Reverse Kron Reduction

• Reverse Kron reduction is a technique to reconstruct eliminated network dynamics and admittance structures from reduced models using Schur complements and memory kernels.
• It employs methods like Volterra convolution and numerical matrix exponentials to recover noise effects and initial conditions in dynamical systems.
• Applications include power system state estimation and multi-phase radial network reconstruction, ensuring robust uncertainty quantification and precise admittance recovery.

Reverse Kron reduction provides a systematic methodology for reconstructing network dynamics or admittance structures that were “eliminated” through standard Kron reduction, a process traditionally used to reduce complex systems—such as power grids or network-coupled oscillators—to a lower-dimensional subspace by “marginalizing out” certain fast or unobservable nodes. The reverse procedure has dual importance: (i) for dynamical systems, it enables the recovery of the original dynamics of eliminated nodes from the reduced model, accounting for all memory and noise transfer effects; (ii) for static network problems, it allows for the exact reconstruction of the full admittance or Laplacian matrix from only partial (Kron-reduced) observations, under structural assumptions. These techniques are foundational in power system identification, reduced modeling of stochastic networks, and admittance matrix estimation in radial electrical networks (Pagnier et al., 2024, Low, 2024).

1. Fundamentals of Kron Reduction and its Inversion

Standard Kron reduction applies to a network encoded by a symmetric or Hermitian Laplacian or admittance matrix partitioned as

$$L = \begin{pmatrix} L_{AA} & L_{AB}\ L_{BA} & L_{BB} \end{pmatrix}$$

where nodes in $A$ are retained (“kept”) and $B$ are eliminated. The reduced matrix is the Schur complement

$$L_{\mathrm{red}} = L_{AA} - L_{AB} L_{BB}^{-1} L_{BA}$$

If $L_{BB}$ (or $Y_{ii}$ for admittance matrices) is invertible—for example, if the eliminated subnetwork is connected and has positive definite properties—the Kron reduction faithfully captures static properties on $A$ such as voltages or power flows. In dynamic settings (e.g., swing equations linearized near stable operation), the effect of elimination extends to the dynamical behavior of the remaining nodes, motivating careful treatment of the inversion process (Pagnier et al., 2024, Low, 2024).

Reverse Kron reduction comprises determining how the eliminated subnet's dynamics or structure can be exactly reconstructed from the observed reduced model, the initial conditions, and possibly the full noise history.

2. Reverse Kron Reduction for Linear Network Dynamics

In dynamical settings described by equations of the form

$\dot{z} = Lz + \eta(t)$

with $z = (x_A, x_B)^T$ and nodal excitations $\eta(t)$, direct Kron reduction is insufficient to characterize the evolution of $A$0-type nodes after their removal. Using the Mori–Zwanzig projection formalism, the influence of the eliminated nodes manifests as a memory kernel and colored noise in the effective dynamics on $A$1:

$A$2

where

• $A$3 is the memory kernel,
• $A$4 encodes dependence on initial conditions,
• $A$5 is direct noise driving $A$6 nodes.

The process of reverse Kron reduction is then implemented by reconstructing $A$7 using

$A$8

This exact operator is a Volterra convolution of the observed $A$9-node dynamics and the noise, with the matrix exponential kernel $B$0 (Pagnier et al., 2024).

Necessary conditions for this inversion are: $B$1 must be Hurwitz (all eigenvalues with strictly negative real part to ensure decay and integrability of the memory kernel), $B$2 invertible, and initial conditions $B$3 known or sampled appropriately.

3. Reverse Kron Reduction of Multi-phase Radial Admittance Matrices

For static network identification problems in radial three-phase systems, the Kron reduction formalism allows recovery of the full admittance matrix $B$4 from reduced data. Here, $B$5 is partitioned,

$B$6

with $B$7 the “measured” (boundary) and $B$8 the “hidden” (interior) buses. Kron reduction produces

$B$9

The reverse procedure is detailed for three-phase radial networks, where the reversible structure arises from sequential elimination of each hidden node and preservation of an "arrow" matrix structure. Each elimination creates a new clique among the neighbors of the eliminated node and updates only the clique’s block entries in the matrix; elsewhere the blocks remain unchanged.

Reverse Kron reduction proceeds iteratively:

• At each step, permute the matrix to bring the current clique to the bottom-right,
• Write the Schur complement equation for that block,
• Recover the eliminated node’s parameters by solving a $$L_{\mathrm{red}} = L_{AA} - L_{AB} L_{BB}^{-1} L_{BA}$$0 block system,
• Update the clique and repeat until the full admittance is reconstructed.

This structure guarantees invertibility for radial networks where every hidden node has degree $$L_{\mathrm{red}} = L_{AA} - L_{AB} L_{BB}^{-1} L_{BA}$$1, all diagonal blocks are nonsingular ($$L_{\mathrm{red}} = L_{AA} - L_{AB} L_{BB}^{-1} L_{BA}$$2 for admittance), and all $$L_{\mathrm{red}} = L_{AA} - L_{AB} L_{BB}^{-1} L_{BA}$$3 intermediate matrices remain full rank (Low, 2024).

