Supremizer Enrichment Strategy

• Supremizer enrichment strategy is a stabilization technique that augments reduced velocity spaces with pressure-derived auxiliary functions to restore the inf-sup condition.
• It constructs supremizers by solving auxiliary elliptic problems, thereby mitigating spurious pressure modes and enabling accurate pressure recovery in ROMs.
• While effective for incompressible flow and optimal control, the method increases online costs and may affect velocity accuracy or long-term robustness in complex simulations.

Supremizer enrichment strategy is a stabilization technique for reduced formulations of saddle-point partial differential equations in which the reduced velocity or primal space is augmented with auxiliary functions constructed from the pressure or dual space. Its central purpose is to restore or improve the reduced inf–sup, or Ladyzhenskaya–Babuška–Brezzi, condition, thereby preventing spurious pressure modes and enabling accurate recovery of pressure-dependent quantities. In reduced-order models for incompressible Stokes and Navier–Stokes equations, and in reduced models for certain optimal control systems, the strategy appears when independently generated reduced spaces fail to preserve the pressure–velocity coupling present in the full-order problem (Stabile et al., 2017, Kean et al., 2019, Davie et al., 2023).

1. Analytical motivation and inf–sup structure

The canonical setting is a mixed formulation with velocity or primal space $V$, pressure or dual space $Q$, and coupling bilinear form $b(v,q)$. Stability requires an inf–sup condition of the form

$$\beta_{\mathrm{LBB}}=\inf_{q\in Q}\sup_{v\in V}\frac{b(v,q)}{\|v\|_V\|q\|_Q}>0.$$

At full-order level this condition may be ensured by a stable finite element pair or by algorithmic pressure–velocity coupling. After reduction, however, the reduced spaces are usually built from snapshots independently, and the reduced inf–sup constant can become effectively zero. In finite-volume POD–Galerkin models for incompressible Navier–Stokes equations, this loss of stability manifests as spurious pressure modes and inaccurate pressure approximation; in POD models built from discretely divergence-free snapshots, the pressure term may drop out of the reduced momentum equation altogether, producing a velocity-only ROM and making post hoc pressure recovery necessary (Stabile et al., 2017, Kean et al., 2019).

This mechanism is not restricted to fluid flow. In parametrized optimal control problems with saddle-point structure, naive reduced spaces can annihilate the reduced coupling operator and thereby violate the analogue of the LBB condition. The reduced problem then becomes unstable unless the primal space is enriched or the state and adjoint spaces are modified by another stabilization device (Davie et al., 2023).

A recurring misconception is that full-order stability automatically transfers to the reduced model. The reported results show the opposite: even when the full-order discretization is stable, Galerkin projection onto reduced spaces generally does not preserve inf–sup stability, and the inability to approximate the pressure field accurately is a direct consequence (Ali et al., 2020).

2. Supremizer construction and mathematical definition

The supremizer is the Riesz representative of the pressure-coupling functional. In reduced basis notation, for a given pressure mode $q$, the supremizer operator is

$S=R_V^{-1}B^T,$

so that $s(q)\in V$ satisfies

$(s(q),v)_V=b(v,q)\qquad \forall v\in V.$

When the $V$-inner product is chosen as an energy or $H^1$-type inner product, this becomes an elliptic auxiliary problem aligned with the velocity operator (Mueller et al., 21 Aug 2025).

In the finite-element and finite-volume incompressible-flow literature represented here, the standard implementation is a stiffness-based or $Q$0-Riesz variant. For each pressure basis function $Q$1, or for each pressure snapshot, one solves

$Q$2

which in strong form becomes

$Q$3

The same construction is used for filtered pressures in Leray or LES-filtered reduced models. At the discrete level, the auxiliary solve is written as

$Q$4

with $Q$5 the velocity stiffness operator and $Q$6 the pressure-gradient coupling (Stabile et al., 2017, Girfoglio et al., 2021).

Two broad variants are distinguished. In the exact variant, one auxiliary problem is solved for each retained pressure basis function. In the approximate variant, one solves the auxiliary problem for each pressure snapshot, collects supremizer snapshots, and then compresses them by POD or another basis-generation procedure. The approximate construction reduces offline cost but does not provide the same formal stability guarantee as exact mode-by-mode enrichment (Stabile et al., 2017).

