Your difficulties extend beyond erf. CVX does have log_normcdf(x), which approximates the cumulative standard Normal distribution function, into which erf is convertible`. However, the variable product stuff going on looks like a total no-go. I don’t think CVX is the tool for this problem.