A graph drawn with finitely many line segments has a derivative except at finitely many corners. Polynomials are smoother still, and they approximate every real continuous function on a compact interval uniformly. It is tempting to conclude that differentiability is therefore typical among continuous functions. The conclusion does not follow: density says that every neighborhood contains a smooth function, not that most points of that neighborhood share its regularity.
This chapter makes a different notion of size precise. In the complete uniform-norm space , a dense set consists of functions whose local absolute difference quotients are unbounded at every point. In particular, the functions having a finite derivative even somewhere form a meager set. The proof will use only polygonal approximation, compactness of the interval and Baire's theorem. It does not require a formula for a special nowhere differentiable function.
Motivation
Uniform distance controls vertical error. It does not control how rapidly a graph changes over a short horizontal distance. A triangular oscillation of height can have slope if its base has length . Keeping small and making smaller allows a tiny uniform perturbation to create very large slopes.
A single such perturbation does not destroy differentiability everywhere. A polygon still has finite slopes on its segments. The useful observation is that, for each fixed slope threshold, polygons with sufficiently steep segments can avoid a particular closed set. Baire's theorem lets us avoid all the thresholds simultaneously in a limit. The distinction between one threshold and countably many thresholds is the center of the argument.
The word typical below always refers to Baire category in a stated topology. We put no probability distribution on the function space.
Definitions
Definition 2.3.1
(The ambient space and category)
Let , with distance . A set is nowhere dense if its closure has empty interior, and meager if it is a countable union of nowhere dense sets. A set is comeager if its complement is meager. A set is a countable intersection of open sets.
The space is complete. Indeed, a uniformly Cauchy sequence has a real pointwise limit because is complete; the Cauchy estimate passes to that limit and gives uniform convergence. The uniform limit is continuous. Baire's theorem therefore applies: every countable intersection of open dense subsets of is dense in .
For , define the anchored slope bound
This bound concerns chords drawn from one fixed anchor . It does not assert a Lipschitz estimate between every pair of points.
Definition 2.3.2
(Closed classes to be avoided)
For each positive integer , set
Thus means that every anchor has at least one chord with absolute slope greater than . The witness chord may depend on .
At endpoints, limits of difference quotients are taken from within . A finite derivative there means the corresponding finite one-sided derivative. At interior points we mean the usual finite two-sided derivative.
Theorem/proposition
Theorem 2.3.3
(A dense of unbounded local slopes)
The set
is a dense and is comeager in . Each of its functions has no finite derivative at any interior point and no finite one-sided derivative at either endpoint. Consequently, the set of continuous functions possessing a finite derivative at at least one point is meager.
Corollary 2.3.4
(Rough functions can satisfy finitely many prescribed values)
Fix and real values . In the nonempty affine space
with its inherited uniform metric, is a dense and is comeager. Thus any continuous interpolant can be approximated uniformly, with all those sample values unchanged, by a function with unbounded local absolute slopes everywhere.
The theorem proves a stronger local-slope condition than merely failure of a finite derivative. It does not classify all possible infinite one-sided derivatives or signed upper and lower Dini derivatives.
Proof sketch or proof idea
A finite derivative implies a finite anchored bound
Proof
(A finite derivative gives a finite anchored bound)
Suppose has a finite derivative at . By the definition of the derivative, there are and such that
For the remaining points, compactness gives boundedness of , and
Thus is finite, and for some integer . The same argument applies to a finite one-sided derivative at an endpoint.
More generally, any finite local bound on the absolute quotients gives a finite global anchored bound by this far-point estimate. Therefore is equivalent to the displayed local limsup being infinite. Large global anchored quotients must occur arbitrarily close to the anchor: away from it they are bounded by the same estimate. This proves the equality of the two descriptions of .
The slope classes are closed: the anchor is allowed to move
Proof
(Closedness of the slope classes)
Suppose and uniformly. Choose anchors with
The compact interval supplies a subsequence with . Along it,
For an arbitrary fixed , passage to the limit in the anchored inequality gives . The same limiting anchor works for every , because the original inequality held for all along the chosen subsequence. Hence and is closed.
Proof X-ray
Two kinds of convergence control the moving anchor
Uniform convergence controls the replacement of by even at the moving point . Continuity of the fixed limit then controls the replacement of by its limit . Compactness belongs to the interval of anchors, not to a ball in the infinite-dimensional function space.
Every class has empty interior: add small, steep tents
Proof
(Steep tents give empty interior)
Fix and . Uniform continuity allows a polygonal interpolant with : choose a sufficiently fine finite partition and join adjacent function values by straight lines. On each cell, the interpolant is a convex combination of the two endpoint values, both close to the value at an interior point.
