Optimal Kron Reduction (Opti-KRON)

• The paper introduces an MILP formulation that balances voltage error against the fraction of reduced nodes, achieving up to 94% reduction with minimal voltage deviation.
• Opti-KRON extends classical Kron reduction by dynamically selecting super nodes and applying graph-locality and voltage-error constraints to non-zero-injection buses in AC networks.
• Scalability and structure preservation are enhanced via a two-stage reduction process with a radialization step, enabling faster OPF computations in large-scale feeder networks.

Searching arXiv for the core Opti-KRON papers and closely related Kron-reduction literature. Optimal Kron-based Reduction of Networks, commonly abbreviated Opti-KRON, denotes an optimization-based extension of classical Kron reduction in which the reduced model is not determined solely by a fixed boundary/interior partition, but by a decision process that balances model size against fidelity of the reduced network. In the electric-power setting, the framework was introduced as a mixed-integer linear programming formulation that chooses super nodes and current aggregations using the full AC network together with a library of AC load-flow data, and it was later extended with scalability mechanisms and a radiality-preservation step for radial distribution feeders (Chevalier et al., 2022, Mokhtari et al., 20 Aug 2025). In a broader technical sense, Opti-KRON sits at the intersection of Schur-complement-based network elimination, graph Laplacian structure, and realizability constraints inherited from linear, nonlinear, directed, polyphase, and dynamic Kron-reduction theory (Schaft et al., 2024, Dorfler et al., 2011).

1. Definition and conceptual scope

Classical Kron reduction eliminates interior nodes while preserving the terminal behavior seen at boundary nodes. In linear resistive networks with Laplacian/admittance matrix $L$, partitioned into interior and boundary blocks, the reduced admittance is the Schur complement

$$L_{\mathrm{red}} = L_{BB} - L_{BC}L_{CC}^{-1}L_{CB},$$

and the mapping $z_B \mapsto J_B$ is exactly preserved (Schaft et al., 2024). This exact boundary equivalence is the benchmark against which later “optimal” formulations are measured.

Opti-KRON retains that Kron-reduction backbone but changes the design problem. Instead of taking the retained set as fixed, it chooses which buses remain as super nodes and how reduced-node injections are reassigned, subject to graph-locality and voltage-error constraints. In the 2022 formulation, “optimal” means balancing worst-case intra-cluster voltage deviation against the fraction of reduced nodes through a MILP grounded in the physics of the full network (Chevalier et al., 2022). In the 2025 extension, the same label denotes a more scalable and structure-preserving framework with a zero-injection reduction stage, tightened Big-$M$ bounds, and a radialization procedure that restores a radial reduced topology after Kron-based densification (Mokhtari et al., 20 Aug 2025).

A common ambiguity is whether Opti-KRON refers only to those MILP formulations or to a larger research program. The narrower meaning is the optimization framework introduced in 2022 and extended in 2025. A plausible broader implication is a family of optimization problems constrained by Kron-reduction realizability conditions, Laplacian structure, and boundary-behavior preservation results established across the Kron-reduction literature (Schaft et al., 2024, Sugiyama et al., 2022).

2. Mathematical foundations and feasibility conditions

The structural core of Opti-KRON is Schur-complement elimination on Laplacian-type operators. In undirected loopy Laplacian settings, Kron reduction preserves the loopy-Laplacian class, irreducibility, and effective resistance between retained nodes, while typically densifying the boundary graph and propagating self-loops from interior nodes to boundary nodes (Dorfler et al., 2011). Those properties matter operationally because any optimization over reduced networks must respect the fact that exact elimination is not arbitrary: it produces a specific class of reduced matrices with nonpositive off-diagonals, nonnegative diagonal structure, and graph-theoretic fill-in.

In directed networks, feasibility is no longer automatic. For directed weighted graphs, Schur-complement existence is characterized by a reachable-subset condition: the relevant retained set must be reachable from all eliminated nodes, and strong connectivity or quasi-strong connectivity is inherited by the reduced graph under the stated hypotheses (Wang et al., 2023). A related directed-graph formulation shows that the reduced loopy Laplacian remains a directed Laplacian-type object and that effective resistance, generalized through Markov-chain arguments, is invariant under the directed Kron reduction for loop-less graphs (Sugiyama et al., 2022). This suggests that Opti-KRON on directed networks must treat reachability and, when needed, weight balance as hard feasibility constraints rather than secondary design preferences.