The same idea generalizes beyond incompressible flow. In the optimal-control setting, the supremizer is defined on the product state–control space $Q$7 with inner product induced by the block matrix

$Q$8

and the auxiliary solve is

$Q$9

Here the role of the pressure basis is played by reduced adjoint or Lagrange multiplier vectors, and the enriched space is the reduced state–control space rather than a pure velocity space (Davie et al., 2023).

3. Enriched reduced spaces and pressure recovery mechanisms

Once the supremizers are computed, they are appended to the reduced velocity or primal basis. In the classical POD–Galerkin Navier–Stokes case, the enriched velocity space takes the form

$b(v,q)$0

All reduced operators must then be reassembled on the enlarged basis. The mass, diffusion, divergence, gradient, and convective operators are no longer those of the original velocity POD space alone; they are defined on the union of velocity and supremizer modes (Stabile et al., 2017).

The enriched formulation can be used in two conceptually distinct ways. One option is a coupled mixed ROM, in which pressure remains an unknown in the reduced saddle-point system. Another option is sequential pressure recovery. In the momentum-equation recovery formulation, one first computes the reduced velocity from the velocity-only ROM, then recovers the reduced pressure from the momentum equation by testing against the supremizer space $b(v,q)$1. In the finite-element analysis of this method, the reduced pair $b(v,q)$2 satisfies a discrete inf–sup condition

$b(v,q)$3

and the sequential recovery is equivalent to a coupled enriched ROM, but with smaller online linear systems (Kean et al., 2019).

The Leray reduced model implemented through the Evolve–Filter algorithm introduces a more elaborate variant because two velocity fields and two pressure-like fields coexist: the evolve pair $b(v,q)$4 and the filtered pair $b(v,q)$5. In that setting, two enrichment strategies were compared. SUP1 enriches each velocity space only with supremizers associated with its own pressure field. SUP2 performs cross-enrichment: both the evolve and the filtered velocity spaces are enriched with supremizers associated with both pressure fields. The reported conclusion is that, in this tightly coupled EF setting, enriching only the “matching” velocity space is not sufficient; accurate pressure reconstruction requires enriching both velocity spaces with both sets of supremizers (Girfoglio et al., 2021).

This result suggests that supremizer design is not purely local to one reduced block when the reduced model contains two-way coupled mixed subproblems. The auxiliary spaces must reflect the actual coupling pattern of the ROM, not merely the field labels inherited from the full-order equations.

4. Offline and online realization

The standard workflow has a clear offline–online structure. Offline, one solves the full-order model over the training time interval or parameter set, stores velocity and pressure snapshots, constructs POD or reduced-basis spaces, solves the auxiliary elliptic supremizer problems, compresses or orthonormalizes the resulting supremizers, and then assembles all reduced operators on the enriched basis. Online, one solves the reduced mixed system or the reduced velocity equation plus pressure-recovery system (Stabile et al., 2017, Kean et al., 2019).

In finite-volume POD–Galerkin formulations, pressure and velocity snapshots are typically treated with $b(v,q)$6-based POD because the discrete fields are piecewise constant. Since the velocity POD basis and the supremizer basis are generated independently, their union is not globally orthogonal in $b(v,q)$7. Initial conditions and reduced solves are therefore handled through the reduced mass matrix rather than by assuming an orthonormal basis (Stabile et al., 2017).

The EF Leray ROM adds two implementation details. First, nonhomogeneous Dirichlet boundary conditions are handled by divergence-free lifting functions: the lifts are subtracted from velocity snapshots before POD and added back during reconstruction. Second, the filtered pressure requires its own supremizer family, built by solving

$b(v,q)$8

The reported offline choices retained approximately $b(v,q)$9 POD energy and used $$\beta_{\mathrm{LBB}}=\inf_{q\in Q}\sup_{v\in V}\frac{b(v,q)}{\|v\|_V\|q\|_Q}>0.$$0, $$\beta_{\mathrm{LBB}}=\inf_{q\in Q}\sup_{v\in V}\frac{b(v,q)}{\|v\|_V\|q\|_Q}>0.$$1, $$\beta_{\mathrm{LBB}}=\inf_{q\in Q}\sup_{v\in V}\frac{b(v,q)}{\|v\|_V\|q\|_Q}>0.$$2, $$\beta_{\mathrm{LBB}}=\inf_{q\in Q}\sup_{v\in V}\frac{b(v,q)}{\|v\|_V\|q\|_Q}>0.$$3; for SUP2 the authors used $$\beta_{\mathrm{LBB}}=\inf_{q\in Q}\sup_{v\in V}\frac{b(v,q)}{\|v\|_V\|q\|_Q}>0.$$4 and $$\beta_{\mathrm{LBB}}=\inf_{q\in Q}\sup_{v\in V}\frac{b(v,q)}{\|v\|_V\|q\|_Q}>0.$$5 (Girfoglio et al., 2021).

