Reduced-Space Formulation in Computational Modeling

• Reduced-space formulation is a method that reexpresses complex models on a smaller set of variables while maintaining key physical and mathematical structures.
• It is used in applications such as PDE reduction, constraint elimination, and surrogate modeling to achieve computational gains and accurate structure preservation.
• This approach enables high-fidelity systems to be approximated with lower-dimensional models that remain faithful to the original governing operators and constraints.

Searching arXiv for recent and relevant papers on “reduced-space formulation” and closely related formulations across application domains. Reduced-space formulation denotes a class of reformulations in which a governing problem is expressed on a smaller set of variables, basis functions, invariant coordinates, or quotient degrees of freedom than those of the original model. In the cited literature, the term covers projected finite-dimensional subspaces for PDEs, elimination of state variables through implicit constraints, slow invariant manifolds for open quantum systems, reduced input parameter spaces for surrogate modeling, reduced phase spaces obtained by symmetry quotient, and one-body reduced descriptions in quantum many-body theory. Taken together, these usages suggest that the expression does not designate a single algebraic recipe, but a family of constructions whose common aim is to retain the operative physics, constraints, or variational structure while replacing the full description by a lower-dimensional or less redundant one.

1. Principal meanings of reduced-space formulation

Across the literature, the reduced object can be a projected modal space, a control space after eliminating states, a slow manifold, a quotient by symmetry directions, or a reduced observable space. The operative distinction is not merely smaller dimension, but the choice of a reduced description that remains faithful to the governing structure.

Setting	Reduced object	Representative formulation
Thermomechanical vibration	Structural and thermal modal subspaces in symmetric state space	MMS with residual flexibility (Ahn et al., 2021)
Composite open quantum systems	Invariant slow manifold $D_\epsilon$ with slow coordinates $x_d$	Adiabatic elimination of fast Lindblad modes (Régent et al., 2023)
AC optimal power flow	Control space $\mathbf u$ with states $\mathbf x(\mathbf u)$ eliminated	Feasible-path reduced OPF (Pacaud et al., 2021)
Space-time PDE discretization	Low-dimensional space-time or spatial subspace within all-at-once formulations	Matrix-oriented and reduced-basis space-time methods (Henning et al., 2019)
Uncertainty quantification	Reduced input parameter space $\boldsymbol{\theta}_r$	DGSM-driven surrogate construction (Vohra et al., 2018)
Geometric mechanics	Reduced presymplectic or reduced phase space after quotienting symmetries	Maxwell–Vlasov and BK-dust gravity reductions (Colombo, 25 Nov 2025)

A recurring feature is that reduction is defined relative to a governing operator or constraint set. In some works the full model is projected onto chosen subspaces; in others the reduced description is obtained by solving constraints exactly, by quotienting gauge or relabeling directions, or by replacing many-body states with reduced density matrices or densities. This suggests that “reduced-space formulation” is best understood structurally rather than dimensionally.

2. Projection-based reduced spaces in PDEs and multiphysics

A prominent usage arises in projection-based model reduction for PDEs and multiphysics systems. In thermomechanical vibration, the full coupled thermoelastic finite-element model is first recast in a symmetric state-space form with state vector $$\mathbf d = [\mathbf u_s,\dot{\mathbf u}_s,\boldsymbol{\Theta}_T]^T$$, and the reduced-space formulation is then built by projecting structural and thermal variables onto reduced modal subspaces while preserving multiphysics coupling through residual flexibility. The final reduced system has the form $$\mathbf A_p \dot{\mathbf d}_p + \mathbf B_p \mathbf d_p = \mathbf f_p$$, with $\mathbf d_p \in \mathbb R^{2N_d+N_a}$, and the thermal subspace is updated by structural residual flexibility rather than chosen independently (Ahn et al., 2021).

For space-fractional diffusion, the reduced finite-element formulation is constructed by POD on snapshots of the full FE solution. The reduced space $W^d = \mathrm{span}\{\psi_1,\dots,\psi_d\}$ is defined in the fractional energy inner product, and the full dense system generated by the nonlocal operator is replaced by a $d\times d$ system for POD coefficients. The reduced solution takes the form $x_d$0, and the paper proves unconditional stability together with an error decomposition separating FE discretization error from POD truncation error (Sun et al., 2018).

For the 2D incompressible Navier–Stokes equations in stream function–vorticity form, the reduced-space formulation is likewise Galerkin but explicitly uses distinct reduced spaces and distinct reduced coefficients for $x_d$1 and $x_d$2. The reduced expansions $x_d$3 and $x_d$4 are built from a global POD basis over time and parameter samples, and the projected dynamics yield a small reduced system involving mass, diffusion, Poisson, and nonlinear convection tensors (Girfoglio et al., 2022).

