Lipschitz & One-Insight Problems

Convexity closed with a challenge: when the exam hands you a fresh problem and never says which tool to use, how do you find it fast? This last calculus lesson is the sharpest version of that challenge.

Every year, GATE slips in a question that looks brutal on first read. It presents:

• an abstract function f with no formula
• a strange-looking inequality
• a request for one specific number

There is nothing to differentiate, nothing to plug in. It seems unfair.

It is not. These problems are one-trick ponies: each hinges on a single insight that, once you see it, collapses the whole thing to a one-line answer. This lesson teaches the most famous example — and, more usefully, how to recognise the species on sight. That way, the next abstract f does not rattle you.

The star concept here, the Lipschitz bound (a cap on how fast a function can change), is not mere exam trivia. Bounding a function’s or a gradient’s Lipschitz constant is how modern machine learning keeps training stable. It underpins:

• WGAN’s weight clipping
• certified robustness bounds
• the convergence guarantees for gradient descent

The signature problem: a squared bound

Here is the setup that has appeared more than once. You are told a function satisfies

|f(x) − f(y)| ≤ C · (x − y)²      for all real x, y

for some constant C. There is no formula for f. The question asks for something like f(1) − f(0). It looks hopelessly underspecified — but the squared (x − y)² is quietly doing all of the work.

Read the inequality in plain words before touching any algebra: over a stretch of width h, the function is allowed to move by at most C·h².

And h² is punishingly small next to h once h is small — at h = 0.01 the allowance is C/10000, a ten-thousandth rather than a hundredth. A function permitted to move only by the square of the distance travelled can never accumulate any slope at all. Everything below is that observation made precise.

The one idea: divide, then take the limit

The derivative of f at a point x is the limit of the difference quotient — the very definition from the differentiability lesson. So look at that quotient and feed it the bound.

Dividing is the natural move here, and it is worth knowing why: the thing you want — a slope — is a ratio, while the fact you were handed is a plain size comparison. Division is what turns one into the other. For y ≠ x, divide both sides by |x − y|:

| f(x) − f(y) |
| ----------- |  ≤  C · |x − y|
|    x − y    |

The left side is the difference quotient whose limit (as y → x) is |f'(x)|. The right side C·|x − y| goes to 0 as y → x. Squeezed between, the slope has nowhere to go:

0 ≤ |f'(x)| ≤ lim (C·|x − y|) = 0      ⇒      f'(x) = 0   for every x

A function whose derivative is zero everywhere is constant. So f takes the same value at every point, and therefore any difference vanishes:

f'(x) = 0 everywhere   ⇒   f is constant   ⇒   f(1) − f(0) = 0

How GATE asks this

Almost always a NAT whose answer is a clean 0. The wrapper changes. It might ask for:

• f(2) − f(−3)
• “how many such non-constant f exist?”
• the bound as |f(x) − f(y)| ≤ 5(x − y)²

But the engine never changes: divide by |x − y|, let y → x, conclude f' ≡ 0, so f is constant and the difference is 0. The skill GATE is really testing is recognition: an abstract f plus a squared (or higher-power) bound is the tell, every time.

Worked example — a real GATE DA question

A function f satisfies |f(x) − f(y)| ≤ (x − y)² for all real x, y. Find f(1) − f(0).

Run the pattern. Fix x, take any y ≠ x, and divide the given bound by |x − y|:

| f(x) − f(y) |
| ----------- |  ≤  |x − y|
|    x − y    |

Now let y → x. The left side tends to |f'(x)| (the definition of the derivative), and the right side |x − y| → 0. By the squeeze,

|f'(x)| ≤ 0   ⇒   f'(x) = 0      for every x.

So f has zero slope everywhere and is therefore constant. A constant function takes the same value at 1 and at 0, hence

f(1) − f(0) = 0.

The answer is 0 — daunting on first read, a single line once you spot that the squared bound kills the derivative. That recognition, not the algebra, is the mark.

A question to carry forward

That closes the calculus chapter. Step back and notice what every lesson in it shared: a world of smooth, continuous change:

• curves that bend
• slopes that vary by infinitesimal amounts
• quantities approached but never quite reached

It is the mathematics of the analog world. Yet the data scientist’s daily instrument is nothing like a smooth curve. It is a handful of lines of code that must run exactly — discrete, literal, and unforgiving of a single off-by-one.

The next chapter changes register completely, from the continuous to the computational. Here is the thread onward: when you are handed a short Python snippet and asked what it prints, how do you read it the way the machine does? Trace each step precisely enough to predict the output character for character, with no rounding and no “approximately”?

In one breath

• Some GATE calculus problems are one-insight: an abstract f, a strange bound, and a single idea that collapses it to a one-line answer.
• The signature: |f(x) − f(y)| ≤ C(x − y)² ⇒ divide by |x − y|, let y → x, get |f'(x)| ≤ 0 ⇒ f' ≡ 0 ⇒ f constant ⇒ any difference like f(1) − f(0) = 0.
• The answer is almost always a clean 0, and the constant C never matters.
• The square is everything: a linear bound ≤ C|x − y| only gives |f'| ≤ C (bounded slope, Lipschitz) — f(x) = x meets it and is not constant. It needs power > 1.
• The real skill is recognition: abstract f + squared (or higher) bound ⇒ reach for divide-and-limit.

Practice