SAV-ZEC Reformulation: Models & Methods

• SAV-ZEC Reformulation is a family of auxiliary-variable methods that encode challenging structures via low-dimensional variables and closure constraints.
• In mixed-integer convex optimization, it aggregates unit-level variables and applies perspective reformulation with zero-enforced closure to yield exact convex hulls.
• In gradient flows and CHNS systems, the method transforms nonlinear energy contributions into scalar or low-rank corrections, ensuring modified energy laws and efficient numerical schemes.

Searching arXiv for the cited papers to ground the article in current preprints. “SAV-ZEC Reformulation” is not a single established acronym with a uniform technical meaning across the arXiv literature. Rather, it designates several auxiliary-variable reformulations whose common role is to encode structure that is difficult to handle directly: symmetry and on/off disjunctions in mixed-integer convex optimization, segment activation in piecewise-convex MINLP relaxations, nonlinear energy contributions in gradient flows, exact global constraints in constrained dissipative systems, or cancellation of nonlinear couplings in multiphysics phase-field models. A plausible unifying interpretation is that these reformulations introduce low-dimensional auxiliary quantities and pair them with a closure, constraint, or cancellation mechanism so that the transformed model admits an exact convex-hull description, a modified-energy law, or a linear decoupled discretization, depending on the application (Wu et al., 4 Feb 2026, Trindade et al., 2022, Zhang et al., 17 Jun 2026, Cheng et al., 2019, Sun et al., 28 Jul 2025).

1. Terminological status and domain-specific meanings

The acronym is explicitly nonstandard in several of the relevant papers. In “Variable Aggregation-based Perspective Reformulation for Mixed-Integer Convex Optimization with Symmetry,” the authors state that SAV-ZEC is not an established acronym and adopt the mapping SAV = Symmetry-Aware Variable aggregation and ZEC = Zero-Enforced Closure (Wu et al., 4 Feb 2026). In “Scalar-Tracking SAV Schemes with Pullback Corrections for Gradient Flows,” the acronym ZEC does not appear, and the paper instead presents the pullback-corrected SAV framework as the relevant reformulation of SAV (Zhang et al., 17 Jun 2026). In “Scalar auxiliary variable approach for conservative/dissipative partial differential equations with unbounded energy,” the acronym likewise does not appear; the supplied interpretation labels the extended quadratic gradient-system representation as “zero-energy-conserving/zero-energy-consistent” in an explanatory sense rather than as paper terminology (Kemmochi et al., 2021). In “Global constraints preserving SAV schemes for gradient flows,” ZEC is interpreted as “Zero-Error Constraints,” again as an explanatory label for exact discrete constraint preservation (Cheng et al., 2019). By contrast, in the CHNS setting the acronym is tied directly to the “zero energy contribution” technique (Sun et al., 28 Jul 2025).

Domain	SAV meaning	ZEC meaning or status
Symmetric mixed-integer convex optimization	Symmetry-Aware Variable aggregation	Zero-Enforced Closure
Piecewise-convex MINLP relaxations	Single auxiliary epigraph variable per convex segment	Zero-enforcing segment constraints
Gradient flows with pullback corrections	Scalar auxiliary variable	ZEC not used
Variational PDEs with unbounded energy	Scalar auxiliary variable	Interpretive “zero-energy-conserving/consistent” view
Constrained gradient flows	Scalar auxiliary variable	Zero-Error Constraints
CHNS	Scalar auxiliary variable	Zero energy contribution

This multiplicity of meanings is important. A persistent misconception is that SAV-ZEC denotes one canonical algorithmic template; the literature instead shows several non-equivalent constructions that share only a family resemblance.

2. Symmetry-aware variable aggregation and zero-enforced closure in mixed-integer convex optimization

In the mixed-integer convex optimization setting, the reformulation is defined for symmetric groups of identical units. For a group $G$ of $n$ identical units indexed by $i=1,\ldots,n$, each unit has an on/off indicator $y_i \in \{0,1\}$ and a continuous operating vector $x_i \in \mathbb{R}^d$. The per-unit cost $f(x_i)$ is closed convex and separable, and all units in the group share identical bounds and constraints. Valid aggregation requires identical cost and per-unit constraints, exchangeability of units, boundedness of per-unit feasible sets, and a mild interiority assumption for perspective tightness (Wu et al., 4 Feb 2026).