A related space-time interpretation appears in least-squares and Petrov–Galerkin formulations. In the parabolic reduced-basis construction, a high-fidelity space-time space $x_d$5 is replaced by a low-dimensional space $x_d$6, and the reduced solution is represented as $x_d$7. The reduced formulation preserves symmetry, uniform coercivity, and continuity of the underlying bilinear form, which permits offline–online decomposition and absolute and relative error bounds in the discrete space-time norm (Hinze et al., 29 Jan 2026).

The same theme appears in matrix-oriented space-time Petrov–Galerkin discretizations, where the all-at-once space-time solution matrix $x_d$8 is approximated by $x_d$9, with $\mathbf u$0 spanning a rational Krylov subspace. The resulting reduced matrix equation $\mathbf u$1 replaces the full $\mathbf u$2-dimensional problem by one of dimension $\mathbf u$3 (Henning et al., 2019).

A dual-space variant is developed for parametrized advection–reaction problems in an ultraweak optimal-test formulation. There the reduced space is built in the dual solution space $\mathbf u$4, solving reduced normal equations

$\mathbf u$5

and the reduced primal space arises automatically as $\mathbf u$6. The necessary computations are performed entirely in the space of dual solutions, and the paper proves exponential convergence of the Kolmogorov $\mathbf u$7-width together with greedy quasi-optimality (Engwer et al., 2023).

3. Constraint elimination, condensation, and mixed-variable reduction

A second major meaning of reduced-space formulation is elimination of state variables governed by equality constraints. In AC optimal power flow, the state–control partition $\mathbf u$8 is chosen so that the power-flow equations $\mathbf u$9 determine the state locally as $\mathbf x(\mathbf u)$0. The resulting reduced-space OPF minimizes $\mathbf x(\mathbf u)$1 subject only to inequality constraints in $\mathbf x(\mathbf u)$2, while the physical power-balance equations are satisfied by construction at every iteration. This feasible-path property is central to the proposed real-time algorithm, which operates directly in the reduced space and computes reduced gradients and reduced Hessians through adjoint and adjoint–adjoint methods rather than by forming $\mathbf x(\mathbf u)$3 explicitly (Pacaud et al., 2021).

The same elimination principle is pursued at the level of interior-point linear algebra. In condensed reduced-space IPMs, variables are partitioned into control $\mathbf x(\mathbf u)$4 and state $\mathbf x(\mathbf u)$5, with state equations $\mathbf x(\mathbf u)$6. Depending on whether reduction is performed after or before linearization, the paper derives linearize-then-reduce and reduce-then-linearize variants. In both cases, the large sparse KKT system is condensed to a dense system $\mathbf x(\mathbf u)$7, whose size is proportional to the number of degrees of freedom $\mathbf x(\mathbf u)$8, making dense GPU factorization attractive when $\mathbf x(\mathbf u)$9 is relatively small (Pacaud et al., 2022).

For SCOPF and related power-grid problems, reduced-space IPM is combined with quasi-Newton updates and adjoint gradients. The optimizer works in the control space while the power-flow states are recovered scenario by scenario, and the gradient is assembled from adjoint solves that are naturally parallel across contingencies. This transfers the main cost from a monolithic full-space KKT factorization to many smaller forward and adjoint solves, a structure explicitly motivated by high-performance computing (Kardos et al., 2020).

An electromagnetic analogue appears in the reduced $\boldsymbol{\theta}_r$0-$\boldsymbol{\theta}_r$1 formulation for superconductors. There the vector field $\boldsymbol{\theta}_r$2 is retained only in conducting regions, while nonconducting regions are described by a scalar magnetic potential $\boldsymbol{\theta}_r$3 satisfying $\boldsymbol{\theta}_r$4. The method further separates the magnetic field into a source field $\boldsymbol{\theta}_r$5 and a reaction field $\boldsymbol{\theta}_r$6, so that the time-dependent nonlinear solve is restricted to the superconducting body and a reduced surrounding domain. This is a reduced-space formulation in the sense of mixed variables, source–reaction decomposition, and domain reduction (Arsenault et al., 2021).

4. Slow manifolds, reduced phase spaces, and symmetry quotient

Reduced-space formulation is also used for reductions that are not subspace projections in the linear-algebraic sense. In composite open quantum systems governed by Lindblad equations, the reduced model is built by adiabatically eliminating fast dissipative modes and representing the dynamics on an invariant slow manifold $\boldsymbol{\theta}_r$7 of the same dimension as the stationary subspace $\boldsymbol{\theta}_r$8. The reduced coordinates are $\boldsymbol{\theta}_r$9, defined by invariant operators $$\mathbf d = [\mathbf u_s,\dot{\mathbf u}_s,\boldsymbol{\Theta}_T]^T$$0, and the effective dynamics is a low-dimensional linear ODE

$$\mathbf d = [\mathbf u_s,\dot{\mathbf u}_s,\boldsymbol{\Theta}_T]^T$$1

The construction avoids the full tensor-product Hilbert space and yields accessible models for one-, two-, and three-cat-qubit gates (Régent et al., 2023).

