An Exact, Finite Dimensional Representation for Full-Block, Circle Criterion Multipliers

Basic Information

• Paper ID: 2511.20995
• Title: An Exact, Finite Dimensional Representation for Full-Block, Circle Criterion Multipliers
• Authors: Felix Biertümpfel (University of Michigan & TU Dresden), Bin Hu (University of Illinois at Urbana-Champaign), Geir Dullerud (University of Minnesota), Peter Seiler (University of Michigan)
• Classification: eess.SY (Systems and Control), cs.SY
• Submission Date: November 26, 2025 (arXiv submission)
• Paper Link: https://arxiv.org/abs/2511.20995

Abstract

This paper provides, for the first time, a finite-dimensional characterization of the complete set of full-block circle criterion multipliers. The study considers discrete-time linear time-invariant systems interconnected with non-repeated sector-bounded nonlinear feedback. Sufficient conditions for stability and performance can be derived through: (i) dissipativity inequalities and (ii) quadratic constraints (QCs) on constrained nonlinear input/output pairs. Larger classes of QCs (or multipliers) can reduce the conservatism of the conditions. Full-block circle criterion multipliers define the complete set of all possible QCs for non-repeated sector-bounded nonlinearities, providing the least conservative conditions. However, full-block multipliers are defined by uncountably infinite constraints, making the original formulation computationally infeasible. The key theoretical insight of this paper is: the set of all input/output pairs for non-repeated sector-bounded nonlinearities equals the set of all incremental pairs of appropriately constructed piecewise linear functions. The new characterization requires only a finite number of matrix copositive constraints, enabling exact computational implementation for problems with nonlinear input/output dimension ≤ 4.

Research Background and Motivation

Problem Definition

This paper studies systems formed by the interconnection of a known discrete-time linear time-invariant (LTI) system with static memoryless nonlinear feedback, with the goal of analyzing the stability and induced gain of the interconnected system.

Problem Significance

1. Fundamental Robust Control Problem: Sector-bounded nonlinearities are ubiquitous in practical control systems (such as saturation, dead-zone, ReLU activation functions, etc.), and their stability analysis is a core problem in control theory
2. Neural Network Verification: For small-scale recurrent neural networks (e.g., using ReLU activation functions), this method can be applied to incremental stability analysis
3. Theoretical Completeness: Providing non-conservative analysis tools is essential for understanding system performance limits

Limitations of Existing Methods

1. Diagonal Multipliers (M_d): Computationally simple but overly conservative, unable to fully exploit the structural properties of nonlinearities
2. Convex Relaxation (M_c): Improves conservatism by imposing constraints at vertices of the sector hypercube, but remains incomplete
3. Full-Block Multipliers (M_fb): Theoretically provide a complete set of QCs, but are defined by uncountably infinite constraints (requiring one constraint for each point in the sector α,β^m), making them computationally completely infeasible
4. Existing Approximation Methods: Based on generalizations of Polya's theorem or finite grid approximations, but cannot guarantee exactness

Research Motivation

A significant gap exists between theoretical optimality and computational practice: the theoretically optimal full-block multipliers are computationally infeasible. This paper aims to find an equivalent but finite-dimensional characterization that makes the theoretically optimal solution computationally tractable.

Core Contributions

1. Theoretical Breakthrough: Proves that the set of all input/output pairs for non-repeated sector-bounded nonlinearities equals the set of all incremental pairs of specific piecewise linear functions (Lemma 5)
2. Finite-Dimensional Characterization: Provides a new finite-dimensional characterization of full-block circle criterion multipliers (Corollary 1), transforming the infinite-dimensional constraint problem into 4^m copositive constraints
3. Exact Computational Method: For m ≤ 4 dimensional problems, leverages known exact copositive relaxations to achieve exact computational implementation of the complete QC set
4. Practical Verification: Demonstrates the advantages of the new method over existing approaches through numerical examples, verifying larger stability margins and tighter performance bounds

Detailed Methodology

Task Definition

System Setup: Consider the feedback interconnection Fu(G,ΔΦ) shown in Figure 1:

• Nominal System G: Discrete-time LTI system

  x(k+1) = Ax(k) + B₁w(k) + B₂u(k)
  v(k) = C₁x(k) + D₁₁w(k) + D₁₂u(k)
  y(k) = C₂x(k) + D₂₁w(k) + D₂₂u(k)
  

• Perturbation ΔΦ: Defined by non-repeated scalar nonlinearities, w(k) = Φ(v(k)), where Φ(v) = φ₁(v₁),...,φₘ(vₘ)ᵀ