For unbalanced polyphase power systems, feasibility rests on rank properties of the compound nodal admittance matrix. Under weak connectivity and strictly lossy branches, every proper diagonal subblock is full rank, so Kron reduction is feasible for any nonempty set of zero-injection nodes, and the reduced compound admittance matrix retains the same block-rank property (Kettner et al., 2017). In practice, this means that polyphase Opti-KRON cannot be reduced to single-phase intuition: eliminability is block-structured at the node level, and the admissible reduced models must preserve interphase coupling encoded in the compound admittance blocks.

These results delimit the feasible set. Opti-KRON is not merely a clustering problem on a graph; it is an optimization over realizable Schur-complement reductions and admissible approximations of them.

3. Core Opti-KRON formulation as a MILP

The 2022 Opti-KRON formulation starts from the AC network, its bus admittance $Y_b$, the bus impedance $Z_b=Y_b^{-1}$, and a library of AC load-flow voltages and currents. Binary variables select super nodes and define an aggregation matrix $A\in\{0,1\}^{n\times n}$, where $A_{i,j}=1$ means that the injection of bus $j$ is assigned to super node $i$ (Chevalier et al., 2022). The physical locality constraint uses the graph Laplacian to restrict feasible aggregations to neighbors of non-reduced nodes.

The later formulation makes the assignment constraints explicit in a nodal-admittance view. Each bus is assigned to exactly one super node,

$$L_{\mathrm{red}} = L_{BB} - L_{BC}L_{CC}^{-1}L_{CB},$$0

a reduced node cannot act as a super node,

$$L_{\mathrm{red}} = L_{BB} - L_{BC}L_{CC}^{-1}L_{CB},$$1

and assignments are restricted to adjacency,

$$L_{\mathrm{red}} = L_{BB} - L_{BC}L_{CC}^{-1}L_{CB},$$2

with $$L_{\mathrm{red}} = L_{BB} - L_{BC}L_{CC}^{-1}L_{CB},$$3 the adjacency matrix (Mokhtari et al., 20 Aug 2025). Current aggregation is

$$L_{\mathrm{red}} = L_{BB} - L_{BC}L_{CC}^{-1}L_{CB},$$4

and the full-network voltage under aggregated injections satisfies

$$L_{\mathrm{red}} = L_{BB} - L_{BC}L_{CC}^{-1}L_{CB},$$5

Voltage accuracy is measured through the per-scenario assignment error

$$L_{\mathrm{red}} = L_{BB} - L_{BC}L_{CC}^{-1}L_{CB},$$6

from which the Maximum Intra-Cluster Error is defined as

$$L_{\mathrm{red}} = L_{BB} - L_{BC}L_{CC}^{-1}L_{CB},$$7

The objective balances reduction against accuracy: $$L_{\mathrm{red}} = L_{BB} - L_{BC}L_{CC}^{-1}L_{CB},$$8 together with a hard voltage-magnitude constraint

$$L_{\mathrm{red}} = L_{BB} - L_{BC}L_{CC}^{-1}L_{CB},$$9

Because the raw problem is a MINLP, the formulation is converted to a MILP through Big-$z_B \mapsto J_B$0 linearization, rectangular decomposition of complex equations, and linearized magnitude-error inequalities (Mokhtari et al., 20 Aug 2025).

The conceptual shift is decisive. Classical Kron reduction is exact but restricted to zero-injection elimination. Opti-KRON generalizes it by permitting non-zero-injection buses to be reduced approximately through controlled reassignment of injections to neighboring super nodes. Accordingly, exact electrical equivalence is traded for explicit voltage-error control (Chevalier et al., 2022, Mokhtari et al., 20 Aug 2025).

