I'm interested in evaluating the following integral $$ I(n) =\int_{-\infty}^0\int_{-\infty}^{z_2 \sqrt{\frac{3-2}{3}}}\cdots \int_{-\infty}^{z_{n-1}\sqrt{\frac{n-2}{n}}}\exp\left(-\sum_{k=2}^n z_k^2\right)dz_n\cdots dz_3dz_2 $$ where $n\geq 2$. Clearly for $n=2$, this is just (half of) a standard Gaussian integral, and gives $$ I(2) = \frac{\sqrt{\pi}}{2} $$ In addition, Mathematica gives $$ I(3) = \frac{\pi}{3\cdot 2} $$ Comparison with numerical results seems to suggest that in general $$ I(n) = \frac{\pi^{\frac{n-1}{2}}}{n!} $$ I.e. that this simplex captures a fraction (1/n!) of the entire Gaussian weight ($\pi^{\frac{n-1}{2}})$. This seems like a very clean result, but I'm absolutely at a loss for how to prove this. A proof by induction doesn't pan out (unless there's some neat integration trick that simplifes the resulting integral of a Gaussian and an error function...), and I'm not aware of any symmetry arguments that could simplify this. Any ideas are welcome.
Edit: If you'd like to test this in Mathematica, you can use the following function
(*Define the function to compute the integral*)
SimplexIntegral[n_Integer] :=
Module[{vars, integrand, limits, I, sum,
i},(*Create a list of variables z2,z3,...,zn*)
vars = Table[Symbol["z" <> ToString[k]], {k, 2, n}];
(*Define the integrand as the exponential of the sum of squares*)
integrand = Exp[-Sum[vars[[k - 1]]^2, {k, 2, n}]];
(*Set up the integration limits:first from-\[Infinity] to 0,
then based on the previous variable with the sqrt factor*)
limits =
Join[{{vars[[1]], -Infinity, 0}},
Table[{vars[[i]], -Infinity,
vars[[i - 1]] Sqrt[(i - 1)/(i + 1)]}, {i, 2, Length[vars]}]];
(*Perform the numerical integration*)
I = NIntegrate[integrand, Evaluate[Sequence @@ limits],
Method -> "GlobalAdaptive", MaxRecursion -> 15,
PrecisionGoal -> 5, WorkingPrecision -> 25];
(*Return the result*)I]
and you can plot the comparison via
DiscretePlot[(Factorial[n] SimplexIntegral[n])/\[Pi]^((n - 1)/2), {n, 2, 5}]
An Idea: Due to the spherical symmetry of the Gaussian kernel, we can view $I(n)$ as the total Gaussian weight times the ratio $$ \frac{\Omega_{simplex}}{\Omega_{n-1}}\equiv\tilde{\Omega}_{simplex} $$ where $\Omega_{simplex}$ refers to the solid angle of the vertex of the simplex (defined by $-\infty \leq z_2 \leq 0,-\infty \leq z_3\leq z_2\sqrt{\frac{3-2}{3}},\cdots,-\infty \leq z_n\leq z_{n-1}\sqrt{\frac{n-2}{n}}$) at the origin, and $\Omega_{n-1}$ is the solid angle of $\mathbb{R}^{n-1}$. This answer suggests that the integral for $\tilde{\Omega}_{simplex}$ may be written as a multivariate Taylor series $$ \begin{align*} \tilde{\Omega}_{simplex}&= \frac{n^{n/2}}{n!(4\pi)^{\frac{n-1}{2}}}\sum_{\boldsymbol{a}\in \mathbb{N}^{\binom{n-1}{2}}}\left[\prod_{i=1}^{n-1}\Gamma\left(\frac{1 + \sum_{m>i}a_{im} + \sum_{m<i}a_{mi}}{2}\right)\right]\prod_{i<j}\frac{s_{ij}^{a_{ij}}}{a_{ij}!}\\ &\equiv \frac{n^{n/2}}{n!(4\pi)^{\frac{n-1}{2}}}S(n) \end{align*} $$ where $$ s_{ij} = -2 \sqrt{\frac{i}{j}\frac{n-j}{n-i}}\hspace{5mm}\text{for }i<j $$ and note that the vector $\boldsymbol{a}$ consists of only the superdiagonal entries of a $(n-1)\times (n-1)$ matrix, hence it is defined as $\boldsymbol{a} = \{a_{ij}\}_{i,j=1, i<j}^{n-1}$ (note that $\binom{n-1}{2}$ is the number of such superdiagonal entries). This result comes from the fact that the normalized vectors $\overline{\boldsymbol{v}}_1,\cdots, \overline{\boldsymbol{v}}_n $ that define the simplex are given by $$ \overline{\boldsymbol{v}}_i = \sum_{k=i}^{n-1}\sqrt{\frac{i n}{k(k+1)(n-i)}}\textbf{e}_k $$ which implies that, if we let $\boldsymbol{V}\in \mathbb{R}^{(n-1)\times(n-1)}$ be the matrix with the $\overline{\boldsymbol{v}}_i$ as columns $$ \det(\boldsymbol{V}) = \frac{n^{n/2}}{n!} $$ The question then becomes how to evaluate this multivariable Taylor series, which admittedly doesn't seem feasible. To obtain the numerically observed result, we'd have to have $$ S(n) = \frac{(4\pi)^{\frac{n-1}{2}}}{n^{n/2}} $$ For $n=3$, we have $$ S(3) = \sum_{a_{12}=0}^\infty \Gamma\left(\frac{1 + a_{12}}{2}\right)^2\frac{(-1)^{a_{12}}}{a_{12}!} $$ and Mathematica gives $S(3)=\frac{4\pi}{3^{3/2}}$. For $n=4$, the sum is much more complicated: $$ S(4) = \sum_{a_{12}\geq 0}\sum_{a_{13}\geq 0}\sum_{a_{23}\geq 0}\Gamma\left(\frac{1 + a_{12} + a_{13}}{2}\right)\Gamma\left(\frac{1 + a_{23} + a_{12}}{2}\right)\Gamma\left(\frac{1 + a_{13} + a_{23}}{2}\right)\frac{\left(-\frac{2}{\sqrt{3}}\right)^{a_{12} + a_{23}}}{a_{12}!a_{23}!}\frac{\left(-\frac{2}{3}\right)^{a_{13}}}{a_{13}!} $$
In Terms of Joint Probability: I don't immediately see how this is useful, but the integral can be neatly written as $$ I(n)/\pi^{\frac{n-1}{2}} = P\left(Z_2>0, Z_3>Z_2 \sqrt{\frac{3-2}{3}}, \cdots, Z_n > Z_{n-1}\sqrt{\frac{n-2}{n}}\right) $$ where $Z_2,\cdots, Z_n =\mathcal{N}(0,1)$ are independent and identically distributed. And somehow we have, coincidently $$ I(n)/\pi^{\frac{n-1}{2}} = \frac{1}{n}P\left(Z_2>0, Z_3>Z_2, \cdots, Z_n > Z_{n-1}\right) $$ For $n=2$, we have $$ I(2)/\sqrt{\pi} = P(Z_2>0)=\frac{1}{2}\equiv \frac{1}{2!} $$ for $n=3$ $$ I(3)/\pi = P\left(Z_2>0, Z_3>Z_2\sqrt{\frac{3-2}{3}}\right) = \frac{1}{4} - T\left(0, \sqrt{\frac{3-2}{3}}\right) = \frac{1}{6}\equiv \frac{1}{3!} $$ where $T(h, a)$ is Owen's T function