Real Analysis Preview

Pick up the function f(x) = 3x + 1 and the point x = 2. By eye, \lim_{x \to 2} f(x) = 7. The reasoning: plug in x = 2, get f(2) = 7. Of course the limit is 7.

Now pick up f(x) = \frac{\sin x}{x} at x = 0. Plug in x = 0 and you get 0/0 — no answer at all. But the limit does exist and equals 1. You computed it in the article on limits by squeezing.

Now pick up a nastier function:

The function wiggles violently as x approaches 0 — infinitely often, in fact. At every scale you zoom in, you find more oscillation. Does this function have a limit at x = 0? Is it continuous there? The intuitive pictures — "f(x) settles down to one number as x closes in" or "you can draw the graph without lifting your pen" — break under this kind of pressure. A function can wiggle infinitely and still have a limit. A function can be defined everywhere and still fail to be continuous in ways no picture reveals.

The response of mathematics, around the 1850s, was to rebuild the foundation. Stop relying on pictures. Turn "close to" into a statement with quantifiers and inequalities — something a computer, or a sceptic, could check. The resulting framework is called real analysis, and this article is your first look at it.

You will still use the pictures — they are indispensable for intuition. But every picture will now have a precise companion: a statement that says what "close" means, in language that does not depend on eyes or hands.

The problem with "close"

The informal definition of a limit says: \lim_{x \to a} f(x) = L means "f(x) gets close to L as x gets close to a."

Listen carefully to what this sentence actually commits to. How close is close? Does "gets close" mean it eventually reaches L? That it comes within 0.1 of L? Within 0.0001? Nothing in the sentence pins it down. The phrase "gets close to" is an action verb without a destination.

The breakthrough — usually credited to Cauchy in the 1820s and made precise by Weierstrass in the 1850s — was to recast the definition as a game between two players.

Player 1 (the sceptic) picks a tolerance. They say: "I want f(x) to be within 0.01 of L. Can you guarantee that?"

Player 2 (the one asserting the limit) picks a distance. They say: "Yes. As long as x is within 0.003 of a (but not equal to a), I promise f(x) is within 0.01 of L."

Then Player 1 tries a tighter tolerance. "What about 0.0001?" Player 2 has to find a new — probably smaller — \delta that works for this tighter \varepsilon. And so on.

The limit \lim_{x \to a} f(x) = L holds exactly when Player 2 can always win — no matter what tolerance Player 1 names, Player 2 can produce a matching distance. There is no cheating, no eventual victory: for every \varepsilon, some \delta must exist.

This is the epsilon-delta definition. In symbols:

Read it slowly.

• \varepsilon > 0 is the tolerance in the output. "f(x) must be within \varepsilon of L."
• \delta > 0 is the tolerance in the input. "Every x within \delta of a is close enough."
• The condition 0 < |x - a| says x \neq a; you are not checking what happens at a, only near it.
• The arrow \Longrightarrow is a promise: if x is close enough in the input sense, then f(x) is close enough in the output sense.
• The universal quantifier ("for every \varepsilon") demands the promise hold for arbitrarily small tolerances, not just convenient ones.

This definition does not contain the word "close" anywhere. It has been replaced by explicit inequalities. Everything that follows in real analysis is built on this one statement.

The geometry of the definition

There is a picture for the epsilon-delta definition, and it is worth drawing.

On the y-axis, mark the interval (L - \varepsilon, L + \varepsilon). This is the "tolerance band" — a horizontal strip of width 2\varepsilon around the horizontal line y = L. The sceptic says: "stay inside this strip."

On the x-axis, you want to find an interval (a - \delta, a + \delta) — the "safe zone" — such that every x in this interval (except possibly x = a itself) has its f(x) inside the strip.

For the limit to exist, you must be able to do this for every horizontal strip the sceptic draws. If the strip gets thinner, you shrink your vertical band to match. If no \delta works for some \varepsilon, the limit fails to exist at a.

A first proof from the definition

Time to use the definition on a specific function. Take f(x) = 3x + 1 and prove \lim_{x \to 2} f(x) = 7.

