3
$\begingroup$

When the variables $X, Y$ are independent, then the PDF of $Z = X + Y$ can be computed using convolutions: $$ f_Z(z) = \int_{-\infty}^{\infty} f_X(x)f_Y(z - x) dx $$

When the variables are dependent, apparently you can use $$ f_Z(z) = \int_{-\infty}^{\infty} f_{XY}(x, z - x) dx $$

I am wondering where the expression came from for the dependent case? It looks very similar to the independent case except you can't separate the joint distribution into marginals.

$\endgroup$
5
  • $\begingroup$ The derivation/source for both equations is exactly the same, except for independent r.v.'s you can factor the density as a last step. $\endgroup$ – Aaron Hendrickson May 13 at 21:43
  • $\begingroup$ @AaronHendrickson Hmm based on Snoop's answer belong, I don't see how it's the same. I can derive the independent case purely from conditional probability: $$ F_{Z}(z) = P(X + Y \leq z) = \int_{-\infty}^{\infty} P(X + Y \leq z | X = x) f_X(x) dx \\ = \int_{-\infty}^{\infty} P(x + Y \leq z) f_X(x) dx \\ = \int_{-\infty}^{\infty} P(x + Y \leq z) f_X(x) dx \\ = \int_{-\infty}^{\infty} P(Y \leq z - x) f_X(x) dx \\ = \int_{-\infty}^{\infty} \int_{-\infty}^{z - x} f_Y(y) f_X(x) dydx \\ $$ Then I think we just take the derivative with respect to $z$ and arrive at $f_Z(z)$? $\endgroup$ – student010101 May 13 at 21:48
  • $\begingroup$ @AaronHendrickson Oh I guess based on leonbloy's solution, I could do the same $\endgroup$ – student010101 May 13 at 21:48
  • $\begingroup$ I'm not saying there are the same. What I am saying is that the derivation for both formula are with the exception that in the independent case you can factor the joint density at the end. $\endgroup$ – Aaron Hendrickson May 13 at 21:50
  • $\begingroup$ I am surprized that this question has 5 answers whereas it is addressed in many probability textbooks or on MathsSE here for example with 2 nice short proofs by Did. $\endgroup$ – Jean Marie May 13 at 22:55
5
$\begingroup$

From the cumulative distribution function: $$F_Z(z)= P(Z\le z ) = P(X+Y \le z) = P(Y \le z -X) = \int_{-\infty}^{\infty} \int_{-\infty}^{z-x} f_{XY}(x,y)\, dy\, dx $$

Now, $$f_Z(z) = \frac{d F_Z(z) }{ d z} = \int_{-\infty}^{\infty} \frac{d }{ d z} \left( \int_{-\infty}^{z-x} f_{XY}(x,y) dy\, \right) dx = \int_{-\infty}^{\infty} f_{XY}(x,z-x) dx $$

If $X$ and $Y$ are independent we can further write

$$f_Z(z) = \int_{-\infty}^{\infty} f_X(x) f_Y(z-x) dx $$

$\endgroup$
2
  • 1
    $\begingroup$ My calculus is a little rusty. Could you show the derivative step? Did you just use the fundamental theorem of calculus? $\endgroup$ – student010101 May 13 at 21:50
  • $\begingroup$ Yes, the derivative goes inside the outer integral, and then you use en.wikipedia.org/wiki/Leibniz_integral_rule $\endgroup$ – leonbloy May 13 at 22:56
2
$\begingroup$