In the Maxwell–Vlasov system, reduced-space formulation means starting from a unified presymplectic Skinner–Rusk system on $$\mathbf d = [\mathbf u_s,\dot{\mathbf u}_s,\boldsymbol{\Theta}_T]^T$$2, generating the constraint hierarchy by the Gotay–Nester–Hinds algorithm, and then quotienting symmetry directions such as phase-space relabeling. The reduction by the diffeomorphism group produces a reduced presymplectic manifold $$\mathbf d = [\mathbf u_s,\dot{\mathbf u}_s,\boldsymbol{\Theta}_T]^T$$3, whose dynamics reproduces both the Euler–Poincaré formulation and the Marsden–Weinstein/Morrison–Greene Lie–Poisson structure (Colombo, 25 Nov 2025).

A related but distinct reduced-phase-space construction appears in asymptotically flat gravity coupled to Brown–Kuchař dust. There the dust fields provide physical clocks and rods, the Hamiltonian and diffeomorphism constraints are solved, and the reduced phase space is coordinatized by Dirac observables $$\mathbf d = [\mathbf u_s,\dot{\mathbf u}_s,\boldsymbol{\Theta}_T]^T$$4 defined on dust space. The physical Hamiltonian density is

$$\mathbf d = [\mathbf u_s,\dot{\mathbf u}_s,\boldsymbol{\Theta}_T]^T$$5

and it generates time evolution with respect to the dust clock. After imposing asymptotically flat boundary conditions directly on the reduced variables, the authors construct conserved symmetry charges whose reduced-phase-space Poisson algebra acquires a central extension and whose suitable quotient is closely related to the BMS algebra at spatial infinity (Han et al., 2023).

These examples show that reduced-space formulation can refer to invariant manifolds or symmetry quotients rather than to truncation of basis functions. This suggests a broader classification: some reduced spaces are chosen by approximation, whereas others arise from exact elimination of gauge, fast, or redundant degrees of freedom.

5. Reduced inputs and reduced observables

The notion also appears when the reduction is applied to the input or observable side rather than to the state equations. In surrogate modeling under uncertainty, a reduced-space formulation means replacing the full uncertain parameter vector $$\mathbf d = [\mathbf u_s,\dot{\mathbf u}_s,\boldsymbol{\Theta}_T]^T$$6 by an active subset $$\mathbf d = [\mathbf u_s,\dot{\mathbf u}_s,\boldsymbol{\Theta}_T]^T$$7, identified by derivative-based global sensitivity measures. The resulting reduced-space surrogate $$\mathbf d = [\mathbf u_s,\dot{\mathbf u}_s,\boldsymbol{\Theta}_T]^T$$8 is built only in the lower-dimensional space, with unimportant parameters fixed at nominal values, and the screening metric is the normalized quantity $$\mathbf d = [\mathbf u_s,\dot{\mathbf u}_s,\boldsymbol{\Theta}_T]^T$$9 derived from DGSMs and Poincaré-type constants (Vohra et al., 2018).

In quantum many-body theory, the reduction can be even more radical: one passes from the full $$\mathbf A_p \dot{\mathbf d}_p + \mathbf B_p \mathbf d_p = \mathbf f_p$$0-fermion Hilbert space to a Banach space $$\mathbf A_p \dot{\mathbf d}_p + \mathbf B_p \mathbf d_p = \mathbf f_p$$1 of one-body reduced density matrices with finite kinetic energy, and then further to a density space $$\mathbf A_p \dot{\mathbf d}_p + \mathbf B_p \mathbf d_p = \mathbf f_p$$2 through a rigorously defined diagonal map $$\mathbf A_p \dot{\mathbf d}_p + \mathbf B_p \mathbf d_p = \mathbf f_p$$3. The universal reduced density matrix functional $$\mathbf A_p \dot{\mathbf d}_p + \mathbf B_p \mathbf d_p = \mathbf f_p$$4 and the induced density functional $$\mathbf A_p \dot{\mathbf d}_p + \mathbf B_p \mathbf d_p = \mathbf f_p$$5 are convex, lower semicontinuous objects on these reduced spaces, and the ground-state energy is expressed by Legendre–Fenchel duality with respect to the dual spaces of Hermitian forms and scalar potentials (Fredheim et al., 14 Oct 2025).