Let be the largest absolute slope of the finitely many segments of . Choose and a cell-length bound so small that
Subdivide each original segment into finitely many equal cells, with actual cell length . Different original segments may use different cell lengths. On each cell, let be the triangular tent that is zero at both ends and equals at the midpoint. These tents join continuously. Their slopes are and , and .
Set . On each of its finitely many linear pieces, its slope has absolute value at least . For any anchor , choose a different point in one linear piece meeting , as close to as necessary. If is a vertex, choose a one-sided adjacent piece; if it is an endpoint, choose the piece inside the interval. The chord slope is the slope of that piece, so it exceeds in absolute value. Thus .
Finally, . Every ball therefore meets . Since is closed, is open and dense, and is nowhere dense.
Apply Baire and retain every threshold
Proof
(Baire gives a dense comeager class)
Baire's theorem makes dense. It is a by definition, and its complement is the countable union of the closed nowhere dense sets . It is therefore comeager. A function with a finite derivative somewhere belongs to that meager union by the first step, so the set of all such functions is itself meager: a subset of a nowhere dense set is nowhere dense, and a subset of a countable union can be decomposed into the corresponding intersections.
It is useful to visualize the existence argument as a sequence of polygons without treating one polygon as the final answer. Start inside any prescribed open ball. Choose a polygon in that ball and in , and a closed ball of positive radius around it still contained there. Inside its open interior choose a polygon , then a smaller closed ball lying in the old interior and . Continue, choosing radii at most at stage . Density supplies the next polygon, and openness leaves a positive margin for the next ball.
The centers are uniformly Cauchy. Completeness gives a continuous uniform limit, and closedness of the nested balls keeps it in every earlier . Every individual polygon still has finite slopes, but the limit avoids every finite threshold. Derivatives do not pass through this uniform limit. This is the same completeness mechanism as in Baire's theorem; none of the function-space balls was assumed compact.
Proof
(Interpolation constraints survive the tents)
For the interpolation corollary, is closed, since evaluations are continuous, and is nonempty because a polygon interpolates the data. It is therefore complete. For , choose the approximating polygon with a partition containing all . Its values at those nodes are exactly . Add tents that vanish at every partition node. The perturbation preserves every prescribed value and still avoids . Thus is relatively open and dense. Apply Baire inside this complete affine space.
Worked examples
Worked example 2.3.5
(1. An anchored bound is weaker than a Lipschitz constant)
For and ,
Consequently , where the value at is a supremum approached as . At the anchor zero, , so . Yet the best global Lipschitz constant is two, since chord slopes can approach two near one. Membership of asks for one suitable anchor, not a global Lipschitz constant of one.
Worked example 2.3.6
(2. A single fine polygon avoids one threshold)
Take a uniform mesh of cell length and a tent of height on each cell. Its slopes are . Choosing produces a function in whose uniform distance from zero is only .
Nevertheless this function is globally Lipschitz with constant , so it belongs to for every integer . It is differentiable in the interior of each segment. Increasing the number of small tents proves density of each individual ; only the Baire intersection produces a function avoiding all .
Worked example 2.3.7
(3. Smooth and rough approximations coexist)
Given and , real Stone–Weierstrass gives a polynomial with . The category theorem gives with as well. Thus every uniform neighborhood contains both a polynomial and a function with no finite derivative anywhere.
These two density statements do not conflict. The functions differentiable somewhere form a meager set, while is comeager. Density alone cannot decide category. Nor does comeagerness supply a probability: there is no probability measure in this argument.
Counterexample mode
Bounded anchored quotients do not imply a derivative
Set and for . It is continuous on , and , so . Its right difference quotient at zero is . Along it equals one, while along it equals minus one; both sequences tend to zero. Hence no right derivative exists there.
The function is not globally Lipschitz either. At , its classical derivative is . A globally Lipschitz function has derivative bounded by its Lipschitz constant wherever that derivative exists. These unbounded values rule out a global constant. Thus neither global Lipschitz regularity nor differentiability at the anchor can be read from the definition of .
Common mistakes
Common mistake
(Avoid bounded slopes at every anchor)
Showing that every neighborhood contains a function with one steep chord is insufficient: one must avoid the condition of a bounded slope at any anchor. The polygon construction checks every point, including vertices and endpoints. Likewise, the anchor cannot be held fixed when proving closedness of ; it may vary with the approximating function.
Common mistake
(A category statement names its space and metric)
The theorem concerns the uniform topology on all real continuous functions. It does not say that a typical continuously differentiable function is nowhere differentiable, or that polynomial approximation controls derivatives. The ambient space and its metric are part of every category statement.
