Exact Reformulation and Optimization

• Exact Reformulation and Optimization (ERO) is a family of techniques that transform complex optimization problems into alternative, exact reformulations exposing exploitable structure.
• These methods employ mechanisms such as auxiliary-variable lifting and conic dualization to preserve minimizers and feasible sets, ensuring mathematical equivalence.
• ERO approaches facilitate algorithmic advances in sparse optimization, robust control, and bilevel programming, while addressing scalability and precision challenges.

Exact Reformulation and Optimization (ERO) denotes a family of methodologies in which a difficult optimization problem is converted into an alternative formulation that preserves the original problem in a mathematically explicit sense while exposing structure that can be exploited algorithmically. In the arXiv literature, the term has been used for minimizer-preserving reformulations of $\ell_0$-sparse models, exact conic reformulations of adjustable robust problems, exact hull reformulations of generalized disjunctive programs, exact continuous reformulations of logic constraints, and exact-penalty reductions of bilevel and direct metric optimization problems (Bechensteen et al., 2019, Woolnough et al., 2020, Gusev et al., 22 Aug 2025, Wehbeh et al., 7 Jan 2026, Zheng et al., 31 May 2026, Peng et al., 21 Jul 2025). This suggests that ERO is not a single technique but a research idiom centered on equivalence-preserving problem transformation.

1. Exactness as an optimization concept

Within ERO, “exact” is not used in only one sense. In some works it means feasible-set and optimal-value equivalence after projection, as in exact SDP/SOCP reformulations of adjustable robust optimization and exact deterministic convex reformulations of linear stochastic optimal control with linear chance constraints (Woolnough et al., 2020, Dokania et al., 19 Mar 2026). In others it means correspondence of local or global minimizers between the original model and a lifted model, as in exact $\ell_0$ reformulations for sparse least squares and general constrained sparse optimization (Bechensteen et al., 2019, Kanzow et al., 3 Sep 2025). In exact-penalty settings, exactness means that a finite penalty parameter, rather than an asymptotic limit, is sufficient to recover solutions of the original constrained problem (Zheng et al., 31 May 2026, Tin et al., 2021). In still other settings it means that the transformation preserves the mathematical problem while changing only its representation, as in LP variable permutation learned for solver acceleration (Li et al., 2022).

This plurality matters. ERO is not synonymous with convexification, relaxation, or approximation. Several neighboring papers make the boundary explicit. The binary reformulation and linearization framework for bounded polynomial programming is described as an approximate method with provable lower and upper bounds and asymptotic convergence as error tolerances approach zero, rather than as an exact reformulation (Norman, 2012). PolyFormer is presented as a learned polyhedral approximation framework whose feasible-set surrogate may be inner, outer, or intermediate, and therefore is not a classical exact reformulation in general (Wen et al., 9 Mar 2026). Likewise, in stochastic control the linear chance-constrained part is exact, whereas the quadratic chance-constrained part is handled only by conservative convex approximations (Dokania et al., 19 Mar 2026).

A recurring ERO distinction is therefore between exact reformulation and exact optimization. Several papers prove exactness for minimizers or feasible sets while using algorithms that are guaranteed only to converge to critical points or stationary points of the reformulated model. This gap is explicit in sparse biconvex reformulation, in exact-penalty sparse optimization, and in prox-linear bilevel methods (Bechensteen et al., 2019, Kanzow et al., 3 Sep 2025, Zheng et al., 31 May 2026).

2. Principal reformulation mechanisms

The literature uses a small number of recurring mechanisms to obtain exactness.

Mechanism	Representative reformulation	Exactness notion
Auxiliary-variable lifting	Replace $\|x\|_0$ by $\min_{-{\bf 1}\le u\le {\bf 1}}\|u\|_1$ subject to $\|x\|_1=\langle u,x\rangle$ (Bechensteen et al., 2019)	Correspondence of minimizers after exact penalization
Conic dualization via the $\mathcal S$-lemma	Convert robust quadratic inequalities under ellipsoidal uncertainty into LMIs, and in the separable subclass into SOC constraints (Woolnough et al., 2020)	Equivalent feasible sets after projection and identical optimal values
Exact hull closure for disjunctions	Replace the perspective closure of a quadratic disjunctive constraint by $$\mathbf v^\top Q\mathbf v+(\mathbf c^\top\mathbf v)y+dy^2\le 0$$ (Gusev et al., 22 Aug 2025)	Equality of the reformulated set with the conventional exact hull set
Logic-to-CNF continuous reformulation	Rewrite logic into CNF, encode OR by $\min$, conjunction by $\max$, then smooth exactly with auxiliary simplex variables (Wehbeh et al., 7 Jan 2026)	Same feasible set as the original logic-constrained model
Gradient-residual exact penalty	Replace $y\in S(x)$ by $\ell_0$0 under a PL-type condition and a distance-bound argument (Zheng et al., 31 May 2026)	Same global solutions for sufficiently large finite $\ell_0$1
Representation-preserving permutation	Permute LP variable clusters while keeping the mathematical LP unchanged (Li et al., 2022)	Feasible-set and optimal-value preservation by construction

