Mixed formulations and inf-sup stability

Let V and Q be Hilbert spaces with declared norms, let a be bounded on V times V, and let b be bounded on V times Q. Take bounded load functionals F in V′ and G in Q′. The equations are a(u,v)+b(v,p)=F(v) and b(u,q)=G(q). Define the constraint kernel Z as primal functions annihilated by b against every multiplier. A sufficient Brezzi framework requires a to be coercive on Z and b to satisfy a positive inf-sup condition. Full-space coercivity of a is stronger than necessary in this framework.

The inf-sup expression asks, for each nonzero multiplier, whether a primal direction detects it with uniformly controlled norm. It rules out an undetectable multiplier kernel and supplies a stable way to enforce the constraint. These are analytic conditions tied to norms, not a count of unknowns and equations. A bounded system can have the correct dimensions but nearly undetectable multiplier directions, yielding a very large solution response to small data perturbations.

For conforming pairs V_h and Q_h, the discrete kernel consists of functions whose constraints vanish only against Q_h. It can be larger than the intersection of V_h with the continuous kernel. Consequently continuous kernel coercivity does not automatically imply discrete kernel coercivity. A stable discrete family needs suitable lower bounds for both the discrete kernel-coercivity constant and beta_h, independent of the mesh parameter in the physical function-space norms.

An algebraic singular-value calculation represents this condition only after the coefficient norms have been matched to those function-space norms, typically through Gram matrices. Raw coordinate scaling can change Euclidean singular values without changing the underlying functions. A Fortin operator is one common stability tool: it maps continuous primal tests into the discrete space, preserves the b pairing with every discrete multiplier, and has a norm bound independent of h. That bound transfers continuous inf-sup control to the discrete pair.

In a finite-dimensional Euclidean model, take the coupling matrix diagonal with entries one and epsilon. For every positive epsilon the two-by-two matrix is invertible, yet its smallest singular value and inf-sup constant equal epsilon when epsilon is at most one. A load in the weak multiplier direction can therefore require a multiplier of order one divided by epsilon. The equality of primal and multiplier dimensions has not supplied a useful uniform lower bound.

This simple example explains why a numerical solve on one mesh cannot certify a mixed element family. One needs a theorem, a norm-aware stability analysis, or a justified stabilization design that remains controlled under refinement. Stabilization changes the discrete equations and its consistency must also be checked. Adding degrees of freedom can improve stability, but an enrichment name or dimension inequality is not itself an inf-sup proof. The mixed framework separates these requirements so that each can be verified.