Having two Ito processes

$dX_{t} =z_{1} dt + Y_{t} dB_t $

$dX^{'}_{t} =z^{'}_{1} dt + Y^{'}_{t} dB_t $

I am analyzing a proof of the product rule

$d(X_t X_t^{'})=X_t dX_t^{'}+ X_t^{'} dX_t + Y_t Y_t^{'}dt$

In the equation

$df(X_t)= \frac{\partial f}{\partial X_t} \partial X_t + \frac{1}{2} \frac{\partial^2 f}{\partial X_t^2} (dX_t)^2$

$(dX_t)^2$ was replaced with $Y_t^2 dt$. I don't quite follow this transition.

This property was used later in the proof as

$d(X_{t} + X^{'}_{t})^2 = (Y_t + Y^{'}_{t})^2 dt$

Could anybody clarify me the substitution of $(dX_t)^2$ with $Y_t^2 dt$?

  • $\begingroup$ where do you have this from? I tried to apply Ito but I don't think that his is true. Where does the factor 2 go? Is this homework? $\endgroup$
    – Richi Wa
    Commented Apr 26, 2016 at 7:17
  • $\begingroup$ your equality is false $\endgroup$ Commented Apr 26, 2016 at 8:56
  • 1
    $\begingroup$ Note that $d\langle X+X', X+X'\rangle_t = (Y_t+Y'_y)^2dt$. $\endgroup$
    – Gordon
    Commented Apr 26, 2016 at 13:08
  • $\begingroup$ It indeed seems as the textbook you are using uses the notation $(dX_t)^2$ instead of the usual $d\langle X \rangle_t$ for the quadratic variation. I am not particulary fond of the former notation, which I find very confusing... $\endgroup$
    – Quantuple
    Commented Apr 26, 2016 at 13:38

1 Answer 1


$X_t$ being a stochastic process, one cannot use ordinary calculus to express the differential of a (sufficiently well-behaved) function $f$ of $t$ and $X_t$.

Instead one should turn to Itô's lemma, one of the key results of stochastic calculus, which stipulates (assuming $X_t$ is here a continuous, square integrable stochastic process) $$ df(t,X_t) = \frac{\partial f}{\partial t}(t,X_t) dt + \frac{\partial f}{\partial x}(t,X_t) dX_t + \frac{\partial^2 f}{\partial x^2}(t,X_t) d \langle X,X\rangle_t $$ where the quantity $$ \langle X,X \rangle_t $$ represents the quadratic variation of the process $X_t$ over $[0,t]$ defined as $$ \langle X,X\rangle_t := \lim_{\Vert P \Vert \rightarrow 0} \sum_{i=1}^N (X_{t_{i}}-X_{t_{i-1}})^2 $$ with $P$ representing a generic partition $\{t_0 = 0 < \dots < t_N = t\}$ of the interval $[0,t]$ and $\lim_{\Vert P \Vert \rightarrow 0}$ suggests the limit (when it exists) is taken in probability as $\max(\{t_{i}-t_{i-1} \vert i=1,\dots,N\}) \rightarrow 0$.

For Itô processes, that is, stochastic processes of the form $$X_t = X_{0} + \int_0^t \mu(t,X_t) dt + \int_0^t \sigma(t,X_t) dW_t$$ or equivalently in differential form (this is an abusive notation, essentially used for convenience) $$dX_t = \mu(t,X_t) dt + \sigma(t,X_t) dW_t$$ where $X_t$, $\mu(...)$ and $\sigma(...)$ are adapted (generally to the natural filtration of $W_t$) and the integrands verify the usual integrability conditions, it can be demonstrated (cf. any good stochastic calculus book) that: $$ \langle X,X \rangle_t = \int_0^t \sigma^2(t,X_t) dt $$ or in differential form (this is an abusive notation, essentially used for convenience) $$ d\langle X,X \rangle_t = \sigma^2(t,X_t) dt $$

In your case, because $X_t$ is the unique solution of the SDE $$ dX_t = z_1 dt + Y_t dW_t $$ applying the above result with $\mu(t,X_t) = z_1$ and $\sigma(t,X_t) = Y_t$ $$ d\langle X,X \rangle_t = Y_t^2 dt $$ assuming the usual conditions are met.

As far as the second issue is concerned, as @Gordon mentioned, using the commutativity + bilinearity of quadratic variations: \begin{align} d\langle X + X^{'} , X + X^{'} \rangle_t &= d \langle X, X \rangle_t + 2 d \langle X, X^{'} \rangle_t + d \langle X^{'} , X^{'} \rangle_t \\ &= Y_t^2 dt + 2 Y_t Y_t^{'} dt + (Y_t^{'} )^2 dt \\ &= (Y_t + Y_t^{'})^2 dt \end{align}

  • $\begingroup$ in terms of the notation is the $d \langle X,X\rangle_t$ the same as $(dX_t)^2$? assuming equal increments $t_{i}-t_{i-1}$ the quadratic variation = variance of the increments = t, correct? $\endgroup$
    – Michal
    Commented Apr 26, 2016 at 22:16
  • 1
    $\begingroup$ @Michal, $d\langle X,X \rangle_t$ can indeed be thought of as the local conditional variance of $X$ at $t$, ie the variance of the increment $dX_t $ conditionally on information available at $t$. However one should not confuse quadratic variation $\langle X,X \rangle_t$ and variance of $X_t $ (usually denoted $\text{var}(X_t)$). $\endgroup$
    – Quantuple
    Commented Apr 27, 2016 at 7:21
  • 1
    $\begingroup$ The former is a process, while the latter is just a number (statistical measure of the dispersion of a random variable). Compare the processes solutions of $dX_t=dB_t $ and $dY_t = -aY_t dt + dB_t $ for instance. Both exhibit the same quadratic variation over any fixed interval, but $\text {var}(X_t) = t\ (= \langle X,X\rangle_t) $ and $\text {var}(Y_t)=\frac {1-e^{-2at}}{2a}$. $\endgroup$
    – Quantuple
    Commented Apr 27, 2016 at 7:27
  • 1
    $\begingroup$ The last 3 lines are the correct answer to the above - and well explained together with the intro! $\endgroup$
    – Richi Wa
    Commented Apr 27, 2016 at 8:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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