Local Optimization Quantities

• Local Optimization Quantities are defined metrics that capture decay-of-influence, local curvature, basin topology, and error rates across diverse optimization problems.
• They enable scalable distributed algorithms by measuring how local perturbations impact global performance in networked, dynamical, and quantum systems.
• This framework informs algorithm design and convergence assessment by linking geometry, local hardness, and Pareto optimality to practical computational efficiency.

Local Optimization Quantities

Local optimization quantities provide a rigorous framework for understanding and quantifying the behavior and performance of local versus global solution strategies across convex, nonconvex, discrete, and quantum optimization problems. These quantities encompass decay-of-influence metrics, measures of local curvature and conditioning, basin topography, locality rates, error exponents, and specialized notions in fields such as dynamical mean field theory and quantum circuit optimization. They form the foundation for both theoretical analyses and practical algorithms that exploit locality to balance accuracy, communication, and computational effort.

1. Graph-Based Locality and Decay-of-Influence Metrics

A core advance in multi-agent and networked optimization is the quantification of locality via graph-based metrics. In problems of the form $\min_{x\in\mathbb{R}^n} f(x)=\sum_{i=1}^n f_i(x_i)$ subject to $A x = b$ with $A$ full-row rank, each variable $x_i$ is typically coupled to a limited local neighborhood of decision variables. Three undirected graphs are defined:

• Decision-variable graph: $G_{dec}=(V^{(p)},E_{dec})$, with $(i,j)\in E_{dec}$ iff variables $x_i$, $x_j$ appear together in some constraint.
• Constraint graph: $G_{con}=(V^{(d)},E_{con})$, with $(k,\ell)\in E_{con}$ iff constraints $A x = b$0, $A x = b$1 are coupled via $A x = b$2.
• Coupling graph: $A x = b$3, with $A x = b$4 iff $A x = b$5.

The k-hop neighborhood $A x = b$6 denotes all nodes within graph distance $A x = b$7 of node $A x = b$8. Critically, the influence of variables or data outside a k-hop neighborhood decays exponentially, enabling truncated local subproblem solutions to accurately approximate their respective components of the global optimum.

The locality rate is defined as

$A x = b$9

where $A$0. Smaller $A$1 indicates more rapid decay with neighborhood radius, with error bounded as

$A$2

The necessary neighborhood radius to reach error $A$3 is logarithmic: $A$4

These theoretical results translate directly into scalable distributed optimization algorithms with communication and computation that depend on locality rather than total system size (Brown et al., 2020).

2. Sensitivity, Correlation Decay, and Bias-Variance in Network Optimization

Local optimization quantities also arise through the sensitivity of optimal points to constraint perturbations. For min-cost network flow with node-imbalance $A$5 and strongly convex separable costs, the sensitivity matrix is given by

$A$6

where $A$7.

The entry $A$8 quantifies the local correlation between variable $A$9 and constraint $x_i$0. On graphs with sufficient expansion, such local correlations decay exponentially with graph-theoretical distance. For a finite-support perturbation, the gradient of the energy at a given edge with respect to a distant node's data decays as

$x_i$1

with $x_i$2 the minimal graph distance (Rebeschini et al., 2015).

Localized reoptimization algorithms exploit this decay, solving restricted subproblems and incurring a trade-off between computational bias (due to locality) and variance (algorithmic convergence). Bounds on the bias and variance as functions of subgraph size and distance quantify the efficacy and limits of localized procedures.

3. Local Subgradient Variation: Finer-Scale Complexity Measures

Recent work introduces two local subgradient-variation quantities for nonsmooth optimization:

• Bounded Maximum Local Variation (Grad-BMV):

$x_i$3

for all $x_i$4, all unit $x_i$5, and fixed radius $x_i$6.

• Bounded Mean Oscillation of the Subgradient (Grad-BMO):

$x_i$7

for $x_i$8.

Here, $x_i$9 and $G_{dec}=(V^{(p)},E_{dec})$0 replace global Lipschitz and smoothness constants in complexity bounds, enabling finer oracle complexity results that interpolate between worst-case and benign cases. The mean width of the local subdifferential

$G_{dec}=(V^{(p)},E_{dec})$1

further refines complexity in piecewise linear or polyhedral problems (Diakonikolas et al., 2024).