A few things are worth noticing. First, the proof is constructive — it does not just argue that some \delta exists, it tells you exactly what \delta to pick: \delta = \varepsilon/3. Second, the slope 3 is what shows up in the denominator of \delta. If the slope had been 100, you would need \delta = \varepsilon/100. The steeper the function, the more tightly you have to control the input. Third, you never actually evaluated f at x = 2; the limit is about the values near 2, not at 2 itself.

From limits to continuity

Once you have the epsilon-delta definition of a limit, continuity is just two words.

Notice the differences from the limit definition:

• The condition 0 < |x - a| becomes |x - a| < \delta — you are now allowed to take x = a itself, because you want the actual value f(a) to appear on the right.
• The candidate limit L has been replaced by the actual value f(a). Continuity says: the function's value at a is exactly what the limit suggests it should be — there is no jump, no hole, no surprise.

A function is continuous on an interval I if it is continuous at every point of I. This is the rigorous version of "you can draw the graph without lifting your pen." But unlike the pen picture, the rigorous version handles weird cases. The function x \sin(1/x) extended by f(0) = 0 is continuous at 0 — even though no pen on Earth could literally draw its infinite wiggles — because for every \varepsilon > 0, you can pick \delta = \varepsilon and the definition works:

The picture lies (or at least, cannot be drawn). The definition does not.

A harder example

Take f(x) = x^2 and prove \lim_{x \to 3} f(x) = 9.

The last comment in the figure caption is the seed of the next big idea. For linear functions, the \delta that works at one point works everywhere. For x^2, the \delta you need depends on where you are — the function is steeper far from the origin, so the same output tolerance requires a tighter input tolerance. That dependence on location is what motivates the next definition.

Uniform continuity

Continuity at a point is a local condition: for each point a, you need to produce a \delta that depends on both \varepsilon and a. Usually it also depends on a, not just \varepsilon.

Sometimes, though, you can do better. Sometimes one \delta works for every point a in the domain at once. That stronger condition has its own name.

The crucial change from ordinary continuity: the \delta depends only on \varepsilon, not on the location. A single \delta must work everywhere on the interval simultaneously.

Compare the two definitions carefully:

• Continuity at each point. For each a and each \varepsilon > 0, there exists a \delta > 0 (which may depend on a) such that |x - a| < \delta implies |f(x) - f(a)| < \varepsilon.
• Uniform continuity. For each \varepsilon > 0 there exists a \delta > 0 (which does not depend on a) such that for every pair x, y, |x - y| < \delta implies |f(x) - f(y)| < \varepsilon.

The quantifier order has changed. In ordinary continuity, "for each a" comes before "there exists \delta" — so \delta can depend on a. In uniform continuity, "for each \varepsilon" comes first and then "there exists \delta" — and only after that do you quantify over points.

An example where the two definitions part ways. The function f(x) = x^2 is continuous on all of \mathbb{R}, but not uniformly continuous on \mathbb{R}.

Why? Consider the pair of points x_n = n and y_n = n + \frac{1}{n}. Their distance is |x_n - y_n| = 1/n, which goes to zero as n \to \infty. But

So even when you shrink the input distance to arbitrarily small values by going far out on the real line, the output distance stays above 2. No single \delta can make the output tolerance smaller than, say, \varepsilon = 1 for all pairs. So x^2 is not uniformly continuous on \mathbb{R}.

The reason is that x^2 gets steeper and steeper as x grows. Far from the origin, even tiny input changes produce large output changes. Uniform continuity fails when a function can get arbitrarily steep.

When does uniform continuity hold? A famous theorem, which you will prove in a real analysis course, says: if f is continuous on a closed, bounded interval [a, b], then f is automatically uniformly continuous on [a, b]. Closed-and-bounded is sometimes called compact. Compact intervals are special: continuity on them is always uniform. That is why x^2 fails on the whole real line but works fine on [0, 10] — on a bounded interval, the slope cannot run away to infinity.