These constructions indicate that reduced-space formulation need not preserve the original variable type. A plausible implication is that reduction may consist in replacing a detailed state description by a lower-order observable description, provided an exact or variationally controlled map connects the two levels.

6. Preserved structure, computational gains, and limitations

A central technical issue is what the reduction preserves. In the thermomechanical state-space setting, the reduction is built to preserve symmetry of the projected system, with $$\mathbf A_p \dot{\mathbf d}_p + \mathbf B_p \mathbf d_p = \mathbf f_p$$6 and $$\mathbf A_p \dot{\mathbf d}_p + \mathbf B_p \mathbf d_p = \mathbf f_p$$7 inherited by the reduced formulation; the thermal update via residual flexibility is specifically designed to preserve multiphysics coupling effects (Ahn et al., 2021). In space-time least-squares reduced-basis methods, the reduced space is constructed so that symmetry, coercivity, and continuity of the full bilinear form are retained, enabling residual-based certification with absolute and relative error bounds (Hinze et al., 29 Jan 2026). In reduced OPF, the feasible-path algorithm preserves the physical power-flow constraints at every iteration because the state is recomputed from the control at each trial point (Pacaud et al., 2021).

The literature also reports explicit computational advantages. For the 3D flange-pipe thermomechanical example, full coupled mode superposition costs about $$\mathbf A_p \dot{\mathbf d}_p + \mathbf B_p \mathbf d_p = \mathbf f_p$$8 s, uncoupled reduction about $$\mathbf A_p \dot{\mathbf d}_p + \mathbf B_p \mathbf d_p = \mathbf f_p$$9 s, and the proposed MMS about $\mathbf d_p \in \mathbb R^{2N_d+N_a}$0 s, so the reduced multiphysics formulation is much cheaper than the full coupled solution while remaining close to it in accuracy (Ahn et al., 2021). In the heat-equation space-time Petrov–Galerkin experiments, RKSM takes about $\mathbf d_p \in \mathbb R^{2N_d+N_a}$1 s versus about $\mathbf d_p \in \mathbb R^{2N_d+N_a}$2 s for Crank–Nicolson at $\mathbf d_p \in \mathbb R^{2N_d+N_a}$3, and about $\mathbf d_p \in \mathbb R^{2N_d+N_a}$4 s versus about $\mathbf d_p \in \mathbb R^{2N_d+N_a}$5 s at $\mathbf d_p \in \mathbb R^{2N_d+N_a}$6, illustrating the advantage of low-rank matrix-oriented reduction over time stepping in those cases (Henning et al., 2019). In real-time OPF tracking, reported update times are about $\mathbf d_p \in \mathbb R^{2N_d+N_a}$7 s for case1354, $\mathbf d_p \in \mathbb R^{2N_d+N_a}$8 s for case2869, and $\mathbf d_p \in \mathbb R^{2N_d+N_a}$9 s for case9241, which is presented as compatible with minute-level control cycles (Pacaud et al., 2021). In GPU-condensed reduced-space IPM, the reduced algorithms are reported to be up to $W^d = \mathrm{span}\{\psi_1,\dots,\psi_d\}$0 times faster than Knitro when the relative number of degrees of freedom is small, while performing poorly when that number is large (Pacaud et al., 2022).

The same papers emphasize that reduced-space formulations are conditional rather than universal. The thermomechanical MMS assumes linear thermoelasticity, small deformations, and a symmetric coupling relation $W^d = \mathrm{span}\{\psi_1,\dots,\psi_d\}$1; residual flexibility also makes the updated thermal matrix denser (Ahn et al., 2021). Reduced OPF depends on nonsingularity of $W^d = \mathrm{span}\{\psi_1,\dots,\psi_d\}$2, so the implicit mapping $W^d = \mathrm{span}\{\psi_1,\dots,\psi_d\}$3 can break down near singular power-flow Jacobians (Pacaud et al., 2021). Adiabatic-elimination reductions require a spectral gap, exponential convergence of the unperturbed dynamics, and weak coupling so that perturbation theory in $W^d = \mathrm{span}\{\psi_1,\dots,\psi_d\}$4 remains valid (Régent et al., 2023). DGSM-based reduced input spaces require differentiability of the quantity of interest, smooth parameter dependence, and independent inputs in the stated framework (Vohra et al., 2018). Condensed reduced-space IPMs may lose their advantage when the number of controls is not small relative to the state dimension, because the dense reduced system and its assembly cease to be favorable (Pacaud et al., 2022).

In consequence, reduced-space formulation is best viewed as a structural strategy rather than a single method. Its success depends on identifying a smaller space that is not merely lower-dimensional, but also compatible with the analytic, geometric, or computational invariants that matter for the application.