The aggregated variables are

$$s=\sum_{i=1}^n y_i \in \{0,1,\ldots,n\}, \qquad X=\sum_{i=1}^n x_i \in \mathbb{R}^d.$$

In the multi-block setting, the group-level variables are $Y^s=\sum_{i=1}^n y_i^s$ and $X_s=\sum_{i=1}^n x_{is}$. Aggregation replaces the Cartesian product of $n$ copies of the per-unit feasible set by the Minkowski sum $n$0, and the perspective reformulation is then applied at the aggregated level. The natural aggregated objective term becomes $n$1 for $n$2, with the closure at $n$3 enforcing $n$4 in bounded-domain cases. This is the “Zero-Enforced Closure” component: the closure of the perspective epigraph captures the limiting behavior at $n$5 and forces the aggregate operating vector to vanish when no unit is on (Wu et al., 4 Feb 2026).

The central structural result is the exact convex hull theorem

$n$6

under boundedness and Assumption 1. The paper states that aggregation and perspective commute in convex hull, so that aggregating first and then applying perspective yields the same convex hull as applying per-unit perspective first and then aggregating. In the paper’s comparative notation, the relaxation bounds satisfy

$n$7

This formalizes the claim that aggregation plus perspective gives maximal tightness for symmetric groups and dominates standard aggregation without perspective (Wu et al., 4 Feb 2026).

The canonical quadratic case makes the closure explicit. For $n$8 with $n$9, the aggregated perspective epigraph is

$i=1,\ldots,n$0

with closure at $i=1,\ldots,n$1 enforcing $i=1,\ldots,n$2 and $i=1,\ldots,n$3 when the domain is bounded. The paper gives the rotated-cone representation

$i=1,\ldots,n$4

where $i=1,\ldots,n$5. Piecewise-linear convex costs are modeled by linear epigraph inequalities $i=1,\ldots,n$6, and general convex costs are handled through epigraph scaling $i=1,\ldots,n$7 for $i=1,\ldots,n$8 (Wu et al., 4 Feb 2026).

The reformulation is applied to unit commitment with identical generators, the line cover problem with convex quadratic costs, and separable mixed-integer convex quadratic programs. The reported computations are specific: on UC (Ostrowski instances), the aggregated perspective formulation “solved all 20 instances within 36 seconds,” while the 3-bin and per-unit perspective formulations “often timed out at 1200s”; the UC-agg relaxation bounds matched those of UC-per up to numerical tolerance and were significantly tighter than 3-bin. On line cover instances with $i=1,\ldots,n$9 and $y_i \in \{0,1\}$0–$y_i \in \{0,1\}$1 identical replicas, LC-agg was consistently faster with fewer B&B nodes than LC and LC-per, and on separable mixed-integer convex QPs with $y_i \in \{0,1\}$2, SQP-agg showed substantial speedups and node reductions versus SQP and SQP-per (Wu et al., 4 Feb 2026).

The exactness is not unconditional. The paper states that symmetry is essential, heterogeneous units must be partitioned into homogeneous classes, and nonseparable constraints across units or time can obstruct exact aggregation. The result also depends on boundedness and on the interiority assumption used for per-unit perspective tightness (Wu et al., 4 Feb 2026).

3. Piecewise-convex perspective reformulations and segmentwise zero-enforcing constraints

A different optimization usage appears in “Comparing perspective reformulations for piecewise-convex optimization.” There, the relevant structure is a lower-bounding subproblem arising in Sequential Convex MINLP for separable nonconvex univariate functions. The supplied mapping identifies the SAV-ZEC-style formulation with the Multiple-Choice Model strengthened by perspective epigraphs and zero-enforcing interval constraints (Trindade et al., 2022).

For a function $y_i \in \{0,1\}$3 on a partitioned domain, convex segments are reinforced by perspective, while concave segments are replaced by piecewise-linear upper bounds. In the paper’s notation, the MCM uses segment variables $y_i \in \{0,1\}$4, activation binaries $y_i \in \{0,1\}$5, and epigraph auxiliaries $y_i \in \{0,1\}$6. For convex segments $y_i \in \{0,1\}$7,

$y_i \in \{0,1\}$8

while the zero-enforcing interval constraints are

$y_i \in \{0,1\}$9

These constraints force $x_i \in \mathbb{R}^d$0 when $x_i \in \mathbb{R}^d$1 and ensure that exactly one segment is active (Trindade et al., 2022).