Common confusions

• "Epsilon and delta are specific numbers." They are not. \varepsilon is an arbitrary positive number chosen by the sceptic; \delta is a positive number chosen in response. Both are "for every" / "there exists" variables, not constants. The definition describes a relationship between them, not their values.

• "If you can find one \varepsilon and \delta that work, the limit exists." No. You need the property to hold for every \varepsilon. Finding a \delta that works for \varepsilon = 0.1 tells you nothing; the definition demands a \delta for \varepsilon = 0.0001, for \varepsilon = 10^{-100}, and for every positive number in between.

• "Continuity means the graph has no jumps." That is an informal picture. The rigorous definition handles jumps, but it also handles cases no picture can show (like x \sin(1/x)). Trust the definition over the picture.

• "Uniform continuity is the same as being continuous at every point." Very different. A function can be continuous at every point of an interval without being uniformly continuous on it. x^2 on \mathbb{R} is the standard counterexample. Uniform continuity is a stronger, global condition.

• "The \delta has to be the largest one that works." It does not. If \delta = 0.1 works, then any smaller positive number also works. You just need some valid \delta; you do not need the optimal one. In proofs, you take the most convenient \delta that makes the algebra go through.

Going deeper

Sequential continuity

There is an equivalent way to state continuity that uses sequences instead of epsilon-delta. A function f is continuous at a if and only if: for every sequence x_n \to a, the sequence f(x_n) \to f(a).

This is sometimes more convenient for proofs. Instead of chasing down an explicit \delta, you argue: let x_n be any sequence approaching a; show f(x_n) approaches f(a). The two formulations are logically equivalent — a theorem of real analysis, proved using the completeness of the real numbers.

Sequential continuity is often how you detect failure of continuity. If you can find a sequence x_n \to a for which f(x_n) does not approach f(a), the function is not continuous at a.

Why "real" analysis?

The word "real" in real analysis refers to the real numbers — as opposed to complex numbers, which give rise to complex analysis. But it also hints at something deeper: the definitions above are really about the real number line, with its ordering, its absolute value, and its completeness property.

The completeness property says: every non-empty set of real numbers that has an upper bound has a least upper bound. This single property is what makes the real numbers work — it is why limits converge, why continuous functions on [a, b] attain their maximum, why the intermediate value theorem is true. Before the nineteenth century, mathematicians had been using these facts on faith; completeness made them provable.

You will meet completeness in a first course on real analysis. It is the soil everything else grows in.

Historical context, briefly

The shift from informal calculus to rigorous analysis took about a century and a half, starting from Cauchy in the 1820s. Two centuries of spectacularly successful but unrigorous calculus had accumulated a list of paradoxes and counterexamples — functions that were continuous everywhere but differentiable nowhere, infinite series that could be rearranged to converge to any desired sum, "theorems" that were actually false for exotic functions.

Each counterexample forced the definitions to be sharpened. The epsilon-delta definition and its descendants were the end of that process — a foundation rigorous enough that the exotic functions were handled cleanly and no new counterexamples broke things.

Two centuries on, real analysis underlies not only calculus but also probability theory, functional analysis, the mathematical side of physics, and the theoretical foundations of machine learning. Every single theorem in those fields traces its rigour back to the definition in the box above.

Where this leads next

You now know how real analysis approaches the definitions you have been using intuitively. The next articles build on this foundation to tackle more sophisticated objects — functions that are themselves limits of other functions, and power series that express complicated functions as polynomials.

• Sequences and Series of Functions — a new object to analyse: a sequence of functions f_n. Does it approach some limit function f? In what sense? Pointwise or uniformly? This is where epsilon-delta really earns its keep.
• Taylor and Maclaurin Series — writing familiar functions as infinite polynomials. The convergence theorems all depend on the definitions in this article.
• Limit — revisit the earlier intuitive article with your new rigorous eyes. Everything you already knew is still true; it now has a foundation.
• Continuity — the same story for the informal continuity article.
• Uniform Continuity — the full treatment of the property introduced briefly above, including the proof that continuity on a compact interval is automatically uniform.