Concept lensStructural
Category describes a space together with its topology
The uniform norm records heights while permitting arbitrarily steep small perturbations. A norm that also controls the first derivative excludes those perturbations from a small ball. The same inequality-defined sets can therefore have empty interior in one function space and nonempty interior in another. Baire's theorem is unchanged; the open sets to which it applies have changed.
Common mistake
(The comeager class is not every nowhere differentiable function)
One must also distinguish the theorem's sufficient roughness condition from a complete classification of nondifferentiable functions. Failure of a finite derivative can happen even with bounded absolute difference quotients, as the anchored counterexample shows. We proved that the stronger unbounded-slope class is comeager; we did not assert that every nowhere differentiable function lies in that particular .
Summary
Finite differentiability somewhere forces membership in the union of the closed anchored-slope classes. Small, steep polygonal perturbations show that each class has empty interior. Completeness then allows all thresholds to be avoided simultaneously. The resulting category conclusion is compatible with dense smooth approximations and even survives finitely many fixed interpolation values.
Exercises
Quick check
(Put and . Prove uniform convergence to , with error at most . Explain why this sequence does not converge in the norm on continuously differentiable functions.)
Solution · Answer
For and , . This proves the uniform bound, and each is smooth. The limit has left derivative minus one and right derivative one at . Moreover, . Its pointwise limit is minus one to the left, zero at , and one to the right. The derivatives cannot converge uniformly, because a uniform limit of continuous functions is continuous. They are not uniformly Cauchy either, by completeness of the continuous-function space. Thus the original functions do not converge, or even form a Cauchy sequence, in the stated derivative-controlling norm. Any limit would also have to equal the uniform limit , which is not .
Quick check
(Fix . Prove that finitely many rational points and a number satisfy)
Deduce that the uniform ball of radius around lies in .
Solution · Answer
For a fixed anchor , membership of gives a point with . Continuity in and density of the rationals give a rational witness with the same strict inequality. For each rational , the anchors satisfying that strict inequality form an open set. These sets cover , so compactness yields finitely many witnesses whose open sets still cover. The continuous function is positive everywhere on the compact interval and therefore has a positive minimum . Changing by less than in uniform norm changes each displayed expression by less than . At every anchor at least one strict witness remains, so the new function stays in .
Quick check
(Give a proof that a continuous function passing through , and can be approximated arbitrarily closely in the uniform norm by another such function with no finite derivative anywhere. State exactly which space is complete and which sets are dense.)
Solution · Answer
Use the closed affine subspace . Evaluation maps are continuous, so is complete; a tent through the three data makes it nonempty. Each is relatively open. To prove density, approximate a given member of by a polygon whose partition includes , and add small steep tents that vanish at every partition node. The perturbed polygon remains in and lies in , by the same slope estimate as in the theorem. Baire in makes dense there. Its members satisfy all three values and the unbounded local-slope condition, hence have no finite derivative, with one-sided interpretation at the endpoints.
Quick check
(Let be the real polynomials restricted to , with the uniform metric, and let consist of degrees at most . Prove that each is closed and nowhere dense in . Deduce that is not complete despite being dense in .)
Solution · Answer
In fact is closed even in . Suppose polynomials of degree at most converge uniformly to . Evaluate at distinct points. The coefficient vectors are obtained from those values by one fixed invertible Vandermonde matrix, so they converge to a coefficient vector of degree at most . Coefficient convergence on the bounded interval gives uniform convergence to that polynomial; uniqueness of the uniform limit identifies it with . For any , adding with arbitrarily small nonzero leaves while remaining in . Thus its relative interior in is empty. Since , this nonempty space is meager in itself and cannot be complete by Baire. Stone–Weierstrass gives its density in , which is a separate property.
Quick check
(Prove that is complete in . Show that, for each positive integer , has nonempty interior in that norm. Explain which step of the roughness proof therefore fails in this new space.)
Solution
For a -Cauchy sequence, the functions converge uniformly to a continuous , and their derivatives converge uniformly to a continuous . Pass to the limit in . It follows that , so the fundamental theorem of calculus gives and . Both uniform errors tend to zero, proving completeness in the stated norm. If , then , and the mean value theorem gives for all pairs. Hence the nonempty open ball of radius about zero lies in . These closed classes no longer have empty interior. The polygonal approximants and tents generally have corners and are not , so they are not admissible approximants in this space. Moreover, a differentiable perturbation with a large derivative cannot have small norm here: small height alone does not suffice. The density argument fails; Baire's theorem itself remains valid.