The paper compares three standard formulations: the Incremental Model (IM), the Multiple-Choice Model (MCM), and the Convex Combination Model (CCM). The critical result is that, unlike the piecewise-linear case, these formulations are not equivalent when perspective is used on the convex segments. MCM describes the convex envelope of each $x_i \in \mathbb{R}^d$2, and CCM has an equivalent continuous relaxation. IM is weaker in general. Example 1, based on $x_i \in \mathbb{R}^d$3 over $x_i \in \mathbb{R}^d$4, shows that the IM relaxation can be strictly weaker than the MCM relaxation. Proposition 1 gives a narrow equivalence case: if the function has exactly two segments ordered concave then convex, the continuous relaxations of IM and MCM are equivalent; the equivalence fails when the order is convex then concave (Trindade et al., 2022).

The computational evidence is correspondingly sharp. For nonlinear continuous knapsack instances with four intervals and $x_i \in \mathbb{R}^d$5, the paper reports an IM root relaxation gap of approximately $x_i \in \mathbb{R}^d$6 versus approximately $x_i \in \mathbb{R}^d$7 for MCM, with total times of approximately $x_i \in \mathbb{R}^d$8s versus $x_i \in \mathbb{R}^d$9s. For nonlinear uncapacitated facility location, MCM always outperformed IM in relaxation gaps and frequently in running time; in the hardest three-interval cases, MCM reduced the root relaxation gap from approximately $f(x_i)$0 to approximately $f(x_i)$1–$f(x_i)$2 (Trindade et al., 2022).

In this setting, “ZEC” is not a closure at a zero count, as in symmetry-aware aggregation. It is the segmentwise implication $f(x_i)$3, embedded in an exactly-one-alternative structure. The reformulation is therefore algebraically closer to on/off convexification than to energy-based SAV methods.

4. Scalar auxiliary variables for gradient flows and variational PDEs

In gradient-flow numerics, SAV denotes the scalar auxiliary variable method, but the ZEC label is absent in some of the most relevant papers. “Scalar-Tracking SAV Schemes with Pullback Corrections for Gradient Flows” considers gradient flows

$f(x_i)$4

with $f(x_i)$5 self-adjoint and nonpositive and $f(x_i)$6 self-adjoint and positive semidefinite. In first-order SAV, the scalar auxiliary variable $f(x_i)$7 produces, after eliminating the scalar, a semi-implicit state equation augmented by a rank-one positive semidefinite correction

$f(x_i)$8

The paper’s main reformulation, PB-SAV, keeps a single scalar $f(x_i)$9 tracking the total nonlinear energy but replaces the rank-one correction by a pullback correction

$$s=\sum_{i=1}^n y_i \in \{0,1,\ldots,n\}, \qquad X=\sum_{i=1}^n x_i \in \mathbb{R}^d.$$0

which remains positive semidefinite and has rank at most the number of components. This decouples the number of scalar variables from the correction rank. The paper proves modified-energy dissipation laws for fixed and step-dependent decompositions, gives a refinement identity in which the gain over rank-one SAV is an explicit weighted variance, shows that $$s=\sum_{i=1}^n y_i \in \{0,1,\ldots,n\}, \qquad X=\sum_{i=1}^n x_i \in \mathbb{R}^d.$$1, and identifies the correction with the Gauss-Newton matrix of a least-squares representation of the nonlinear energy in finite dimensions. The low-rank perturbation is implemented by a Sherman-Morrison-Woodbury formula (Zhang et al., 17 Jun 2026).

The numerical evidence in that paper is mixed in a precise sense. Along smooth trajectories, PB-SAV preserves first-order consistency and often changes the first-order error constant; when component forces have distinct directions or scales, it can substantially improve trajectory accuracy. The reported tests include finite-dimensional gradient flows, Allen-Cahn dynamics, and nonlocal Cahn-Hilliard models, with regimes in which higher-rank PB-SAV reduced final $$s=\sum_{i=1}^n y_i \in \{0,1,\ldots,n\}, \qquad X=\sum_{i=1}^n x_i \in \mathbb{R}^d.$$2 error markedly relative to rank-one SAV (Zhang et al., 17 Jun 2026).

