|z_2|+|z_3|$, applying the two-term case twice; induction gives $\left|\sum z_k\right|\le\sum|z_k|$. (Check, values arbitrary: $z_1=1+i,z_2=-2,z_3=3i$: $\sqrt{17}\approx4.12\le\sqrt2+2+3\approx6.41$, room to spare, as usual.)
Reversed. How small can $|z-w|$ be? Feed $z=(z-w)+w$ into the triangle inequality: $|z|-|w|\le|z-w|$; swap roles for $|w|-|z|\le|z-w|$; then $-A\le X\le A\iff|X|\le A$ welds the two into $$\big||z|-|w|\big|\le|z-w|.$$ ($z=3+4i$, $w=1$: $4\le\sqrt{20}\approx4.47$.)
Key takeaway: $\left|\sum z_k\right|\le\sum|z_k|$; $\big||z|-|w|\big|\le|z-w|$: the reverse form is the tool for bounding things below.
The ML-estimate, derived
Lesson 5 used the ML-estimate as a black box: $\left|\int_\gamma h(z)\,dz\right|\le M\cdot L$. Time to open the box: the proof is the triangle inequality plus one recognition. Through the parametrization $\gamma:[a,b]\to\mathbb{C}$: $$\left|\int_\gamma h(z)\,dz\right| = \left|\int_a^b h(\gamma(t))\,\gamma'(t)\,dt\right| \le \int_a^b \big|h(\gamma(t))\big|\,\big|\gamma'(t)\big|\,dt$$ (the triangle inequality for integrals; an integral is a limit of sums). Now bound $|h(\gamma(t))|\le M$ and pull it out; what's left, $\int_a^b|\gamma'(t)|\,dt$, is speed integrated over time: the arc length $L$, exactly as in vector calculus. Hence $\left|\int_\gamma h\,dz\right|\le M\cdot L$. $\blacksquare$
Worked: $\oint_{|z|=1}z^2\,dz$: $M=1$, $L=2\pi$, bound $2\pi$ (true value: $0$, a bound's job is to hold, not to be close).
Practice, with a subtlety: $\oint_{|z|=2}z^{-2}\,dz$: on $|z|=2$, $M=\frac14$, $L=4\pi$, bound $\pi$. But $z^{-2}$ blows up at $z=0$, inside the contour — does that poison the estimate? No: ML only looks at values on the curve itself. What singularities inside do to the integral's actual value is the business of residues, Lesson 7's topic.
Key takeaway: $\left|\int_\gamma h\,dz\right|\le M\cdot L$: triangle inequality for integrals, bound by $M$, recognize arc length. It only needs the bound on the curve, nowhere else.
Geometric series, comparison, and Lesson 6's doorstep
Multiply $S_n=1+r+\dots+r^n$ by $r$ and subtract to get $S_n=\frac{1-r^{n+1}}{1-r}$. For $|r|<1$, the term $r^{n+1}\to0$ (an $\varepsilon$–$N$ fact now, not hand-waving), and watching only that term vanish leaves $$\sum_{k=0}^\infty r^k = \frac{1}{1-r}, \qquad |r|<1.$$ (Check: $r=\tfrac12$ gives $2$.)
Comparison. If $0\le c_k\le d_k$ and $\sum d_k$ converges, then $\sum c_k$ converges: its partial sums are increasing and bounded above by $\sum d_k$. Bounded monotone sequences converge. Completeness again, doing the actual work. The picture: $d_k$ is an envelope; the $c_k$ can wobble underneath it, and the envelope's convergence drags theirs along.
Now aim it at a power series. If $|a_k|\le2$ and $|z|\le\tfrac12$, then $|a_kz^k|\le2(\tfrac12)^k$, a convergent geometric envelope (sum $4$), so $\sum a_kz^k$ converges absolutely. (Practice: $|a_k|\le3$, $|z|\le\tfrac13$: envelope sums to $3\cdot\tfrac32=\tfrac92$. Converges.)
That argument (bound the coefficients, compare against a geometric envelope) is, almost verbatim, how Lesson 6 proves every holomorphic function is locally a power series. There the coefficient bound won't be handed to us; it will come from Cauchy's estimates, which is exactly why this lesson and Lesson 5 had to come first.
Key takeaway: $\sum_{k=0}^\infty r^k=\frac{1}{1-r}$ for $|r|<1$. Comparison: $0\le c_k\le d_k$ with $\sum d_k$ convergent $\Rightarrow\sum c_k$ convergent. Geometric envelopes + Cauchy's estimates = Taylor series, next lesson.