Is it approximating the gradient or approximating the solution? If it’s the gradient, doesn’t it have the real gradient in expectation, so you can run as many

Is it approximating the gradient or approximating the solution? If it’s the gradient, doesn’t it have the real gradient in expectation, so you can run as many iterations to get epsilon precision up to what your machine will support?