Calculus intuition · Lesson 4

Limits, continuity, and optimization

Limits ask what a function approaches. Continuity asks whether the approach matches the value. Critical points and linear approximations turn those ideas into search and error control.

Core ~27 minDeep study ~78 minWhen does a limit exist but a derivative fail?Why can a linear approximation lie even when the derivative exists?

Before this: Rates of change, The derivative

How this idea was born

Derivatives already lean on a quiet promise: as the zoom window shrinks, a slope number settles. That promise is a limit. Once you trust limits, a second human need appears: can I walk the graph without falling through a hole? That is continuity. A third need follows immediately in design and learning: where does a smooth score stop climbing, and how wrong is a straight-line forecast?

Nineteenth-century analysis spent decades turning those intuitions into checkable language. Karl Weierstrass (1815-1897) pressed continuity and limits into inequality form in his lectures; MacTutor sketches the career. Historians caution against a cartoon that credits one person alone (Cauchy and others matter too). Steal the ethic carefully: if a claim about nearness matters, make the nearness negotiable and testable.

For a modern rereader, this chapter sits between one-variable derivatives and many-variable learning. You already use limits whenever you write

f'(a)

. Here you name them, connect them to continuity, then spend them on optimization and on honesty about linear approximation error.

First principles

A limit asks what y approaches as x approaches a, even if the function is undefined exactly at a.

Informally,

\lim_{x \to a} f(x) = L

means: values of

f

get arbitrarily close to

L

whenever

x

gets sufficiently close to

a

(possibly excluding

a

itself). The function may have a hole at

a

; the limit still cares about the approach.

Continuity at $a$ asks for three gifts at once:

f(a)

defined, the limit exists, and

\lim_{x \to a} f(x) = f(a).

No jump, no removable mismatch, no wild oscillation that refuses a single landing value.

f

is differentiable at

a

, then

f

is continuous at

a

. The converse fails: sharp corners and cusps can be continuous yet refuse a unique tangent slope.

For a smooth enough

f

, critical points are places where

f'(x)=0

f'

fails to exist. On a closed interval, continuous functions attain maxima and minima; candidates live at critical points and endpoints. That is the classical search checklist, not a promise that every critical point is a peak.

Near

a

, the linear approximation

f(x) \approx f(a) + f'(a)(x-a)

is the best first-order stand-in. The numerical error is the leftover

R(x) = f(x) - \bigl(f(a) + f'(a)(x-a)\bigr).

When

f''

exists and is continuous nearby,

|R(x)|

typically shrinks like

(x-a)^2

(Taylor remainder intuition). Small steps keep linear forecasts honest; large steps do not.

Worked intuition

Think of a limit as a negotiation. You name how close to

L

you demand (

\varepsilon

). Nature (or the definition) answers with how close to

a

you must stay (

\delta

). Continuity adds: the negotiated landing value must match the posted value

f(a)

Optimization on a smooth hill: walk until the slope meter reads zero, then check neighbors (or second-derivative sign) to see whether you found a crest, a trough, or a shelf. Endpoints matter when your domain is boxed.

Linear approximation is a tangent flashlight. In a tiny room it lights the floor well. Across a canyon it lies. Machine learning's local steps inherit that flashlight: useful nearby, dangerous when you leap.

Algebrica's limits, continuous-functions, and max-min pages are the formal companions. Read the pictures first; then let inequalities harden what your eye already suspects.

Common confusions

Limit equals value automatically. Only continuity forces that handshake. A hole can have a limit with no function value, or a value that disagrees with the limit.

Continuity means "looks smooth." Continuity forbids jumps and certain wild mismatches. Differentiability is the smoother upgrade.

|x|

is continuous at

0

and not differentiable there.

$f'(c)=0$ means global optimum. It means a horizontal tangent candidate. Context, endpoints, and second-derivative (or direct comparison) decide the story.

Linear approximation error is "rounding noise." It is structured leftover from curvature and higher terms. Doubling the step often multiplies error by roughly four when the quadratic term dominates.

Limits are only for calculus class. Every instantaneous rate, every continuous loss landscape assumption, and every finite-difference diagnostic leans on limit thinking.

"The function blows up so the limit is infinity." Infinite limits are a useful extension of language; they are not finite limits. Say what you mean.

History & stories

Weierstrass rebuilt analysis so continuity and limits could be checked with inequalities rather than gestures. MacTutor's Weierstrass biography is the short stop. Cauchy's earlier courses already pushed limit language hard; the dual history is richer than a single inventor plaque. Steal clarity, not a false monopoly.

The practical payoff arrived wherever engineers and scientists needed to know whether a quantity settles under zoom, whether a model jumps, and whether a peak search is honest.

Real world

A temperature sensor that chatters: average in a shrinking window and ask whether the reading settles. A stock chart with a trading halt: continuity failed in the recorded series even if the underlying price "should" have moved smoothly. A bike route elevation profile: critical points mark local climbs and descents; endpoints are the start and finish.

Anytime you trust a straight-line forecast for a curved relationship, you are betting that the remainder

R(x)

is small enough for your decision.

Open when you want the machine-learning connection. Skip freely.

Field notes, extra examples, and glossary live here when you want more than the core path.

Carry this forward

Limits make instantaneous rates honest; continuity makes graphs traversable.
Optimization hunts critical points and boundaries.
Linear steps are local tools; error control is part of the craft.
Next: pack slopes into gradients for many knobs at once.

Go deeper when curiosity hits - videos, essays, and primary trails.

referenceMacTutor - Karl Weierstrass
video3Blue1Brown - Limits (essence of calculus)
articleAlgebrica - LimitsApproach language before epsilontic polish.
articleAlgebrica - Continuous functionsLimit equals value; discontinuity types.
articleAlgebrica - Maximum, minimum, and inflection pointsCritical-point studio for smooth curves.
bookSteven Strogatz - The Joy of XChange chapters that make nearness feel physical.