Let say, we have 2 IIDs $X, Y \sim N \left( 0, 1 \right) + \eta $

Now $\eta$ has a discrete distribution with values 0(-10) with probabilities 0.991 & 0.009 respectively.

Also assume that $Z = X+Y$.

I need to find the value of $z$ such that $P \left[ Z < z \right] = 1\%$

Is there any closed form solution available for the value of $z$?

  • $\begingroup$ There is no closed form solution, but the value is relatively easy to find numerically by finding the (unique zero of the) function $\Pr(Z \lt z) - 1/100.$ $\endgroup$
    – whuber
    Nov 2, 2020 at 23:06

1 Answer 1


Another way to express these quantities is to let $W$ be a standard Normal variable and $U$ be a Bernoulli$(0.009)$ variable. Both $X$ and $Y$ have the distribution of $W - 10U.$ Thus, $Z=X+Y$ has the distribution of (a) the sum of two iid standard Normal variables plus $-10$ times (b) the sum of two iid Bernoulli$(0.009)$ variables.

It is elementary that (a) has a Normal$(0,\sqrt{2})$ distribution and (b) has a Binomial$(0.009, 2)$ distribution. This latter takes on three values $0,1,2$ with chances $(1-p)^2, 2p(1-p),$ and $p^2,$ respectively (writing $p=0.009$). Subtracting $10$ times their value exhibits $Z$ as a mixture of three Normal variables with means $0$, $0-10(1)=-10,$ and $0-20(1)=-20.$ The mixture weights are $(1-p)^2, 2p(1-p),$ and $p^2,$ respectively.

Here is a plot of this mixture distribution (CDF) $F.$ I use a semi-log scale because there's a fairly large range of relevant probabilities:


The three Normal components centered at $0,-10,-20$ are apparent: these are the locations very close to the modes (where the slope of this plot is locally steepest). The red line shows the value $1\% = 0.01.$ The solution you seek is the value $z$ shown by the vertical gray line, situated where the red line intersects the graph.

Evidently, this solution is the zero of the function $z\to F(z)-1/100.$ Find it using any good univariate root finder. With double precision arithmetic you should obtain $z \approx -9.8006135477.$

This approach extends in an obvious manner to finding and working with distributions of the sums of any finite number of finite mixtures: it comes down to adding any pair of mixture components (such as $X$ and $Y$) and adding the discrete mixing variables (in this case, adding two iid copies of $U$).


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.