Sector Boundedness Definition: A scalar function φ is in sector α,β if:

(φ(x) - αx)(βx - φ(x)) ≥ 0, ∀x ∈ ℝ

Analysis Objectives:

1. Well-posedness
2. Internal stability
3. Finite induced ℓ₂ gain: ‖Fu(G,ΔΦ)‖₂→₂ < γ

Core Theoretical Framework

1. Quadratic Constraints (QC) Fundamentals

Definition: A set G ⊂ ℝ^(2m) satisfies the QC defined by M ∈ S^(2m) if:

zᵀMz ≥ 0, ∀z ∈ G

Application to Nonlinearities: For the graph of Φ, G(Φ) = {v; w : v ∈ ℝᵐ, w = Φ(v)}, if:

[v; w]ᵀM[v; w] ≥ 0, ∀v ∈ ℝᵐ, w = Φ(v)

then M defines a valid QC for Φ.

Complete Set:

G(sec[α,β]ᵐ) := ⋃_{Φ∈sec[α,β]ᵐ} G(Φ)

2. Stability Condition (Theorem 1)

Given assumptions:

1. D₁₁ = 0 (ensures well-posedness)
2. G(secα,βᵐ) ⊂ QC(M) (M is a valid multiplier)
3. ∃P ≽ 0, γ > 0 such that L(P,M,γ²) ≺ 0

where L is an affine function involving system matrices, Lyapunov matrix P, and multiplier M.

Conclusion: The system Fu(G,ΔΦ) is well-posed, internally stable, and ‖Fu(G,ΔΦ)‖₂→₂ < γ.

Key Theoretical Innovation

Piecewise Linear Function Construction

Define a special piecewise linear function:

f_{αβ}(x) = {αx,  x ≤ 0
            {βx,  x > 0

Corresponding repeated function: Fαβ(v) = f_{αβ}(v₁),...,f_{αβ}(vₘ)ᵀ

Incremental Graph:

I(Fαβ) := {[v̄-v̂; w̄-ŵ] : ∃v̄,v̂ ∈ ℝᵐ s.t. w̄=Fαβ(v̄), ŵ=Fαβ(v̂)}

Core Equivalence (Lemma 5)

Key Result: I(Fαβ) = G(secα,βᵐ)

Proof Sketch:

1. I(Fαβ) ⊆ G(secα,βᵐ): For any incremental pair (dv,dw), one can construct a diagonal matrix Γ such that dw_i = γ_i·dv_i, where γ_i ∈ α,β
2. G(secα,βᵐ) ⊆ I(Fαβ): For any input/output pair (v,w), define:

   v̄ᵢ = (γᵢ-α)/(β-α)·vᵢ, v̂ᵢ = (γᵢ-β)/(β-α)·vᵢ
   

   
   where γᵢ = wᵢ/vᵢ (or (α+β)/2 if vᵢ=0), one can verify that v̄-v̂=v and w̄-ŵ=w

Geometric Intuition (Figure 2): Any two-point incremental pair of the piecewise linear function can realize any slope within α,β, thereby covering all input/output pairs of sector-bounded nonlinearities.

New Finite-Dimensional Characterization

Incremental QC Set (Theorem 2)

Define the function:

g_M(Γ̄,Γ̂) = [Γ̄  -Γ̂      ]ᵀ   [Γ̄  -Γ̂      ]
             [rΓ̄+cI -rΓ̂-cI]  M [rΓ̄+cI -rΓ̂-cI]

where c = (α+β)/2, r = (β-α)/2

New Multiplier Set:

M_inc := {M ∈ S^(2m) : g_M(Γ̄,Γ̂) ∈ COP^(2m), 
          ∀Γ̄,Γ̂ ∈ diag({-1,1}ᵐ)}

Theorem: I(Fαβ) ⊂ QC(M) ⟺ M ∈ M_inc

Proof Key Points:

• Utilize the decomposition of |v̄|,|v̂| ∈ ℝ^m_≥0 and sign matrices Γ̄,Γ̂
• Transform incremental QC conditions into copositive conditions
• Copositivity ensures non-negativity for all non-negative vectors

Equivalence Result (Corollary 1)

Core Result: M_inc = M_fb

Proof Chain:

1. By Theorem 2: M ∈ M_inc ⟺ I(Fαβ) ⊂ QC(M)
2. By Lemma 5: I(Fαβ) = G(secα,βᵐ)
3. By Lemma 2: G(secα,βᵐ) ⊂ QC(M) ⟺ M ∈ M_fb