For unfitted finite elements on parameterized domains, the enrichment is transported to a reference configuration. There the velocity Riesz matrix includes an $$\beta_{\mathrm{LBB}}=\inf_{q\in Q}\sup_{v\in V}\frac{b(v,q)}{\|v\|_V\|q\|_Q}>0.$$6-like bulk term together with Nitsche penalty contributions. The practical implementation computes a Cholesky factor $$\beta_{\mathrm{LBB}}=\inf_{q\in Q}\sup_{v\in V}\frac{b(v,q)}{\|v\|_V\|q\|_Q}>0.$$7 of the reference velocity norm matrix $$\beta_{\mathrm{LBB}}=\inf_{q\in Q}\sup_{v\in V}\frac{b(v,q)}{\|v\|_V\|q\|_Q}>0.$$8 and forms the enrichment as $$\beta_{\mathrm{LBB}}=\inf_{q\in Q}\sup_{v\in V}\frac{b(v,q)}{\|v\|_V\|q\|_Q}>0.$$9, followed by $q$0-orthonormalization of the union of the velocity and supremizer bases. The reported theoretical conclusion is that the supremizer operator can be defined on the reference configuration without loss of inf–sup stability, provided the geometric mapping and aggregated discretization satisfy the stated norm-equivalence and coupling assumptions (Mueller et al., 21 Aug 2025).

5. Reported performance and comparison with alternative stabilizations

The empirical record is mixed but consistent in one respect: supremizer enrichment improves pressure stabilization, yet it is not uniformly the most efficient or the most robust option.

In finite-volume POD–Galerkin ROMs for the unsteady incompressible Navier–Stokes equations, the supremizer-enriched ROM improves the reduced inf–sup constant as the number of supremizer modes increases. For the lid-driven cavity test with $q$1, the reported reduced inf–sup constant reaches $q$2 at $q$3. For the circular-cylinder case with $q$4, $q$5, the value increases up to approximately $q$6 at $q$7. In short-time reconstructions the SUP-ROM tends to be more accurate for pressure than the pressure-Poisson alternative, but it may be slightly worse for velocity, and in longer-time cylinder simulations it exhibits energy growth and eventual non-physical patterns whereas the PPE-ROM remains regular, albeit with a small phase shift in vortex shedding (Stabile et al., 2017).

The EF Leray study sharpens this picture. In 2D unsteady flow past a cylinder at Reynolds number $q$8, the unstabilized ROM is unstable or inaccurate, SUP1 reduces velocity errors but leaves pressure errors unacceptable, and SUP2 significantly improves pressure accuracy. PPE yields the lowest velocity errors and is more computationally efficient, while SUP2 gives better filtered-pressure reconstruction. Representative average errors reported for SUP2 versus PPE are approximately $q$9 versus $S=R_V^{-1}B^T,$0 for $S=R_V^{-1}B^T,$1, $S=R_V^{-1}B^T,$2 versus $S=R_V^{-1}B^T,$3 for $S=R_V^{-1}B^T,$4, $S=R_V^{-1}B^T,$5 versus $S=R_V^{-1}B^T,$6 for $S=R_V^{-1}B^T,$7, and $S=R_V^{-1}B^T,$8 versus $S=R_V^{-1}B^T,$9 for $s(q)\in V$0. Lift and drag errors are comparable, with $s(q)\in V$1 for SUP2 versus $s(q)\in V$2 for PPE and $s(q)\in V$3 for SUP2 versus $s(q)\in V$4 for PPE. The online times were reported as approximately $s(q)\in V$5 s for SUP2 and $s(q)\in V$6 s for PPE, against a full-order cost of about $s(q)\in V$7 s (Girfoglio et al., 2021).