A broader SAV reformulation for variational PDEs with unbounded energy appears in “Scalar auxiliary variable approach for conservative/dissipative partial differential equations with unbounded energy.” The paper decomposes

$$s=\sum_{i=1}^n y_i \in \{0,1,\ldots,n\}, \qquad X=\sum_{i=1}^n x_i \in \mathbb{R}^d.$$3

with $$s=\sum_{i=1}^n y_i \in \{0,1,\ldots,n\}, \qquad X=\sum_{i=1}^n x_i \in \mathbb{R}^d.$$4 self-adjoint positive semidefinite and $$s=\sum_{i=1}^n y_i \in \{0,1,\ldots,n\}, \qquad X=\sum_{i=1}^n x_i \in \mathbb{R}^d.$$5 lower-bounded. Two scalar auxiliary variables

$$s=\sum_{i=1}^n y_i \in \{0,1,\ldots,n\}, \qquad X=\sum_{i=1}^n x_i \in \mathbb{R}^d.$$6

yield the coupled system

$$s=\sum_{i=1}^n y_i \in \{0,1,\ldots,n\}, \qquad X=\sum_{i=1}^n x_i \in \mathbb{R}^d.$$7

The paper rewrites this as an extended quadratic gradient system $$s=\sum_{i=1}^n y_i \in \{0,1,\ldots,n\}, \qquad X=\sum_{i=1}^n x_i \in \mathbb{R}^d.$$8 on $$s=\sum_{i=1}^n y_i \in \{0,1,\ldots,n\}, \qquad X=\sum_{i=1}^n x_i \in \mathbb{R}^d.$$9 with quadratic modified energy

$Y^s=\sum_{i=1}^n y_i^s$0

Lemma 1 states that $Y^s=\sum_{i=1}^n y_i^s$1 inherits the conservative or dissipative character of $Y^s=\sum_{i=1}^n y_i^s$2: skew-adjoint $Y^s=\sum_{i=1}^n y_i^s$3 gives energy conservation, and negative semidefinite $Y^s=\sum_{i=1}^n y_i^s$4 gives energy dissipation. On that basis the paper constructs a second-order Crank-Nicolson-type SAV scheme and fourth-order SAV-Runge-Kutta schemes for conservative systems. For the KdV equation, the modified energy was numerically conserved, and the convergence studies gave $Y^s=\sum_{i=1}^n y_i^s$5 solution errors for the CN-type schemes and $Y^s=\sum_{i=1}^n y_i^s$6 for SAV-RK (Kemmochi et al., 2021).

These two papers show that, in the gradient-flow literature, the distinctive role of SAV is not zero-enforcing closure in the mixed-integer sense. It is the transformation of nonlinear energy terms into scalar-tracked or low-rank-corrected structures that preserve a modified energy law and permit linearly implicit solvers.

5. Exact global constraints and zero energy contribution formulations

A more literal ZEC interpretation appears in constrained gradient flows. “Global constraints preserving SAV schemes for gradient flows” considers

$Y^s=\sum_{i=1}^n y_i^s$7

with global constraints $Y^s=\sum_{i=1}^n y_i^s$8. The paper combines SAV with Lagrange multipliers and presents three fully discrete approaches. All of them use constant-coefficient linear solves at each time step together with a small nonlinear algebraic system for multipliers and auxiliary variables. The first approach gives a first-order scheme with exact constraint preservation and a modified-energy inequality; the second introduces a multiplier $Y^s=\sum_{i=1}^n y_i^s$9 and uses a Crank-Nicolson discretization to obtain unconditional energy stability with respect to the original energy; the third determines $X_s=\sum_{i=1}^n x_{is}$0 first and then $X_s=\sum_{i=1}^n x_{is}$1, retaining unconditional energy stability while improving robustness. The paper states that the first and third approaches are more efficient and robust than the second approach. Applications include the phase-field vesicle membrane model, where volume and surface area are preserved exactly, and the optimal partition model, where each norm constraint is preserved exactly (Cheng et al., 2019).

