Derivative Bounds and Lipschitz Estimates

From Derivative Information to Quantitative Estimates

The previous tutorial used derivative signs to decide whether a function is strictly increasing or decreasing. Derivative bounds provide more quantitative information: they can limit how far function values separate when inputs are close. This is the purpose of a Lipschitz estimate.

The Corollary (Derivative Bound Implies a Lipschitz Bound) and the Secant-Slope Bounds from Derivative Bounds theorem, established earlier in the course, already give the basic upper estimate. We will use those results rather than prove them again. The new questions are how sharp such a bound can be, what a Lipschitz estimate implies about derivatives, and how a lower derivative bound controls the inverse function.

A Lipschitz estimate compares every pair of inputs in the domain, not just inputs near one fixed point. In particular, it implies uniform continuity: given \(\varepsilon>0\), any \(\delta=\varepsilon/L\) works when \(L>0\); if \(L=0\), the function is constant. Lipschitz continuity is therefore stronger than uniform continuity.

The Least Lipschitz Constant

On an interval, a derivative bound provides an upper bound for every secant slope. Conversely, any Lipschitz constant must bound the derivative wherever the derivative exists. Together, these facts identify the optimal constant exactly.

Proof. In the finite case, the Corollary (Derivative Bound Implies a Lipschitz Bound) gives \(|f(x)-f(y)|\leq M|x-y|\) for all \(x,y\in I\). Thus \(M\) is a Lipschitz constant.

Now let \(L\) be any Lipschitz constant for \(f\), and fix \(a\in I\). For every sufficiently small nonzero \(h\), both \(a\) and \(a+h\) lie in \(I\). The Lipschitz inequality gives

Taking the limit as \(h\to0\), using differentiability at \(a\), yields \(|f'(a)|\leq L\). This holds for every \(a\in I\), so \(M\leq L\). Since every Lipschitz constant is at least \(M\), and \(M\) itself is one, the least constant equals \(M\). In the infinite case, a finite Lipschitz constant \(L\) would imply \(|f'(a)|\leq L\) for every \(a\), contradicting the unboundedness of \(|f'|\). \(\square\)

The supremum \(M\) need not be attained at any point. The theorem concerns the least uniform bound, not necessarily a derivative value that occurs somewhere. If \(M=0\), the estimate says \(|f(x)-f(y)|\leq0\), so \(f\) is constant.

A Lipschitz Estimate Also Bounds Derivatives

The proof of the theorem gives a useful converse that does not require computing a supremum. If a function is \(L\)-Lipschitz and is differentiable at a point, then the magnitude of its derivative there cannot exceed \(L\). This is a pointwise conclusion, so it remains valid even when the function is not differentiable everywhere.

Proof. Since \(a\) is an interior point of \(E\), for all sufficiently small nonzero \(h\), \(a+h\in E\). The Lipschitz inequality gives \(|f(a+h)-f(a)|\leq L|h|\). Dividing by \(|h|\) gives

The difference quotient converges to \(f'(a)\). Continuity of absolute value therefore gives \(|f'(a)|\leq L\). \(\square\)

Lower Derivative Bounds and Inverse Estimates

An upper bound on \(|f'|\) prevents function values from separating too quickly. A positive lower bound on \(f'\), on the other hand, ensures that distinct inputs remain quantitatively separated after applying \(f\). This gives an estimate for the inverse.

Proof. Since \(f'(x)>0\) throughout \(I\), the Positive Derivative Implies Strict Increase theorem shows that \(f\) is strictly increasing, hence one-to-one. Its inverse is therefore defined on \(f(I)\).

Take distinct \(u,v\in f(I)\), and write \(u=f(x)\), \(v=f(y)\). If necessary, exchange the two pairs so that \(x<y\); strict increase then gives \(u<v\). The Mean Value Theorem applies to \(f\) on \([x,y]\): the interval lies in \(I\), \(f\) is continuous there because it is differentiable, and it is differentiable on \((x,y)\). Thus for some \(c\in(x,y)\),

Since \(y-x>0\), this implies \(y-x\leq (v-u)/m\). Recalling that \(x=f^{-1}(u)\) and \(y=f^{-1}(v)\), and restoring absolute values to cover either ordering, we obtain

This is the asserted Lipschitz estimate. \(\square\)

The lower bound must be strictly positive. If the derivative can approach zero, the inverse may still exist, but this theorem supplies no finite Lipschitz constant for it. The conclusion is a global estimate on the entire image \(f(I)\), not merely a local statement near one input.

Using the Bounds Carefully

Upper and lower derivative bounds answer different questions. If \(|f'|\leq M\), then the change in output is at most \(M\) times the change in input. If \(f'\geq m>0\), then the change in output is at least \(m\) times the change in input, and the inverse has Lipschitz constant at most \(1/m\). A lower bound on \(f'\) does not by itself provide an upper Lipschitz bound for \(f\); the derivative may be arbitrarily large.

These results depend on the interval structure of the domain: the Mean Value Theorem compares two inputs by considering the whole segment between them. A derivative bound stated only at scattered points, or on a domain with gaps, cannot automatically be used to compare every pair of inputs. Also distinguish a bound that is attained from a supremum approached only near an excluded endpoint. The least-Lipschitz-constant theorem handles both cases.

In applications, an upper derivative bound quantifies stability: small errors in the input produce controlled errors in the output. A positive lower derivative bound quantifies invertibility and stability in the reverse direction: small changes in the output force the recovered inputs to be close. These estimates are often more informative than continuity alone because they specify an explicit constant.