0
$\begingroup$

For an algorithm, I have to simulate $X_1, \ldots, X_n \sim_{\text{iid}} \text{Logistic}(0,1)$ conditionally to the event $(X_1, \ldots, X_n) \in P$, where $P$ is an affine plane in $\mathbb{R}^n$.

I really need to sample from this distribution; the goal is not to evaluate an integral.

What are the possible methods?

$\endgroup$
1
  • $\begingroup$ @MattF.The affine plane is $P=\{u\alpha+v\beta+b\mid u,v\in \mathbb{R}^2\}$ where $\alpha$ and $\beta$ are two orthogonal vectors in $\mathbb{R}^n$ and $b \in \mathbb{R}^n$. So the event has probability $0$. I need to simulate the conditional distribution. $\endgroup$ – Stéphane Laurent Dec 12 '20 at 17:50
1
$\begingroup$

According to your comment, your plane is $P=\{u\alpha+v\beta+b\colon(u,v)\in\mathbb R^2\}$, where $\alpha$ and $\beta$ are two orthogonal vectors in $\mathbb R^n$ and $b\in\mathbb R^n$. Assuming that the vectors $\alpha$ and $\beta$ are nonzero, let $(a_1,\dots,a_n)$ be any basis of $\mathbb R^n$ such that $a_1=\alpha$ and $a_2=\beta$. Let the random variables $Y_1,\dots,Y_n$ be the coordinates of the random vector $X-b$, where $X:=[X_1,\dots,X_n]^T$, so that $X-b=AY$, where $A$ is the $n\times n$ matrix with columns $a_1,\dots,a_n$ and $Y:=[Y_1,\dots,Y_n]^T$.

Then the joint pdf of $X_1,\dots,X_n$ is given by the formula $$f_X(x)=f(x_1)\cdots f(x_n)\tag{1}$$ for $x=(x_1,\dots,x_n)\in\mathbb R^n$, where $f$ is the pdf of $\text{Logistic}(0,1)$, and the joint pdf of $Y_1,\dots,Y_n$ is given by the formula $$f_Y(y)=f_X(b+Ay)|\det A| \tag{2}$$ for $y=(y_1,\dots,y_n)\in\mathbb R^n$.

So, the joint pdf of the 2D conditional distribution you want to sample from is given by the formula $$f_{Y_1,Y_2|Y_3,\dots,Y_n}(y_1,y_2|y_3=\dots=y_n=0) =\frac{f_Y(y_1,y_2,0,\dots,0)}{f_{Y_3,\dots,Y_n}(0,\dots,0)} \propto f_Y(y_1,y_2,0,\dots,0)$$ for $(y_1,y_2)\in\mathbb R^2$, where $f_{Y_3,\dots,Y_n}$ is the joint pdf of $Y_3,\dots,Y_n$ and $\propto$ means "proportional to".

So, if you don't want to evaluate the integral $$f_{Y_3,\dots,Y_n}(0,\dots,0)=\int_{\mathbb R^2}dy_1\,dy_2\, f_Y(y_1,y_2,0,\dots,0), $$ then to sample from the 2D conditional distribution in question you can use one of the Markov chain Monte Carlo (MCMC) methods, which "create samples from a continuous random variable, with probability density proportional to a known function." In your case, the known function is $(y_1,y_2)\mapsto f_Y(y_1,y_2,0,\dots,0)$, with $f_Y$ computed according to (2) and (1).


See e.g. this answer on how to realize MCMC methods in Mathematica.

$\endgroup$

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.