1
$\begingroup$

The quadratic variation of a function $f : [a,b] \to \mathbb R$ is defined to be $$ \sup_P \sum_{i=1}^n (f(x_i) - f(x_{i-1}))^2 $$ where the supremum is taken over all partitions $P$ such that $a = t_0 < t_1 < \ldots < t_n = b$. If $f : [a,b]$ is continuously differentiable, then its quadratic variation vanishes. I know examples of non-differentiable functions where the quadratic variation is non-zero, for example a typical path of a brownian motion.

Do you know an example of a differentiable function $f : [a,b] \to \mathbb R$ which has non-zero quadratic variation?

$\endgroup$
3
  • 3
    $\begingroup$ You can't take the supremum over all partitions, otherwise no non-constant function could have zero quadratic variation. Are you taking maybe the $\limsup$ as the mesh of the partition tends to $0$? $\endgroup$ – Daniel Fischer Jul 8 '15 at 23:55
  • $\begingroup$ Yes, thank your for pointing out! I just generalised the definition for total variation without thinking about this case, for total variation by the triangle inequality for finer partitions the sum increases, But in the quadratic case $(f(a) - f(c) + f(c) - f(b))^2 = (f(a)-f(c)^2 + (f(c) - f(b))^2 - 2(f(a)-f(c))(f(c)-f(b))$, so it could decrease and therefore just taking the sup does not works here. $\endgroup$ – StefanH Jul 9 '15 at 13:17
  • $\begingroup$ Just one comment: The definition you gave is that of $2$-variation. If you consider paths of a stochastic process, then the paths of the quadratic variation process will be smaller than or equal to the $2$-variation of the paths in value. However, the converse may be false. The classical example here, too, is Brownian motion. See also here: en.wikipedia.org/wiki/P-variation. $\endgroup$ – Kehrwert Jul 18 '20 at 8:17
2
$\begingroup$

Since it's not made clear yet in the question I will here assume the definition $$Q(f,[a,b]) \equiv \limsup\limits_{|\Pi|\to 0}\sum_\Pi (f(x_k)-f(x_{k+1}))^2$$ for the quadratic variation where $\Pi$ denotes a partition of $[a,b]$ and $\limsup$ is taken over all partitions with a given mesh size $|\Pi|$.


For the differentiable function $$f(x) = \left\{\matrix{x^2\sin\left(\frac{1}{x^4}\right) & x \not= 0\\0 & x = 0}\right.$$

and the partition

$$x_{k} = \left\{\matrix{\frac{1}{\sqrt[4]{2\pi(n+k) + \frac{\pi}{2}}} & k\text{ even }\\\frac{1}{\sqrt[4]{2\pi(n+k)}} & k\text{ odd }}\right.$$

we get

$$Q(f,[0,1]) \geq \lim_{n\to\infty}\sum_{k=0}^n(f(x_k)-f(x_{k+1}))^2 = \frac{1}{2\pi}\int_0^1 \frac{{\rm d}x}{1+x} = \frac{\log 2}{2\pi} > 0$$

$\endgroup$
6
  • $\begingroup$ Just for to record, because it took me a while to see that the sums are indeed Riemann-sums for the integral. We have $f(x_k) = 0$ if $k$ is odd, and $f(x_k) = 1/\sqrt{2\pi(n+k)+pi/2}$ otherwise, therefore (suppose $n$ is even, otherwise on extra term needs to be added at the end) <see next comment> $\endgroup$ – StefanH Jul 9 '15 at 11:44
  • $\begingroup$ \begin{align*} \sum_{k=0}^n (f(x_k) - f(x_{k+1})^2 = \sum_{k=0}^n \left(f(x_{2\left \lfloor \frac{k+1}{2} \right\rfloor}\right)^2 & = f(x_0)^2 + \sum_{i=1}^{n/2} 2f(x_{2i})^2 \\ & = \frac{1}{2\pi}\frac{1}{n + \frac{1}{4}} + \frac{1}{2\pi} \sum_{i=1}^{n/2} \frac{2}{n+2i + \frac{1}{4}} \\ & = \frac{1}{2\pi}\left( \frac{1}{1+1/4n}\frac{1}{n} + \sum_{i=1}^{n/2} \frac{1}{1+2i/n+1/4n}\cdot \frac{2}{n}\right) \end{align*} The last sum is a Riemann-sum for a partition where the first interval has length $1/n$ and all the other $2/n$, and the point $2i/n+1/4n$ is selected from each. $\endgroup$ – StefanH Jul 9 '15 at 12:11
  • $\begingroup$ Okay, I see this argument is not working, that are not Riemann-sums, as the "partitions" are to large, try $n=2$ for example, then we have $[0,1/n]$, and $[1/n, 3/n]$ as intervalls, which is by $1/n$ to big, and this is always the case, but luckily this "overlaying" part vanishes for $n\to\infty$. More precisely: \begin{align*} & \frac{1}{1+1/4n}\frac{1}{n} + \sum_{i=1}^{n/2} \frac{1}{1+2i/n+1/4n}\cdot \frac{2}{n} \\ & = \frac{1}{1+1/4n}\frac{1}{n} + \sum_{i=1}^{n/2-1} \frac{1}{1+2i/n+1/4n}\cdot \frac{2}{n} + \frac{1}{1+1+\frac{1}{4n}} \cdot \frac{2}{n} \\ \end{align*}Because <next comment> $\endgroup$ – StefanH Jul 9 '15 at 13:02
  • $\begingroup$ $$\frac{1}{1+1+\frac{1}{4n}} \cdot \frac{2}{n} = \frac{1}{2n}\left( \frac{4}{2+1/4n}\right) = \frac{1}{2n}\left( 1 + \frac{4 - (2+1/4n)}{2+1/4n}\right) $$ we can write the last sum above as $$ \frac{1}{1+1/4n}\frac{1}{n} + \sum_{i=1}^{n/2-1} \frac{1}{1+2i/n+1/4n}\cdot \frac{2}{n} + \frac{1}{1+1}\frac{1}{n} + \frac{1}{2n}\cdot \frac{4 - (2+1/4n)}{2+1/4n}. $$ Now the last three terms consitute indeed a Riemann-sum on $[0,1]$, where in the last interval $[(n-1)/n,1]$ the point $1$ was selected. The fouth summand goes to zero as $n \to \infty$, therefore the whole expression goes to the integral. $\endgroup$ – StefanH Jul 9 '15 at 13:02
  • $\begingroup$ @Stefan Very good. It was so messy adding all the fine details so I left it out, but it's nice to have them here:) $\endgroup$ – Winther Jul 9 '15 at 21:12

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.