The finite-element analysis of momentum-equation pressure recovery provides a theoretical explanation for some of these observations. There, MER pressure recovery via supremizer test spaces is proved stable and convergent, with explicit dependence on POD truncation errors and on the principal-angle constant $s(q)\in V$8. Numerically, the reported pressure error obeys $s(q)\in V$9 when varying the pressure-space dimension and $(s(q),v)_V=b(v,q)\qquad \forall v\in V.$0 when varying the velocity-space dimension. In the same study, PPE errors stagnate around $(s(q),v)_V=b(v,q)\qquad \forall v\in V.$1–$(s(q),v)_V=b(v,q)\qquad \forall v\in V.$2, and time-averaged pressure error fields show PPE inaccuracies concentrated near the boundary, whereas MER errors are more uniformly distributed and notably smaller, approximately $(s(q),v)_V=b(v,q)\qquad \forall v\in V.$3 versus approximately $(s(q),v)_V=b(v,q)\qquad \forall v\in V.$4 (Kean et al., 2019).

Supremizer enrichment also competes with other stabilization philosophies. In stabilized reduced-basis methods for steady Stokes and Navier–Stokes equations, consistent offline-online residual-based stabilization can preserve a modified inf–sup stability without enlarging the velocity basis. For equal-order pairs such as $(s(q),v)_V=b(v,q)\qquad \forall v\in V.$5 and $(s(q),v)_V=b(v,q)\qquad \forall v\in V.$6, stabilization-only often performs comparably to or better than stabilization plus supremizers in velocity, while the combined strategy can improve pressure by about one order of magnitude. For stable Taylor–Hood $(s(q),v)_V=b(v,q)\qquad \forall v\in V.$7 reduced models without online residual stabilization, however, supremizers remain necessary to preserve reduced pressure accuracy (Ali et al., 2020).

In parametrized optimal control, the assessment is less favorable to supremizers. The approximate supremizer method required substantially more basis vectors than aggregation and in several Petrov–Galerkin runs failed to reach the prescribed tolerance. For diffusion control with Galerkin projection and $(s(q),v)_V=b(v,q)\qquad \forall v\in V.$8, the reported basis counts were $(s(q),v)_V=b(v,q)\qquad \forall v\in V.$9 for supremizers versus $V$0 for aggregation at $V$1, corresponding to $V$2 versus $V$3 total columns; for $V$4 and $V$5, the counts were $V$6 versus $V$7, corresponding to $V$8 versus $V$9 columns. The authors concluded that aggregation was more robust and efficient for those control problems (Davie et al., 2023).

6. Limitations, extensions, and scope of applicability

Supremizer enrichment is a targeted remedy for reduced inf–sup failure, not a universal cure for all reduced-model pathologies. Several limitations recur across the reported studies.

First, approximate enrichment may stabilize pressure empirically without furnishing a formal reduced inf–sup guarantee. This is explicit in snapshot-based finite-volume enrichment and in the optimal-control setting, where offline supremizers are parameter-dependent approximations rather than exact online Riesz representatives (Stabile et al., 2017, Davie et al., 2023).

Second, enrichment increases reduced dimension and therefore online cost. In ROMs with polynomial or tensorial nonlinearities, the storage of reduced convective tensors grows cubically with the number of enriched velocity modes. This is manageable in the reported low-dimensional tests but becomes a design constraint for larger bases (Stabile et al., 2017).

Third, improved pressure behavior can coexist with worse velocity accuracy or weaker long-time robustness. The FV cylinder tests indicate that SUP-ROM may be more accurate for short-time pressure reconstruction while PPE-ROM is faster and more robust for long periodic integrations (Stabile et al., 2017). The EF study likewise recommends PPE when computational efficiency and velocity accuracy are paramount and filtered-pressure accuracy is less critical, and SUP2 when pressure fields, especially the filtered pressure $H^1$0, are central (Girfoglio et al., 2021).

Fourth, the correct enrichment depends on problem structure. In EF ROMs, enriching only with “own-pressure” supremizers is inadequate; in optimal control, aggregation can dominate supremizers because the state and adjoint spaces coincide naturally; in unfitted methods, the velocity norm used in the Riesz map must incorporate Nitsche penalty terms to maintain robustness near cut cells (Girfoglio et al., 2021, Davie et al., 2023, Mueller et al., 21 Aug 2025).

Despite these restrictions, the strategy has broadened rather than narrowed in recent work. It has been adapted to residual-based stabilized reduced-basis formulations, to parameterized domains via reference-configuration mappings, to unfitted aggregated finite elements, and to tensor-based reduced bases with hyper-reduction. This suggests that the enduring content of the method is not a particular Poisson solve, but the reduced-space principle it encodes: construct primal enrichment functions as Riesz representatives of the coupling operator so that the reduced saddle-point structure remains compatible with the full-order one (Ali et al., 2020, Mueller et al., 21 Aug 2025).