If I want to fit a quadratic function of two variables to some data, I can use
$$f(x, y) = c_1 x^2 + c_2 xy + c_3 y^2 + c_4 x + c_5 y + c_6$$
$$\frac{\partial}{\partial c_i} \sum_j\left( z_j - f(x_j, y_j) \right)^2 = 0$$
to obtain six equations, and then endeavor to solve them.
I've done it for one variable not two, but I'm guessing the process is straightforward.
If I extend this to more variables, and to higher order than 2, when will the analytical expressions be at risk for having multiple solutions?