The characteristic function of $Z=X+Y$ is $$E[e^{iuZ}]=E[e^{iu(X+Y)}]=\int_{\mathbb{R}^2} e^{iu(x+y)}p_{X,Y}(x,y)dxdy$$ The pdf is found by inverse Fourier Transform. By exchanging integrals $$f_{X+Y}(z)=\frac{1}{2\pi}\int_{\mathbb{R}^2}\bigg(\int_{\mathbb{R}} e^{iu(x+y)}e^{-iuz}du\bigg)p_{X,Y}(x,y)dxdy=$$ $$=\int_\mathbb{R^2}\delta(z-(x+y))p_{X,Y}(x,y)dxdy$$ By symmetry of the Dirac delta $\delta(w)=\delta(-w)$ $$\int_\mathbb{R^2}\delta((x+y)-z)p_{X,Y}(x,y)dxdy=$$ $$(x+y)-z=0 \implies y=z-x $$ $$=\int_\mathbb{R}p_{X,Y}(x,z-x)dx$$


$$P(Z\leq z)=P(X+Y\leq z)=\int P(X+y\leq z|Y=y)p_Y(y)dy=$$ $$=\int_{\infty}^\infty\int^{z-y}_{-\infty}p_{X|Y}(x,y)p_y(y)dxdy=\int_{\infty}^\infty\int^{z-y}_{-\infty}p_{X,Y}(x,y)dxdy$$ $$f_Z(z)=\frac{d}{dz}P(Z\leq z)=\int_{-\infty}^\infty p_{X,Y}(z-y,y)dy$$ By using Leibniz integral rule.

$\endgroup$
2
  • $\begingroup$ This is more involved than I anticipated. Is there a way to just derive this from conditional distribution? $\endgroup$ – student010101 May 13 at 21:42
  • $\begingroup$ added @student010101 $\endgroup$ – Snoop May 13 at 21:51
1
$\begingroup$

There is a property for linear combinations that says if X, Y have joint pdf $f(x, y)$ and $Z=aX+bY+c$, then Z has pdf $$g(z)=\int_{-\infty}^\infty f\left(x, \frac{z-c-ax}b\right)\frac1{|b|}dx$$

For $Z=X+Y$, $$\begin{split}G(z)&=\int_{-\infty}^\infty\int_{-\infty}^{z-x}f(x,y)dydx\end{split}$$

Change of variables $w=x+y$

$$\begin{split}G(z)&=\int_{-\infty}^\infty\int_{-\infty}^{z}f(x,w-x)dwdx\\ &=\int_{-\infty}^z\int_{-\infty}^{\infty}f(x,w-x)dxdw\end{split}$$

Derivative with respect to $z$

$$\begin{split}g(z)&=\left[\frac{d}{dz}(z)\right]\int_{-\infty}^\infty f(x, z-x)dx-0\\ &=1\cdot\int_{-\infty}^\infty f(x, z-x)dx=\int_{-\infty}^\infty f(x, z-x)dx\end{split}$$

$\endgroup$
3
  • $\begingroup$ But how did this come about? $\endgroup$ – student010101 May 13 at 21:40
  • $\begingroup$ Would you like it for the general form or just for X+Y? $\endgroup$ – Stacker May 13 at 21:41
  • $\begingroup$ I think just $X + Y$ $\endgroup$ – student010101 May 13 at 21:41
0
$\begingroup$

A pdf for a random vector $X:\Omega\to\Bbb R^n$ is any function $g:\Bbb R^n\to \Bbb R$ such that $E[h(X)]=\int_{\Bbb R^n}h(x)g(x)\,dx$ for, say, all bounded summable Borel functions $h:\Bbb R^n\to\Bbb R$.

So, if your random vector is $V=(X,Y)$, then, without too many ceremonies, you can use the change of variables $(t,s)= (x,y+x)$ and Fubini-Tonelli in the integral $$E[h(X+Y)]=\int_{\Bbb R^2} h(x+y)f_{X,Y}(x,y)\,dxdy=\int_{\Bbb R^2}h(s)f_{X,Y}(t,s-t)\,dsdt=\\=\int_{-\infty}^\infty f(s)\left(\int_{-\infty}^\infty f_{X,Y}(t,s-t)\,dt\right)\,ds$$

And there you have it.

$\endgroup$

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.