Skip to content

Brownian Motion Is Nowhere Differentiable

Shrinking secants near a Brownian point.
Brownian motion is continuous, but its paths oscillate too violently to have a tangent line. The right derivative is controlled by the secant quotients
B(t+h)B(t)h,h>0.\begin{equation*}\frac{B(t+h)-B(t)}{h}, \qquad h>0.\end{equation*}
For smooth functions these quotients settle as h0h\downarrow0. For Brownian motion they almost surely fail to settle at every time.
PALEY, WIENER AND ZYGMUND
Almost surely, Brownian motion is nowhere differentiable.
DEFINITION
For a function ff, we define the upper and lower right derivatives
Df(t)=lim suph0f(t+h)f(t)h,\begin{equation*}D^* f(t) = \limsup_{h\downarrow0}\frac{f(t+h)-f(t)}{h},\end{equation*}
and
Df(t)=lim infh0f(t+h)f(t)h.\begin{equation*}D_* f(t) = \liminf_{h\downarrow0}\frac{f(t+h)-f(t)}{h}.\end{equation*}
PALEY, WIENER AND ZYGMUND
Almost surely, Brownian motion is nowhere differentiable. Furthermore, almost surely, for every tt,
DB(t)=+orDB(t)=or both,\begin{equation*}D^*B(t)=+\infty \quad \text{or} \quad D_*B(t)=-\infty \quad \text{or both},\end{equation*}
where
DB(t)=lim suph0B(t+h)B(t)h,DB(t)=lim infh0B(t+h)B(t)h.\begin{equation*}D^*B(t)=\limsup_{h\downarrow0}\frac{B(t+h)-B(t)}{h}, \qquad D_*B(t)=\liminf_{h\downarrow0}\frac{B(t+h)-B(t)}{h}.\end{equation*}
Three forced small increments.
Three forced small increments.
It is enough to prove the statement on [0,1][0,1]. Suppose that for some t0[0,1]t_0\in[0,1] both one-sided derivative bounds are finite. Since Brownian motion is bounded, this implies that for some finite MM,
sup0<h1B(t0+h)B(t0)hM.\begin{equation*}\sup_{0<h\leq1}\frac{|B(t_0+h)-B(t_0)|}{h}\leq M.\end{equation*}
We show that this event has probability zero for every fixed MM.
Fix nn and let t0t_0 lie in the dyadic interval
t0(k12n,k2n].\begin{equation*}t_0 \in \left( \frac{k-1}{2^{n}},\,\frac{k}{2^{n}}\right].\end{equation*}
For j=1,2,3j=1,2,3, the triangle inequality gives
B(k+j2n)B(k+j12n)=B(k+j2n)B(t0)+B(t0)B(k+j12n)B(k+j2n)B(t0)+B(t0)B(k+j12n)M2j+12n.\begin{align*}\left|B(\frac{k+j}{2^{n}})-B(\frac{k+j-1}{2^{n}})\right| & =\left|B(\frac{k+j}{2^{n}})-B(t_0)+ B(t_0) - B(\frac{k+j-1}{2^{n}}) \right| \\&\leq \left|B(\frac{k+j}{2^{n}})-B(t_0)\right| + \left|B(t_0)-B(\frac{k+j-1}{2^{n}})\right| \\&\leq M \frac{2j+1}{2^{n}}.\end{align*}
Define the event Ωn,k\Omega_{n,k} that three neighboring increments are all that small,
Ωn,k:={B((k+j)2n)B((k+j1)2n)M(2j+1)2n for j=1,2,3}.\begin{equation*}\Omega_{n,k}:=\left\{\left|B((k+j)2^{-n})-B((k+j-1)2^{-n})\right|\leq M(2j+1)2^{-n}\text{ for } j=1,2,3\right\}.\end{equation*}
By independence of Brownian increments and Brownian scaling,
P(Ωn,k)j=13P{B(1)M2j+12n/2}P{B(1)7M2n/2}3.\begin{align*}\mathbb{P}(\Omega_{n,k}) \leq\prod_{j=1}^{3} \mathbb{P}\left\{ |B(1)|\leq M \frac{2j+1}{2^{n/2}}\right\}\leq \mathbb{P}\left\{|B(1)|\leq \frac{7M}{2^{n/2}}\right\}^{3}.\end{align*}
Since the standard normal density is bounded by 1/21/2,
P(Ωn,k)(7M2n/2)3.\begin{equation*}\mathbb{P}(\Omega_{n,k})\leq (7M2^{-n/2})^3.\end{equation*}
Therefore
P(k=12n3Ωn,k)2n(7M)323n/2=(7M)32n/2,\begin{equation*}\mathbb{P}\left(\bigcup_{k=1}^{2^n-3}\Omega_{n,k}\right)\leq2^n(7M)^3 2^{-3n/2} = (7M)^3 2^{-n/2},\end{equation*}
which is summable over all nn. Hence, by the Borel-Cantelli lemma,
P{there is t0[0,1] with sup0<h1B(t0+h)B(t0)hM}P(k=12n3Ωn,k for infinitely many n)=0.\begin{equation*}\mathbb{P}\left\{\text{there is } t_0\in[0,1]\text{ with }\sup_{0<h\leq1}\frac{|B(t_0+h)-B(t_0)|}{h}\leq M\right\} \leq\mathbb{P}\left(\bigcup_{k=1}^{2^n-3}\Omega_{n,k}\text{ for infinitely many } n\right) =0.\end{equation*}
Taking the countable union over integer MM proves the theorem. \Box