Two additional mechanisms broaden this taxonomy. In direct metric optimization for binary imbalanced classification, the indicator relation $\ell_0$2 is replaced exactly by the continuous piecewise-linear identity

$\ell_0$3

under the non-singularity condition $\ell_0$4, and then relaxed to class-dependent inequalities that remain exact because of monotonicity of the target metrics (Peng et al., 21 Jul 2025). In constrained sparse optimization with free-sign variables, $\ell_0$5 regularization is reformulated exactly through complementarity $\ell_0$6, an auxiliary separable objective $\ell_0$7, and then penalized by $\ell_0$8, avoiding variable splitting and preserving one-to-one local/global minimizer correspondence (Kanzow et al., 3 Sep 2025).

These mechanisms share a common structure. A hard combinatorial, logical, robust, or hierarchical condition is lifted into additional variables, dual multipliers, or structured penalties; the lifted relation is then represented by convex blocks, bilinear couplings, perspective closures, or continuous piecewise-linear identities; and exactness is proved by correspondence theorems rather than by heuristic approximation arguments.

3. Optimization after exact reformulation

ERO is usually valuable only when the reformulated model admits an exploitable algorithmic structure. In sparse least squares, the exact penalty

$\ell_0$9

is biconvex, so the proposed method combines continuation in $\|x\|_0$0, Proximal Alternating Minimization, FISTA for the $\|x\|_0$1-subproblem, and either box/$\|x\|_0$2-projection or a coordinatewise closed form for the $\|x\|_0$3-subproblem (Bechensteen et al., 2019). In the broader constrained sparse setting, the exact penalty subproblems are mildly nonsmooth because of $\|x\|_0$4, and the paper therefore develops both an epigraph-based spectral projected-gradient scheme and a proximal treatment with a closed-form scalar prox for $\|x\|_0$5 (Kanzow et al., 3 Sep 2025).

In bilevel optimization, exact reformulation is paired with algorithm design in two rather different ways. One line uses partial calmness and lower-level value-function reformulation, then solves the resulting smoothed Fischer–Burmeister equation system by a Levenberg–Marquardt method; the numerically critical issue there is the partial exact penalty parameter $\|x\|_0$6 (Tin et al., 2021). Another line uses the exact penalty $\|x\|_0$7, establishes a threshold $\|x\|_0$8, and then applies an exact-penalty prox-linear method; in the simple bilevel case, each prox-linear subproblem admits a dual box-constrained quadratic program solved by dual SPG with explicit primal recovery (Zheng et al., 31 May 2026).

Conic exact reformulations often shift the computational burden into standard solver classes. General QDR-based adjustable robust optimization becomes an SDP, and diagonal QDRs yield an exact SOCP, so the optimization stage is delegated to conic solvers rather than to custom nonlinear routines (Woolnough et al., 2020). Exact hull reformulation for quadratic GDPs preserves quadratic structure, allowing MIQCP/MINLP solvers to operate directly on the exact closure model instead of on an $\|x\|_0$9-approximation with fractional nonlinearities (Gusev et al., 22 Aug 2025). In dual adaptive MPC, strong duality converts predicted set-membership updates into bilinear algebraic constraints inside MPC, producing a nonconvex but finite-dimensional optimization problem implementable with homothetic or flexible tubes (Parsi et al., 2022).

Not all recent work follows the “derive a deterministic equivalent and solve it” pattern. The max-min-max algorithm for large-scale robust optimization is explicitly motivated by the limitations of the reformulation approach; it keeps the exact Lagrangian equivalence

$\min_{-{\bf 1}\le u\le {\bf 1}}\|u\|_1$0

but solves it directly with nested first-order methods using only subgradient and projection oracles, achieving oracle complexity $\min_{-{\bf 1}\le u\le {\bf 1}}\|u\|_1$1 in the nonsmooth case and $\min_{-{\bf 1}\le u\le {\bf 1}}\|u\|_1$2 in the smooth case (Tu et al., 2024). This suggests that ERO also includes exact variational reformulations that are exploited algorithmically without ever materializing a classical deterministic robust counterpart.