4. Scalability and structure preservation

The 2025 extension reformulates Opti-KRON for radial distribution networks with two additions: scalability and radial structure preservation. Scalability is improved through three mechanisms: a cutting-plane restriction that limits the number of nodes reduced per MILP iteration, tightened Big-$z_B \mapsto J_B$1 bounds derived from previous-iteration MICE values, and a Stage 1 zero-injection reduction that eliminates zero-injection nodes before the main optimization stage (Mokhtari et al., 20 Aug 2025).

Stage 1 defines the zero-injection set

$z_B \mapsto J_B$2

reduces those nodes first, and uses voltage-difference tests to assign them to non-zero-injection super nodes. Stage 2 then iteratively solves the MILP with the cutting-plane bound

$z_B \mapsto J_B$3

updates adjacency, and repeats until no further reduction occurs (Mokhtari et al., 20 Aug 2025).

The structure-preserving contribution is radialization. Kron reduction of radial feeders creates cliques among neighbors of eliminated buses, so meshing appears even when the original network is radial. The radialization theorem identifies, for each maximal clique of size at least three in the reduced graph, the minimal subset of previously reduced nodes that must be restored: the degree-$z_B \mapsto J_B$4 nodes in the original spanning tree connecting the clique nodes. Reintroducing those nodes yields a radial reduced network while leaving super-node voltages unchanged (Mokhtari et al., 20 Aug 2025).

Work	Main contribution	Representative reported results
"Towards Optimal Kron-based Reduction Of Networks (Opti-KRON)" (Chevalier et al., 2022)	MILP formulation with iterative scheme and graph-locality constraints	25–85% reduction within seconds; worst-case voltage magnitude deviation within any super node cluster of less than $z_B \mapsto J_B$5 pu
"Structure-preserving Optimal Kron-based Reduction of Radial Distribution Networks" (Mokhtari et al., 20 Aug 2025)	Zero-injection stage, tightened Big-$z_B \mapsto J_B$6, cutting plane, radialization	533-bus: 85% reduction with maximum voltage error below $z_B \mapsto J_B$7 p.u.; 3499-bus: over 94% reduction with maximum voltage errors below $z_B \mapsto J_B$8 p.u.

The 2025 study also reports downstream optimization gains. On the South Alburgh feeder, a 94% reduced radialized network achieved a runtime of about $z_B \mapsto J_B$9 s for inverter-dispatch OPF versus about $M$0 s on the full network, while keeping maximum voltage error below $M$1 p.u. across 168 scenarios (Mokhtari et al., 20 Aug 2025). This does not mean that radialization is electrically exact in a stronger sense than meshed reduction; rather, it shows that sparse radial structure can be computationally preferable for downstream OPF even if it restores some previously reduced nodes.

5. Relation to nonlinear, dynamic, and generalized Kron reduction

Opti-KRON is rooted in a much wider theory of Kron reduction beyond static linear admittance matrices. Nonlinear Kron reduction models a network through strongly convex edge functionals $M$2, forms the global potential $M$3, and eliminates interior nodes by solving $M$4. The reduced functional $M$5 has a Hessian given by a Schur complement of the full Hessian, which remains a Laplacian. Under acyclic reduced graphs, or under an additional integrability condition in cyclic cases, the reduced network is again representable by edge-wise strongly convex functions of the same class (Schaft et al., 2024). The same paper states explicitly that it does not formulate optimization problems, but that its structure directly supports an Opti-KRON methodology; this suggests that realizability, monotonicity, and integrability form part of the admissible set for nonlinear optimal reduction.

Dynamic generalizations make the limitations of purely static reduction explicit. For RL networks, a provably exact time-domain Kron reduction is available without the constant $M$6 restriction by projecting branch-flow dynamics onto the nullspace induced by zero-injection constraints, yielding an exact reduced ODE with preserved input–output behavior (Singh et al., 2022). For generalized electrical networks with differential constitutive relations, exact time-domain reduction is possible under a homogeneity condition on constitutive polynomials, preserving transient as well as steady-state terminal behavior (Caliskan et al., 2012). For AC islanded microgrids with arbitrary topology, hybrid KR constructs reduced RL lines that preserve asymptotic periodic behavior exactly while remaining compatible with plug-and-play controller assumptions (Tucci et al., 2015).