Computational Properties

Constraint Count: M_inc requires 4^m = 2^(2m) copositive constraints (one for each pair Γ̄,Γ̂ ∈ diag({-1,1}ᵐ))

Computational Complexity:

• Copositivity verification is a co-NP-complete problem
• Exact Relaxation: For m ≤ 4, any copositive matrix is completely positive (sum of positive semidefinite and non-negative matrices), making the relaxation exact
• For m ≥ 5, non-completely positive copositive matrices exist (e.g., Horn matrices), requiring conservative approximations

Set Inclusion Relationship:

M_d ⊂ M_c ⊂ M_fb = M_inc

Experimental Setup

System Configuration

• System Dimension: 3rd-order discrete-time LTI system
• Nonlinearity Dimension: m = 3 (three independent sector-bounded nonlinearities)
• Sector Bounds: 0,β, with β varying in 0,1.3
• Application Background: Similar to incremental stability analysis of ReLU activation functions in small-scale recurrent neural networks

State-Space Matrices

The specific matrices (A,B₁,B₂,C₁,C₂,D₁₁,D₁₂,D₂₁,D₂₂) of system G are provided in the paper, noting that D₁₁≠0 (ensuring system well-posedness).

Comparison Methods

1. M_d: Diagonal multipliers (most conservative but fastest to compute)
2. M_c: Convex relaxation (moderate conservatism)
3. M_inc: Newly proposed complete set (least conservative)

Implementation Details

• Solver: MATLAB + CVX + MOSEK
• Hardware: 4.51GHz processor, 24GB memory
• Copositivity Implementation: For m=3, use exact relaxation (positive semidefinite + non-negative decomposition)

Evaluation Metrics

• Gain Upper Bound: γ such that ‖Fu(G,ΔΦ)‖₂→₂ < γ
• Stability Margin: Maximum sector bound β_max for which stability can be verified
• Computation Time: Time required to solve the SDP

Experimental Results

Main Results (β=1 case)

Key Findings:

1. Conservatism: M_inc reduces conservatism by 12.87% compared to M_c, and 47.3% compared to M_d
2. Computational Cost: M_inc computation time is approximately 2× that of M_c, but still within acceptable range (<0.5s)
3. Exactness: M_inc provides theoretically optimal bounds (for static QCs)

Sector Bound Variation Analysis (Figure 3)

Testing 15 equally-spaced points for β ∈ 0,1.3:

Convergence Behavior (β→0):

• All methods converge to nominal gain γ_nom = 1.396
• Validates correctness of the methods

Conservatism Gap Evolution:

• β ∈ 0,0.6: Small differences among three methods
• β > 0.6: Gap between M_d and full-block multipliers increases significantly
• β > 0.9: Gap between M_c and M_inc becomes pronounced

Stability Margin:

Statistical Significance: M_inc provides the tightest bounds at all test points, with the gap being particularly significant at large sector bounds.

Experimental Findings

1. Dimension Sensitivity: For m=3 problems, exact copositive relaxation works well
2. Computational Feasibility: Even for the complete set M_inc, computation time remains practical
3. Conservatism Improvement: Improvements are most significant in critical regions (near stability boundary)
4. Theoretical Verification: Experimental results align with theoretical predictions (M_inc = M_fb provides optimal bounds)

Related Work

Quadratic Constraints and Multiplier Theory

1. IQC Framework: Megretski & Rantzer (2002) established foundational theory of integral quadratic constraints
2. Full-Block Multipliers: Veenman & Scherer (2014-2017) systematically studied properties and applications of full-block multipliers
3. Discrete-Time Systems: Fetzer & Scherer (2017) specifically addressed absolute stability analysis for discrete-time feedback interconnections

Computational Methods

1. Polya Theorem Generalization: Fetzer & Scherer (2017) used Polya's theorem to provide approximations of M_fb
2. Finite Grid: Approximate M_fb by imposing constraints at finite grid points on α,β^m
3. Convex Relaxation: M_c imposes constraints at hypercube vertices (comparison baseline in this paper)

Advantages of This Work

1. Theoretical Completeness: First exact finite-dimensional characterization of M_fb
2. Computational Feasibility: Enables exact computation for m ≤ 4 (previous methods either lack exactness or are computationally infeasible)
3. Unified Framework: Establishes deep connection between sector nonlinearities and piecewise linear functions through incremental graphs