4. Domain-specific realizations

Sparse optimization is one of the clearest ERO domains. In single-molecule localization microscopy, exact biconvex reformulation of penalized and $\min_{-{\bf 1}\le u\le {\bf 1}}\|u\|_1$3-sparse least squares yielded better reconstructions than Iterative Hard Thresholding both visually and numerically; on the simulated ISBI 2013 dataset, the biconvex constrained method achieved Jaccard scores $\min_{-{\bf 1}\le u\le {\bf 1}}\|u\|_1$4 versus $\min_{-{\bf 1}\le u\le {\bf 1}}\|u\|_1$5 at $\min_{-{\bf 1}\le u\le {\bf 1}}\|u\|_1$6 nm and $\min_{-{\bf 1}\le u\le {\bf 1}}\|u\|_1$7 versus $\min_{-{\bf 1}\le u\le {\bf 1}}\|u\|_1$8 at $\min_{-{\bf 1}\le u\le {\bf 1}}\|u\|_1$9 nm, while the biconvex penalized method also dominated penalized IHT, e.g. $\|x\|_1=\langle u,x\rangle$0 versus $\|x\|_1=\langle u,x\rangle$1 at $\|x\|_1=\langle u,x\rangle$2 nm (Bechensteen et al., 2019). The later general sparse-optimization paper extends this line from nonnegative least squares to free-sign constrained models and reports competitive or superior performance in sparse portfolio optimization, sparse dictionary learning, and sparse adversarial attacks (Kanzow et al., 3 Sep 2025).

Adjustable robust optimization provides a different ERO template. Once the recourse is restricted to the parameterized quadratic decision rule

$\|x\|_1=\langle u,x\rangle$3

the robust counterpart over an ellipsoid is exactly representable as an SDP, and in the separable diagonal subclass as an SOCP (Woolnough et al., 2020). In distributionally robust optimization over finite supports, discrete DRO with weighted $\|x\|_1=\langle u,x\rangle$4 or density-ratio ambiguity balls is reformulated into single-layer smooth convex programs, and these reformulations further identify exact links to mean-plus-standard-deviation optimization and CVaR minimization, respectively (Shida et al., 21 Oct 2025). In adaptive MPC, strong duality is used to embed future set-membership identification exactly into the MPC problem, enabling exploration–exploitation trade-offs through predicted uncertainty reduction rather than through ad hoc excitation constraints (Parsi et al., 2022).

Discrete and logic-driven optimization also admit exact reformulation patterns. For quadratic optimization over roots of unity, one contribution is theoretical exactness of the SOS hierarchy at level $\|x\|_1=\langle u,x\rangle$5, and another is an exact binary zonotope reformulation for even $\|x\|_1=\langle u,x\rangle$6 that halves the number of binary variables relative to one-hot encoding; the paper reports solution-time speedups of up to $\|x\|_1=\langle u,x\rangle$7 (Al-Sulami et al., 4 Aug 2025). For logic-constrained nonlinear optimization and optimal control, arbitrary Boolean formulas built from equality and inequality predicates are converted into exact continuous constraints without binary variables, and the method is evaluated on quadrotor obstacle avoidance and a hybrid two-tank system with temporal logic constraints (Wehbeh et al., 7 Jan 2026). For quadratically constrained GDPs, the exact hull reformulation preserves the original quadratic structure and avoids the $\|x\|_1=\langle u,x\rangle$8-approximation that had been commonly used to represent the closure of the perspective function (Gusev et al., 22 Aug 2025).

Machine-learning-related ERO appears in two contrasting forms. In binary imbalanced classification, exact constrained reformulations are derived for fix-precision-optimize-recall, fix-recall-optimize-precision, and $\|x\|_1=\langle u,x\rangle$9-score optimization, replacing indicator equalities by exact piecewise-linear continuous constraints and solving the result by exact penalty methods (Peng et al., 21 Jul 2025). In LP, variable reordering is treated as an exact reformulation because it preserves the problem while changing solver behavior; a reinforcement-learning policy over variable-cluster permutations reduced simplex iteration number by about $\mathcal S$0 and solving time by about $\mathcal S$1 on average across the reported datasets (Li et al., 2022). The two cases are methodologically very different, but both use “reformulation” in an exact, solver-conscious sense.

5. Boundary cases, misconceptions, and controversies

A common misconception is to equate ERO with any algebraic rewriting that improves tractability. The literature is stricter. The polynomial-programming framework based on binary reformulation and linearization is explicitly approximate: the variable encoding with remainder variables is exact at the variable level, but products of multiple remainder terms are replaced by their mean, so the method yields certified upper and lower bounds rather than an exact finite reformulation (Norman, 2012). PolyFormer is likewise a learned polyhedral approximation of feasible sets, with a tunable trade-off between inner and outer approximation and with no general exactness theorem for the original model (Wen et al., 9 Mar 2026). These works are closely related to ERO, but they define its boundary rather than its core.

