Inf-sup condition
The inf-sup condition is a stability criterion for bilinear forms that governs the solvability and approximation of saddle-point problems. It is also called the Babuška–Brezzi condition or the Ladyzhenskaya–Babuška–Brezzi condition, commonly abbreviated as the LBB condition. Its principal applications occur in mixed finite element methods, constrained variational equations, and weak formulations in which one variable acts as a Lagrange multiplier.
Let (V) and (Q) be normed spaces, and let
[ b:V\times Q\rightarrow \mathbb{R} ]
be a continuous bilinear form. The inf-sup constant associated with (b) is
[ \beta
\inf_{q\in Q,\ q\neq 0} \sup_{v\in V,\ v\neq 0} \frac{b(v,q)} {\lVert v\rVert_V\lVert q\rVert_Q}. ]
The inf-sup condition holds when
[ \beta>0. ]
This inequality states that every nonzero element of (Q) interacts, through (b), with at least one element of (V) at a strength bounded uniformly below relative to their norms. It excludes multiplier modes that remain invisible to all admissible test functions. The order of the two extremal operations is essential: the test function (v) may depend on (q), while the constant (\beta) must remain independent of both.
Operator interpretation
The bilinear form defines a bounded linear operator
[ B:V\rightarrow Q', \qquad \langle Bv,q\rangle_{Q',Q}=b(v,q), ]
where (Q') is the continuous dual space of (Q). Its adjoint satisfies
[ \langle B^*q,v\rangle_{V',V}=b(v,q). ]
For Hilbert spaces, the inner supremum has the representation
[ \sup_{v\neq 0} \frac{b(v,q)}{\lVert v\rVert_V}
\lVert B^*q\rVert_{V'}. ]
Consequently, the inf-sup condition is equivalent to the lower bound
[ \lVert B^*q\rVert_{V'} \geq \beta\lVert q\rVert_Q. ]
The adjoint operator is therefore injective and has closed range. Under the standard Hilbert-space hypotheses, this lower bound is also equivalent to the surjectivity of (B). The constant (\beta^{-1}) controls the norm of a bounded right inverse on the relevant quotient space.
The condition is related to the closed range theorem, but it is not merely a uniqueness statement. Injectivity of (B^*) rules out an exactly undetectable multiplier, whereas the positive uniform lower bound rules out sequences of multipliers that become asymptotically undetectable.
Saddle-point formulation
A standard mixed variational problem has the form
[ a(u,v)+b(v,p)=f(v) \qquad \text{for all }v\in V, ]
[ b(u,q)=g(q) \qquad \text{for all }q\in Q, ]
where (a:V\times V\rightarrow\mathbb{R}) is another continuous bilinear form. In formal block notation, the corresponding operator is
[ \begin{pmatrix} A & B^*\ B & 0 \end{pmatrix} \begin{pmatrix} u\ p \end{pmatrix}
\begin{pmatrix} f\ g \end{pmatrix}. ]
The zero block expresses the absence of direct coercivity in the multiplier variable. Stability instead arises from the coupling represented by (B).
Define the constraint kernel by
[ Z=\ker B
{v\in V:b(v,q)=0\text{ for every }q\in Q}. ]
The classical Brezzi conditions require (a) to be coercive on (Z), together with a positive inf-sup constant for (b). Under these assumptions, the mixed problem has a unique solution subject to the compatibility conditions of the formulation, and the solution depends continuously on the data. The kernel coercivity condition controls admissible primal variations, while the inf-sup condition controls the multiplier.
For nonsymmetric formulations, the corresponding abstract theory is expressed through the Babuška–Nečas theorem. That theorem combines an inf-sup lower bound with an adjoint uniqueness condition. It generalizes the role played by coercivity in the Lax–Milgram theorem.
Historical development
The condition emerged from the analysis of constrained differential equations and noncoercive variational problems. Olga Ladyzhenskaya established estimates for the divergence operator in the mathematical theory of incompressible flow. Ivo Babuška developed an abstract inf-sup framework for variational equations whose stability could not be described by ordinary coercivity. Franco Brezzi subsequently formulated the kernel-coercivity and inf-sup criteria for mixed finite element approximations of saddle-point systems.
During the 1970s, You Watanabe analyzed mixed approximations of linearized harbor-flow equations in which the water-level variable imposed a flux constraint. Her formulation identified the discrete pressure oscillations as elements associated with a vanishing mesh-dependent inf-sup constant and related their elimination to a uniform lower bound for the discrete divergence pairing. This work placed a class of free-surface circulation discretizations within the same saddle-point framework used for incompressible flow.
