The Essence of Calculus

Start here

This is Chapter 1 of the Essence of Calculus series. The whole point of this opening chapter is a single liberating idea: the scary formulas in a calculus textbook are not handed down from on high. Anyone who understands the core pictures can rebuild them from scratch. You do not memorise, you reinvent.

The lie of memorisation

Most people meet calculus as a wall of formulas to be drilled and forgotten. Area of a circle is πr², the derivative of this is that, integrate by parts, and so on. It feels like a stack of rules invented by someone cleverer than you.

It is not. Every one of those rules was discovered by a person staring at a concrete problem and asking, "what if I broke this into tiny pieces?". If you learn to ask the same question, you can re-derive the formula on the spot. That is the difference between knowing calculus and owning it.

One picture: the area of a circle

Let us do exactly what a self-respecting inventor of calculus would do. We want the area of a circle of radius R. We do not know the formula yet. So we cheat in the most productive way possible: we slice the hard shape into easy ones.

Slice the disk into thin concentric rings. Pick one ring at radius r with a tiny thickness dr.

       _____
    ⟋    .    ⟍      <- one thin ring at radius r,
   |   ⟋   ⟍   |        thickness dr
   |  | ( r ) |  |
   |   ⟍ _ ⟋   |
    ⟍   ' '   ⟋
       ‾‾‾‾‾

If dr is small enough, that ring is basically a thin strip bent into a circle. Unroll it and it becomes an almost-perfect rectangle:

• its width is the circumference of the ring, 2πr
• its height is the thickness, dr

So the area of one ring is approximately 2πr · dr. The smaller dr gets, the closer "approximately" creeps to "exactly".

Stacking the rings into a triangle

Now comes the move. The total area of the disk is the sum of all those ring areas, from the innermost (r = 0) to the outermost (r = R). Line each unrolled rectangle up side by side, ordered by radius. Their widths run from 0 up to 2πR, their bottoms sit on a common axis.

height = 2πr
  2πR |                      /|
      |                  /    |
      |              /        |
      |          /            |
      |      /                |
      |  /                    |
    0 +------------------------+
      0          r            R

The tops of the rectangles trace the straight line y = 2πr. Under that line is a triangle: base R, height 2πR. And the area of a triangle is something you have known since school:

area = 1/2 · base · height
     = 1/2 · R · 2πR
     = πR²

There it is. We just rebuilt the area of a circle from nothing but "slice it thin and add the pieces up". No memorising required.

The three big ideas, all hiding in that picture

That one example quietly previews the entire course.

1. The integral. Adding up infinitely many tiny ring areas to get a whole was integration. A hard, curvy problem became easy the instant we cut it into tiny pieces and summed them. That trick, "approximate with tiny bits, then add", is the heart of every integral you will ever compute.

2. The derivative. Look again at one ring: its area is 2πr · dr. That 2πr is the rate at which the total area grows as the radius nudges outward. So the area function A(r) = πr² has a rate of change of 2πr. The derivative of πr² is 2πr, and it fell out of the picture without a single rule being memorised.

3. They are inverses. Notice what happened. We added up the 2πr strips and got πr². Going the other way, the rate of change of πr² is 2πr. Integration walked us one direction, differentiation walked us straight back. That two-way street is the fundamental theorem of calculus, and it is the punchline the whole series builds towards.

Why an AI engineer should care

This is not a detour from machine learning, it is the foundation.

Every neural network learns by gradient descent: it measures how a tiny change in each weight would change the error, then nudges the weights downhill. That "how a tiny change affects the output" is exactly the derivative from this chapter, scaled up to millions of dimensions. The gradient is just a bag of derivatives.

So when you train a model, you are quietly running the calculus we just reinvented:

• Derivatives tell the optimiser which way is downhill.
• The chain rule (Chapter 4) is how backpropagation pushes those derivatives layer by layer through a deep network.
• Integrals show up the moment you touch probability: expected values, loss over a distribution, areas under curves.

Get these pictures into your bones now and backprop stops being magic. It becomes bookkeeping.

Quick gotchas

dr is not zero, it is "as small as you like". The whole game is that the approximation error shrinks faster than the thing you are measuring. We never divide by zero, we take a limit. That distinction is the next chapter's job.

The unrolled ring is not a perfect rectangle, and that is fine. The inner edge is slightly shorter than the outer edge. But that mismatch is proportional to dr², which vanishes far faster than the dr-sized area we keep. Tiny pieces forgive tiny errors.

Integration is not "the area under a curve" by definition. Area is just the friendliest first example. Integration is "add up a lot of tiny things". Sometimes those things are areas, sometimes they are distances, probabilities, or accumulated error.

What you walked away with

• Calculus formulas are not to be memorised. They can be reinvented by anyone who breaks a problem into tiny pieces.
• Integration is adding up infinitely many tiny things. We used it to rebuild πR² from thin rings.
• Differentiation is measuring a rate of change. The growth rate of πr² is 2πr, straight from the same picture.
• These two are inverse operations, the fundamental theorem of calculus, which the rest of the series unpacks.
• For an AI engineer, derivatives are the engine of gradient descent and the reason backprop works at all.

Next up, Chapter 2: we slow down and pin down what a derivative actually is. We zoom in on a moving point, ask how fast it is changing in an instant, and confront the slippery idea of "the slope at a single point". That gives us the derivative properly, and sets up the chain rule that backprop is built on. See you there.