Another misconception is that exactness is absolute rather than conditional. In adjustable robust optimization, the reformulation is exact only after recourse has been restricted to the chosen QDR class; the equivalence is with the robust problem under that policy class, not with unrestricted adjustable recourse (Woolnough et al., 2020). In stochastic control, the deterministic convex reformulation is exact for linear chance constraints under Gaussian linear dynamics and affine causal policies, but quadratic chance constraints are only approximated (Dokania et al., 19 Mar 2026). In bilevel optimization with $\mathcal S$2 lower-level gradient penalty, exactness depends on the PL-type condition that turns $\mathcal S$3 into lower-level global optimality, and on the threshold $\mathcal S$4 (Zheng et al., 31 May 2026). In value-function-based bilevel reformulation, partial exact penalization is local and tied to partial calmness, not a universal global statement (Tin et al., 2021).

There is also a persistent algorithmic controversy: exact reformulation does not by itself guarantee efficient solution. Sparse biconvex reformulation proves minimizer correspondence for sufficiently large $\mathcal S$5, but PAM is guaranteed only to converge to a critical point (Bechensteen et al., 2019). Exact hull reformulation strengthens closure modeling and often improves numerics, yet some nonconvex instances still favor Big-M or binary multiplication in runtime (Gusev et al., 22 Aug 2025). The large-scale robust-optimization paper is especially explicit on this point: exact deterministic reformulation may be mathematically elegant but computationally weak because it can introduce high-dimensional or highly constrained models, sometimes changing a quadratic robust constraint into an SDP (Tu et al., 2024).

A further misconception is that “exact” implies algebraic sophistication. The LP reordering paper shows the opposite extreme: a variable permutation is exact simply because it preserves feasible points, objective values, and recoverability of solutions by inverse permutation (Li et al., 2022). This suggests that ERO ranges from deep lifted-equivalence theorems to exact representation-level transformations at the solver interface.

6. Limitations and research directions

The current ERO literature is technically strong but assumption-heavy. Exactness theorems often require full-rank operators, Slater-type conditions, PL/error-bound structure, convexity-concavity, fixed recourse, single ellipsoidal uncertainty sets, or bounded-variable closures. Several papers identify these restrictions directly. The sparse least-squares reformulation assumes $\mathcal S$6 has full rank and requires $\mathcal S$7, while also noting that extension beyond least-squares fidelity remains open (Bechensteen et al., 2019). The QDR robust-optimization paper is confined to affine data dependence, fixed recourse, and a single ellipsoidal uncertainty set, and points to intersections of ellipsoids as future work (Woolnough et al., 2020). The exact-penalty prox-linear bilevel framework is restricted to unconstrained differentiable lower-level problems satisfying a PL-type condition (Zheng et al., 31 May 2026).

Scalability remains an active concern even when exactness is available. Direct metric optimization in imbalanced classification introduces one auxiliary variable and one samplewise constraint per data point and is currently implemented in a deterministic rather than stochastic manner (Peng et al., 21 Jul 2025). Exact hull reformulation for GDPs improves numerical stability relative to $\mathcal S$8-hull, but solver convexity recognition for the resulting quadratic forms is not automatic, and the authors identify better solver interfaces and broader polynomial extensions as open directions (Gusev et al., 22 Aug 2025). In free-sign sparse optimization, the new exact penalty avoids variable splitting but produces nonsmooth subproblems, motivating SPG and proximal inner solvers rather than a fully smooth scheme; differentiable exact penalties for this setting are left as future work (Kanzow et al., 3 Sep 2025).

A broader direction is the extension of exact reformulation beyond the currently dominant model classes. The logic-reformulation paper shows that arbitrary finite Boolean structure can be embedded exactly into continuous NLP/OCP models without binary variables (Wehbeh et al., 7 Jan 2026). The stochastic-control paper shows that exact deterministic convexification is possible for linear chance constraints across multiple time steps (Dokania et al., 19 Mar 2026). Together with exact set-membership dualization in adaptive MPC and exact conic reformulations in robust optimization, these results suggest a research trajectory in which discrete logic, uncertainty, hierarchy, and complementarity are treated by structure-preserving continuous models rather than by generic mixed-integer surrogates (Parsi et al., 2022, Woolnough et al., 2020).

Across these developments, the unifying ERO principle is stable: preserve the original optimization semantics exactly where possible, isolate the true source of nonconvexity or combinatorics, and reformulate it into a representation that exposes exploitable structure. What varies from domain to domain is the equivalence notion—feasible-set equality, minimizer correspondence, exact finite penalty, support-function equivalence, or representation invariance—and the computational vehicle used after reformulation. That plurality is not incidental; it is the defining feature of ERO as it currently exists in the optimization literature.