With three variables the gradient has three entries and the Hessian is a 3 \times 3 symmetric matrix. The question “is this a bowl?” becomes “is the Hessian positive (semi)definite?”: does the function curve upward, or at least not downward, in every direction? Two practical tests answer it: the eigenvalues, or the leading principal minors.
\nabla f = \begin{pmatrix} f_x \\ f_y \\ f_z \end{pmatrix}, \qquad H = \begin{pmatrix} f_{xx} & f_{xy} & f_{xz} \\ f_{xy} & f_{yy} & f_{yz} \\ f_{xz} & f_{yz} & f_{zz} \end{pmatrix} In words: first partials in a column; second partials in a symmetric table.
\begin{align*} d^{\mathsf T} H\, d > 0 \text{ for every } d \ne 0 &\iff \text{all eigenvalues of } H > 0\\ &\iff \text{all leading principal minors} > 0 \end{align*} In words: positive definite means upward curvature in every direction d; check it with eigenvalues, or with the determinants of the top-left 1\times1, 2\times2 and 3\times3 blocks (Sylvester’s test).
f \text{ convex} \iff H \text{ positive semidefinite (all eigenvalues} \ge 0\text{) at every point} In words: no direction ever curves downward. For semidefiniteness the leading minors are not enough; use the eigenvalues (or all principal minors).
Take f = x^2 + y^2 + z^2 + xy + yz - 4x - 6y - 4z. Then \nabla f = \begin{pmatrix}2 x + y - 4\\x + 2 y + z - 6\\y + 2 z - 4\end{pmatrix}, \qquad H = \begin{pmatrix}2 & 1 & 0\\1 & 2 & 1\\0 & 1 & 2\end{pmatrix} . In words: the Hessian is the same at every point, because f is quadratic. Its eigenvalues are 2 - \sqrt{2}, 2, \sqrt{2} + 2 (the smallest is about 0.5858), and its leading principal minors are 2, 3, 4: all positive, so H is positive definite and f is convex with exactly one minimum.
Setting \nabla f = 0 gives (x, y, z) = (1, 2, 1) with value -10. The kernel agrees:
minimize x^2 + y^2 + z^2 + x*y + y*z - 4*x - 6*y - 4*z
(1, 2, 1), value -10, status optimal.
A warning: positive numbers on the diagonal are not enough. For x^2 + y^2 + z^2 + 4xy the Hessian is \begin{pmatrix}2 & 4 & 0\\4 & 2 & 0\\0 & 0 & 2\end{pmatrix}, with eigenvalues -2, 2, 6 and leading minors 2, -12, -24: one negative eigenvalue, so the function is not convex (it falls along the direction (1, -1, 0)).
For a quadratic objective \tfrac12 x^{\mathsf T} Q x + c^{\mathsf T} x the Hessian is Q, so “is Q positive semidefinite?” decides whether ProvenDecisions can treat the problem as a convex QP with a proven global optimum, or must label it nonconvex. The eigenvalue test is the one a computer runs; the minor test is the one you can do by hand.