4. Algorithmic Formulation and Computational Considerations

A general step-by-step recipe for reverse Kron reduction in dynamical systems involves:

1. Identifying kept (A) and eliminated (B) node sets and partitioning the system matrix.
2. Verifying invertibility and the Hurwitz property of the eliminated block.
3. Computing the reduced Laplacian or admittance.
4. Forward simulation on the reduced model with noise corrections.
5. Backward reconstruction of eliminated node states via the Volterra convolution integrals.
6. Utilizing numerical techniques for matrix exponentials (e.g., diagonalization, Krylov or Padé approximations), quadrature for convolutions, and consistent discretization of stochastic convolutions (such as Euler–Maruyama for stochastic differential equations) (Pagnier et al., 2024).

For static network reconstruction, the algorithm consists of:

• Iterative elimination (“Schur peeling”) and identification of sibling groups,
• Solving, at each reverse step, clique-local Schur complement equations of the form (3.12) for unknown block elements,
• Reassembling the original matrix with only $$L_{\mathrm{red}} = L_{AA} - L_{AB} L_{BB}^{-1} L_{BA}$$4 block inversions required per step,
• Maintaining orderings and permutations to keep clique structure,
• Final permutation to yield the original node ordering (Low, 2024).

The computational cost is dominated by $$L_{\mathrm{red}} = L_{AA} - L_{AB} L_{BB}^{-1} L_{BA}$$5 operations (with $$L_{\mathrm{red}} = L_{AA} - L_{AB} L_{BB}^{-1} L_{BA}$$6 eliminated nodes, $$L_{\mathrm{red}} = L_{AA} - L_{AB} L_{BB}^{-1} L_{BA}$$7 maximal clique size), with each step involving local operations on $$L_{\mathrm{red}} = L_{AA} - L_{AB} L_{BB}^{-1} L_{BA}$$8 blocks. Numerical stability is secured as long as all $$L_{\mathrm{red}} = L_{AA} - L_{AB} L_{BB}^{-1} L_{BA}$$9 blocks are well-conditioned, which is typical in three-wire positive-definite lines; small shunt regularizations can be added if required.

5. Structural and Statistical Constraints

The exact invertibility of Kron reduction requires:

• In the static setting, radial topology; otherwise, cycles create non-invertible Schur sequences.
• All eliminated and intermediate diagonal blocks to be invertible and of full rank (physically, positive-definite admittances).
• Elimination sequences chosen such that sibling groups and parental nodes can be identified at each reverse step.

Stochastic modeling introduces correlated effective noise in the reduced system: whereas the original noise $L_{BB}$0 may be uncorrelated across $L_{BB}$1 and $L_{BB}$2, after reduction the noise vector on $L_{BB}$3 is temporally and spatially correlated,

$L_{BB}$4

so that the common white/noise assumptions may be violated, affecting the validity of naive reduced models (Pagnier et al., 2024). This has significant implications for rigorous uncertainty quantification in power grid applications.

6. Illustrative Example and Applications

A worked case is provided for a three-phase, five-node radial network, with three measured leaf nodes and two hidden nodes in series. The procedure successfully reconstructs the original $L_{BB}$5 admittance matrix through the sequential recovery of hidden bus admittance and line admittance parameters, each step governed by explicit $L_{BB}$6 algebraic relations. All intermediate inversions remain $L_{BB}$7 matrices ($L_{BB}$8 floating point operations each), and the process admits practical near-linear runtime for large sparse systems (Low, 2024).

Applications include:

• Full admittance recovery in distribution networks for state estimation and system identification from partial measurements.
• Reconstructing fast node dynamics in reduced-order power grid simulations to assess the impact of load fluctuations on generator dynamics, especially when spatio-temporal noise effects are nontrivial (Pagnier et al., 2024).
• Quantifying the information loss and noise transformation that occur during model reduction.

7. Significance and Research Directions

Reverse Kron reduction bridges model reduction and system identification, providing mathematically exact procedures to “lift” reduced models—static or dynamic—back to full-network representations with all noise and memory effects correctly incorporated. Its significance lies in rigorous network reconstruction (particularly for radial and tree-like topologies), quantification of reduction-induced errors, and robust uncertainty propagation in power system dynamics.

A plausible implication is that incorporating reverse Kron correction terms is essential for high-fidelity modeling of power grids and oscillator networks under stochastic disturbances, especially in future distribution grids with rich multi-phase and unbalanced structures. Open challenges include generalization to meshed (non-radial) topologies, efficient algorithms for networks with large numbers of hidden nodes, and robust handling of noisy and incomplete measurement data (Pagnier et al., 2024, Low, 2024).