W(x)=∑n=0∞ancos(bnπx) — continuous everywhere yet differentiable nowhere. The M-test scene powers the continuity argument, and the graph turns thornier as partial sums grow. The full nowhere-differentiability proof is idea-level only.
Prove via the Weierstrass M-test that the Weierstrass function W(x)=∑n=0∞ancos(bnπx) (a=21, b=13) is continuous on all of R. (Hint: the m-test scene in the series chapter.)
un(x)=ancos(bnπx),∣un(x)∣≤(21)n=:Mn
n=0∑∞Mn=2<∞M-test∑un converges uniformly on R
∣W(x)−WN(x)∣≤n=N+1∑∞Mn=(21)Nfor every x
n=6: tail bound (1/2)6=0.015625 — this curve sits within 0.016 of the true W at every x (W6(0)=1.984375).Show proof
Each term un(x)=ancos(bnπx) is continuous. Since ∣cos∣≤1, ∣un(x)∣≤an=(21)n=:Mn for every x, and ∑n=0∞Mn=∑n=0∞(21)n=2<∞. By the Weierstrass M-test (the m-test scene in the series chapter), ∑un converges uniformly on R.
A uniform limit of continuous functions is continuous: fix x0; then ∣W(x)−W(x0)∣≤∣W(x)−WN(x)∣+∣WN(x)−WN(x0)∣+∣WN(x0)−W(x0)∣. The first and last terms are small by the tail bound (1/2)N, the middle one by the continuity of the finite sum WN. Hence W is continuous at x0, and x0 was arbitrary, so W is continuous on all of R.
Check the numbers: at N=6, ∣W(x)−W6(x)∣≤(1/2)6=0.015625 for every x — the curve on screen sits within 0.016 of the true W. Raising N drives the tail bound (1/2)N to 0, so the on-screen WN approaches W uniformly.
State the theorem: if b is odd and ab>1+23π, then the Weierstrass function W is differentiable at no point (our a=21, b=13 satisfy ab=6.5>5.712). The argument below is an idea-level sketch — it conveys the mechanism faithfully but does not reproduce the classical proof's constant-tracking line by line.
ab=6.5>1+23π≈5.712,b=13 odd
at scale b−n: the n-th harmonic completes a full spike, so some xn→x0 has
xn−x0W(xn)−W(x0)≳(ab)n→∞(idea sketch, not the full proof)
n=8: tail bound (1/2)8=0.00390625 — eight smooth terms added, yet the curve is jagged at every scale (W8(0)=1.99609375, W8(1)=−1.99609375).Show proof
Theorem (Weierstrass–Hardy): if b is an odd integer and ab>1+23π, then W(x)=∑n=0∞ancos(bnπx) is differentiable at no point. We use a=21, b=13, so ab=6.5>1+23π≈5.712 — the condition holds.
The mechanism — zooming in at scale b−n: the n-th harmonic cos(bnπx) completes a full period over every interval of length about b−n, so near any point x0 there is a spike point xn with ∣xn−x0∣≲b−n where the n-th cosine swings nearly trough-to-peak (about 2). Across this spike, the n-th term's difference quotient is an⋅(≈2)/(≈b−n)≈2(ab)n=2⋅6.5n→∞ (about 6.37×106 at n=8).
The remaining terms cannot beat the spike: the terms with k<n are nearly flat at this scale (differentiating brings out akbkπ, a geometric sum with ratio ab and a strictly smaller constant than the spike's), and the terms with k>n are uniformly tiny by the tail bound 2an+1/(1−a). The classical proof chooses the spike points delicately, tracks all three constants, and shows that ab>1+23π is exactly the inequality saying "the spike's constant wins."
Honest label: this is an idea-level sketch. The full proof is the classical Weierstrass–Hardy construction — building the spike points and tracking all three constants to the end — which we do not reproduce line by line here. The takeaway it conveys is this: the difference quotient blows up to infinity along the spike sequence xn→x0, so no finite number can serve as W′(x0); x0 was arbitrary, so W is differentiable nowhere.
Each partial sum WN is a finite sum of cosines, hence smoothly (infinitely) differentiable on R. Yet WN→W uniformly (problem 1) while W is differentiable nowhere (problem 2). Conclude that a uniform limit of differentiable functions need not be differentiable.
WN(x)=n=0∑Nancos(bnπx)smooth on R
WNuniformlyW,∣W−WN∣≤(21)N→0
W differentiable nowhere (problem 2)
n=2: a sum of three smooth terms (W2(0)=1.75, tail bound (1/2)2=0.25) — a uniform limit of such smooth things can fail to be differentiable.Show proof
Each WN is a finite sum, so termwise differentiation is legitimate: WN′(x)=−∑n=0Nanbnπsin(bnπx) exists at every x. For instance W2′(x)=−πsin(πx)−213πsin(13πx)−4169πsin(169πx) (since a1b1π=213π and a2b2π=4169π). Every curve on screen is a smooth function.
But by problem 1's tail bound ∣W−WN∣≤(1/2)N→0, we have WN→W uniformly on R. And by problem 2, the limit W is differentiable at no point.
Hence a uniform limit of differentiable functions need not be differentiable — continuity is preserved under uniform limits (problem 1's argument), but differentiability is not. This is why analysis demands stronger convergence (e.g. uniform convergence of the derivatives) before differentiating termwise.