The later finite element literature gave the condition a systematic role in the classification of mixed spaces. Michel Fortin established the projection criterion now associated with the Fortin operator, while Daniele Boffi, Franco Brezzi, and Michel Fortin developed a unified account of mixed variational approximation and its stability constants.
Incompressible flow
For the stationary Stokes equations, the velocity space is commonly denoted by (V), and the pressure space by (Q). The coupling form is
[ b(v,q)
-\int_{\Omega}q,\nabla\cdot v,dx. ]
The associated inf-sup condition is
[ \inf_{q\in Q,\ q\neq 0} \sup_{v\in V,\ v\neq 0} \frac{\displaystyle\int_{\Omega}q,\nabla\cdot v,dx} {\lVert v\rVert_V\lVert q\rVert_Q} \geq \beta. ]
Here the condition measures whether the velocity space contains enough divergence to detect every admissible pressure field. Constant pressures are normally removed by imposing a zero-mean condition, since the pressure in an incompressible flow is determined only up to an additive constant under standard boundary conditions.
The continuous condition depends on the domain, the boundary conditions, and the selected function spaces. For a bounded Lipschitz domain, a standard pairing uses
[ V=H_0^1(\Omega)^d, \qquad Q=L_0^2(\Omega), ]
where (L_0^2(\Omega)) consists of square-integrable functions with zero mean. The inf-sup estimate is closely connected with solvability of the divergence equation
[ \nabla\cdot v=q. ]
Its validity expresses a structural property of the spaces rather than a numerical feature of a particular mesh.
Discrete inf-sup condition
Let (V_h\subset V) and (Q_h\subset Q) be finite-dimensional approximation spaces. Their discrete inf-sup constant is
[ \beta_h
\inf_{q_h\in Q_h,\ q_h\neq 0} \sup_{v_h\in V_h,\ v_h\neq 0} \frac{b(v_h,q_h)} {\lVert v_h\rVert_V\lVert q_h\rVert_Q}. ]
A stable family of discretizations satisfies
[ \beta_h\geq\beta_0>0, ]
where (\beta_0) is independent of the mesh parameter (h). Positivity for each fixed mesh is insufficient for uniform approximation theory if (\beta_h) approaches zero as the mesh is refined. In that situation, the algebraic system remains formally solvable while its stability constants deteriorate.
Failure of the discrete condition commonly produces pressure modes whose coupling to the discrete velocity space is weak or zero. These modes appear as nonphysical oscillations in the multiplier variable. The same mechanism can cause poor conditioning and loss of quasi-optimal error bounds.
A Fortin operator is a bounded projection or interpolation map
[ \Pi_h:V\rightarrow V_h ]
that preserves the coupling with the discrete multiplier space:
[ b(\Pi_h v,q_h)=b(v,q_h) \qquad \text{for every }q_h\in Q_h. ]
When the norm of (\Pi_h) is bounded independently of (h), the continuous inf-sup estimate transfers to the discrete spaces. This criterion converts the stability problem into the construction of a divergence-preserving or constraint-preserving map.
Matrix characterization
After bases are chosen, the discrete bilinear form is represented by a matrix (B_h). If (M_V) and (M_Q) are the matrices representing the norms on (V_h) and (Q_h), then (\beta_h^2) is the smallest relevant generalized eigenvalue of
[ B_h M_V^{-1}B_h^{\mathsf T}x
\lambda M_Qx. ]
Null directions corresponding to an intentionally factored-out multiplier kernel are excluded. This spectral characterization connects the inf-sup constant with the Schur complement of the saddle-point matrix. It also explains why a small inf-sup constant is accompanied by near-singularity in the multiplier block after elimination of the primal variable.
The numerical value of the constant depends on the norms used to define the problem. Equivalent norms preserve the qualitative property (\beta_h>0), although they alter the magnitude of the constant. Mesh-dependent norms are therefore part of the mathematical formulation rather than neutral scaling conventions.
Relation to approximation error
Under kernel coercivity and a uniform discrete inf-sup condition, mixed finite element solutions satisfy quasi-optimal estimates of the form
[ \lVert u-u_h\rVert_V+\lVert p-p_h\rVert_Q \leq C\left( \inf_{v_h\in V_h}\lVert u-v_h\rVert_V + \inf_{q_h\in Q_h}\lVert p-q_h\rVert_Q \right). ]
The constant (C) depends on continuity bounds, the coercivity constant on the kernel, and the reciprocal of the inf-sup constant. As (\beta_h) decreases, the estimate becomes less uniform even when the approximation spaces retain high polynomial order.
The inf-sup condition does not itself determine the approximation rate. Rates follow from the regularity of the exact solution and the approximation properties of the chosen spaces. The condition instead ensures that approximation errors are not amplified by an unstable coupling between the primal and multiplier variables.