A convex loss function is optimized with gradient descent. What key guarantee does convexity provide?