In the CHNS setting, the SAV-ZEC nomenclature is explicit. “Convergence analysis of a second-order SAV-ZEC scheme for the Cahn-Hilliard-Navier-Stokes system” introduces the SAV variable

$X_s=\sum_{i=1}^n x_{is}$2

together with the zero energy contribution variable $X_s=\sum_{i=1}^n x_{is}$3, defined by

$X_s=\sum_{i=1}^n x_{is}$4

At the continuous level, incompressibility and the boundary conditions imply $X_s=\sum_{i=1}^n x_{is}$5, but at the discrete level the $X_s=\sum_{i=1}^n x_{is}$6-equation is used to cancel nonlinear coupling terms in the energy budget. The reformulated total energy is

$X_s=\sum_{i=1}^n x_{is}$7

The fully discrete scheme combines MAC finite differences, BDF2 in time, Adams-Bashforth extrapolation for nonlinear terms, and pressure correction for the Stokes part, so that only constant-coefficient Poisson-like solvers are needed. The paper proves unconditional stability with respect to a rewritten total energy functional and derives optimal convergence rates: the phase variable is controlled in the $X_s=\sum_{i=1}^n x_{is}$8 norm and the velocity in the $X_s=\sum_{i=1}^n x_{is}$9 norm (Sun et al., 28 Jul 2025).

The constrained-gradient-flow and CHNS papers therefore represent two distinct ZEC mechanisms. In one case, ZEC means exact enforcement of global constraints by solving for Lagrange multipliers at each time step. In the other, it means adding an auxiliary scalar whose discrete evolution neutralizes nonlinear energy contributions that would otherwise obstruct unconditional stability and decoupling.

6. Comparative structure, neighboring SAV formulations, and limitations

Across these literatures, the technical role of the auxiliary variable differs substantially. In the mixed-integer convex optimization papers, the auxiliary quantity is an aggregated count or an epigraph variable, and the key effect is convexification of on/off structure under symmetry or segment activation. In the deterministic and stochastic gradient-flow papers, the auxiliary quantity tracks nonlinear energy, and the key effect is the replacement of nonlinear terms by low-rank or scalar-coupled updates that preserve a modified energy law. In the CHNS paper, the additional ZEC scalar is not an energy tracker but a cancellation variable for nonlinear couplings. This suggests that “SAV-ZEC Reformulation” is best understood as a family of structurally analogous but mathematically non-equivalent devices rather than as a single formulation.

A related neighboring construction appears in “Semi-implicit energy-preserving numerical schemes for stochastic wave equation via SAV approach.” That paper studies the stochastic wave equation with multiplicative noise and introduces a stochastic SAV

$n$0

thereby transforming the system into a higher-dimensional stochastic system with modified energy

$n$1

The paper proposes midpoint and exponential-Euler SAV schemes, both semi-implicit and explicitly solvable after a scalar reduction, and proves that they preserve the averaged modified energy evolution law in discrete time. Under the stated assumptions, the exponential Euler SAV scheme attains strong order $n$2 in time, and the fully discrete FEM extension satisfies

$n$3

The paper does not use ZEC terminology, but it belongs to the same auxiliary-variable lineage in which nonlinear structure is transferred into scalar equations that can be balanced against the linearized state update (Cui et al., 2022).

Several limitations recur. Exact convex-hull claims in the optimization papers depend on symmetry, boundedness, and mild interiority assumptions; heterogeneous units or nonseparable cross-unit couplings break the exact theory unless additional structure is present (Wu et al., 4 Feb 2026). Segmentwise perspective convexification is strongest for MCM/CCM; incremental formulations can be substantially weaker outside the special two-segment concave-then-convex case (Trindade et al., 2022). In the SAV discretization papers, admissible decompositions, positivity of radicands, and regularity assumptions are integral to both stability and convergence arguments (Zhang et al., 17 Jun 2026, Kemmochi et al., 2021). Exact constraint preservation in constrained gradient flows requires solving small nonlinear algebraic systems for multipliers, and the well-posedness or robustness of those systems is not uniformly resolved in every formulation (Cheng et al., 2019). In CHNS, unconditional stability does not remove the usual distinction between stability and accuracy; very large time steps remain inaccurate even when the rewritten energy decays (Sun et al., 28 Jul 2025).

The main conceptual point is therefore negative as much as positive: SAV-ZEC is not a universally fixed acronym. Its substantive content must be read from the host problem class. In symmetric mixed-integer convex optimization it means aggregation plus perspective with closure at zero count; in piecewise-convex MINLP it means perspective epigraphs plus zero-enforcing segment activation; in constrained dissipative PDEs it means scalar auxiliary variables combined with exact discrete constraints; in CHNS it means scalar energy tracking together with a zero energy contribution variable; and in several neighboring SAV papers the ZEC label is absent altogether even though the reformulation is structurally adjacent.