Deterministic and randomized algorithms admit convergence rates with dependence on these local variational quantities and mean width, leading to improved depth vs. error trade-offs, especially for parallel optimization algorithms.

4. Topographic Quantities and Basin Structure in High Dimensions

Characterizing the geometrical complexity of high-dimensional objective functions is achieved via several local-optimization quantities (Deng et al., 2014):

• Number of local minima ($G_{dec}=(V^{(p)},E_{dec})$2)
• Basin-of-attraction volumes ($G_{dec}=(V^{(p)},E_{dec})$3) and their probabilities ($G_{dec}=(V^{(p)},E_{dec})$4)
• Depth-weighted probabilities ($G_{dec}=(V^{(p)},E_{dec})$5), defined to down-weight shallow minima
• Landscape entropy ($G_{dec}=(V^{(p)},E_{dec})$6), with

$G_{dec}=(V^{(p)},E_{dec})$7

serving as a complexity index that combines basin widths and depths

• Curvature metrics (Hessian condition number at minimizers $G_{dec}=(V^{(p)},E_{dec})$8)

These quantities are estimated by random hill climbing (randomized local descent), cluster analysis of discovered minima, and statistical error bounds on the frequency of convergence to distinct basins. The topographic fingerprint guides algorithm selection, smoothing strategies, and assessment of optimization hardness.

5. Local Hardness and Error Exponents in Disordered Systems

In complex disordered systems such as spin glasses, the local hardness metric quantifies the predictive error of locally optimized ground states relative to the global ground state. For a local subsystem of size $G_{dec}=(V^{(p)},E_{dec})$9,

$(i,j)\in E_{dec}$0

where $(i,j)\in E_{dec}$1 are spins in the subsystem ground state and $(i,j)\in E_{dec}$2 in the global ground state.

The average error $(i,j)\in E_{dec}$3 typically obeys a power-law scaling: $(i,j)\in E_{dec}$4 with local hardness exponent $(i,j)\in E_{dec}$5 extracted from log–log scaling fits. This exponent quantifies the rate at which local solutions converge to the global as locality increases (Shen et al., 5 May 2025).

Near critical points, gapless avalanche-like excitations dominate error and exhibit scale-invariant size distributions: $(i,j)\in E_{dec}$6 where $(i,j)\in E_{dec}$7 is avalanche size and $(i,j)\in E_{dec}$8 the avalanche exponent.

Correlation of critical thresholds decays algebraically with distance, establishing the range and speed of local information propagation.

6. Radius of Local Efficiency and Local Pareto Optimality

In multiobjective and vector optimization, the largest ball around a candidate solution where the point remains (locally) Pareto optimal is formalized as the radius of efficiency $(i,j)\in E_{dec}$9. For vector quadratic-fractional programs,

$x_i$0

where $x_i$1 encodes intervals along direction $x_i$2 where a trial point dominates the candidate. First-order Pareto (KKT-type) conditions are used to check the presence or absence of locally improving directions. If $x_i$3, the solution is globally Pareto optimal (Oliveira et al., 2013).

Explicit computation of $x_i$4 informs both stopping criteria and Pareto globality certification, with extensions to local quadratic approximations in non-quadratic settings.

7. Local Optimality in Quantum Circuit Compilation

In quantum circuit optimization, local optimality is realized through segment-optimality and compactness with respect to a cost function (e.g., gate count). A circuit is locally optimal if, for a fixed segment size $x_i$5, no subcircuit of $x_i$6 layers can be improved by a prescribed oracle. The cut-and-meld algorithm recursively decomposes, optimizes, and merges subcircuits, guaranteeing local optimality and tight packing (compactness) of gates.

Complexity is controlled by the linear scaling of oracle calls relative to circuit size, and the improvement per iteration, establishing efficient, high-fidelity local optimization for large-scale circuits (Arora et al., 26 Feb 2025).

---

These local optimization quantities provide mathematically precise, problem-specific tools to analyze, predict, and systematically exploit the structure of high-dimensional or networked optimization problems in diverse domains. They underpin advances in distributed algorithms, benchmarking of landscape complexity, and fundamentally limit or unlock parallelism and communication efficiency across convex, nonconvex, discrete, and quantum computational contexts.