A different extension addresses constant power loads. The projected-incidence formulation gives an explicit ODE reduced model for power networks with proper algebraic constraints, preserving the nonlinear load constraint as an invariant and providing a complementary approach to classical Kron reduction when standard elimination is structurally unsatisfactory (Monshizadeh et al., 2015). Taken together, these results imply that Opti-KRON is not restricted to one MILP architecture; it can be understood more generally as optimization over structurally valid reduced models in settings where exact reduction is governed by projected incidence, dynamic subspace projection, or nonlinear energy-function conditions.

6. Applications, limitations, and technical controversies

The most immediate application domain is electric-power systems, especially distribution feeders, where reduced models support AC OPF, voltage control, hosting-capacity studies, and scenario analysis (Chevalier et al., 2022, Mokhtari et al., 20 Aug 2025). Directed-graph formulations extend Kron reduction to lossless DC power-flow networks with source/sink constraints and reachability conditions (Wang et al., 2023). Polyphase theory extends feasibility guarantees to unbalanced compound admittance matrices (Kettner et al., 2017). Nonlinear resistor and memristor networks show that effective reduced constitutive laws can become nontrivial even in very small boundary reductions (Schaft et al., 2024).

Two misconceptions recur in the literature. The first is that Kron reduction is always exact. That is true for classical elimination of zero-injection nodes under the stated rank or solvability conditions, but Opti-KRON in its power-grid MILP form is explicitly approximate because it allows elimination of non-zero-injection buses through aggregation, and therefore measures voltage error rather than preserving terminal behavior identically (Chevalier et al., 2022, Mokhtari et al., 20 Aug 2025). The second is that eliminated-node dynamics can simply be ignored. For power-grid dynamics, the dynamics and noise of the Kron-reduced part can significantly affect the remaining buses, and the reduced noise becomes correlated even when the original nodal disturbances are independent; the paper addressing this point uses Mori–Zwanzig corrections to account properly for reduced-bus contributions (Pagnier et al., 2024).

Performance under contingencies adds another limitation to naive reduction. In network-reduced swing-equation models, line-fault performance measures depend on whether the faulted line connects two passive buses, two active buses, or one active to one passive bus; in all cases the measures depend quadratically on the original line power times a topology-dependent factor, often involving effective resistance and grounded passive-network terms (Coletta et al., 2017). This indicates that an Opti-KRON chosen only by voltage-clustering criteria may be suboptimal for dynamic contingency performance. A plausible implication is that multi-objective Opti-KRON should incorporate not only voltage-error metrics but also dynamic performance surrogates derived from $M$7-type measures or effective resistance.

7. Open problems and research directions

Several open directions are stated explicitly across the cited works. For nonlinear reduction, the main open questions are explicit criteria for global solvability of the interior equations, checkable conditions for the cyclic-case integrability constraint, computational methods for the mapping $M$8, integration with model-order reduction, data-driven identification of effective nonlinear relations, optimization-based boundary selection, and the synthesis problem for networks realizing prescribed boundary behavior (Schaft et al., 2024). These are natural components of a generalized Opti-KRON agenda.

For the power-grid MILP line, the 2022 work identifies comparison with conventional reductions, global-optimality analysis for radial networks, robust OPF on reduced models, and disaggregation of reduced-network solutions back to full-network controls as future work (Chevalier et al., 2022). The 2025 extension adds three-phase unbalanced feeders, lifting of OPF and DER-control solutions from reduced to full networks, and broader use in planning and real-time operation (Mokhtari et al., 20 Aug 2025). Directed-graph effective-resistance theory further suggests optimization over boundary-node selection using resistance or hitting-probability metrics, since those quantities are Kron-invariant under the directed reduction and induce graph metrics in weight-balanced cases (Sugiyama et al., 2022).

The overall trajectory is therefore clear. Opti-KRON began as an optimization-based way to choose reductions in AC power networks, but the surrounding Kron-reduction literature shows that its mature form is likely to be a constrained optimal-reduction theory in which Schur-complement structure, realizability, graph reachability, passivity, and dynamic fidelity are all first-class design variables rather than afterthoughts.