Related but Distinct Research Directions

1. Repeated Nonlinearities: Lemma 4 proves repeated and non-repeated nonlinearities have the same complete QC set
2. ReLU Network Analysis: Noori et al. (2024) studied complete QC sets for repeated ReLU
3. Regional Analysis: Valmorbida et al. (2018) studied regional analysis for slope-restricted Lurie systems

Conclusions and Discussion

Main Conclusions

1. Theoretical Contribution: Proves that the set of all input/output pairs for non-repeated sector-bounded nonlinearities equals the set of all incremental pairs of specific piecewise linear functions
2. Computational Contribution: Transforms the infinite-dimensional characterization of full-block multipliers into a finite-dimensional one with 4^m copositive constraints
3. Practical Contribution: Provides exact computationally feasible implementation for m ≤ 4 problems
4. Verification Conclusion: Numerical experiments confirm significant advantages of the new method in conservatism and stability margins

Limitations

1. Dimension Restrictions:
   • Exact relaxation only valid for m ≤ 4
   • For m ≥ 5, conservative copositive approximations needed (e.g., positive semidefinite + non-negative decomposition)
   • Constraint count grows exponentially with m (4^m)
2. Computational Complexity:
   • Copositivity verification is inherently co-NP-complete
   • For large-scale problems (m > 4), computational burden may be prohibitive
3. Assumption Conditions:
   • Requires D₁₁ = 0 for well-posedness (can be relaxed under additional assumptions)
   • Applies only to static memoryless nonlinearities
   • Limited to sector-bounded type nonlinearities
4. Conservatism Sources:
   • Uses static QCs (dynamic multipliers may be less conservative)
   • Lyapunov function is quadratic (non-quadratic Lyapunov functions may be superior)

Future Directions

1. High-Dimensional Extensions:
   • Investigate tighter copositive approximations for m > 4
   • Explore structured copositive conditions to reduce computational burden
   • Develop scalable numerical algorithms
2. Dynamic Multipliers:
   • Extend finite-dimensional characterization to dynamic multipliers
   • Investigate exploitation of frequency-domain properties
3. Other Nonlinearity Types:
   • Extend to slope-restricted, monotone, odd-symmetric, and other types
   • Study mixed-constraint nonlinearities
4. Application Extensions:
   • Neural network verification and certification
   • Adaptive control system analysis
   • Networked control systems
5. Theoretical Deepening:
   • Study fundamental differences between repeated and non-repeated nonlinearities
   • Explore applicability of incremental graph methods to other problems

In-Depth Evaluation

Strengths

1. Theoretical Innovation (★★★★★):
   • First solution to the long-standing open problem of finite-dimensional characterization of full-block multipliers
   • Incremental graph equivalence (Lemma 5) provides profound theoretical insight
   • Rigorous and complete proofs with clear logic
2. Practical Value (★★★★☆):
   • Provides exact computationally feasible method for m ≤ 4, covering many practical applications
   • Computation time within acceptable range (<0.5s)
   • Good numerical stability (based on mature SDP solvers)
3. Method Elegance (★★★★★):
   • Clever transformation through construction of special piecewise linear functions
   • Simplifies proofs using sign matrices and absolute value decomposition
   • Unifies perspectives on sector nonlinearities and incremental analysis
4. Experimental Sufficiency (★★★☆☆):
   • Clearly demonstrates advantages over existing methods
   • Covers critical parameter ranges
   • Limited to single numerical example, lacking diversity
5. Writing Quality (★★★★★):
   • Clear structure and rigorous logic
   • Standardized and consistent notation
   • Figure 2 provides excellent geometric intuition

Weaknesses

1. Dimension Scalability:
   • Inability to implement exactly for m > 4 is a serious limitation
   • Limited applicability to modern neural networks (typically m ≫ 4)
   • Exponential constraint count (4^m) is fundamental bottleneck
2. Experimental Limitations:
   • Only one 3-dimensional example, lacking 4-dimensional or higher-dimensional experiments
   • No testing of copositive relaxation conservatism at m = 5
   • Lacks detailed comparison with Polya theorem-based methods
3. Insufficient Computational Details:
   • No discussion of numerical stability of copositive relaxation
   • Missing convergence and precision analysis of SDP solver
   • Lacks computational complexity analysis for large-scale problems
4. Application Scope:
   • Limited to discrete-time systems (continuous-time requires separate treatment)
   • Static nonlinearity assumption excludes dynamic uncertainty
   • Sector-bound assumption may not hold in some applications
5. Theoretical Depth:
   • Unexplored why m ≤ 4 is critical dimension for exact relaxation
   • Insufficient discussion of broader applicability of incremental graph method
   • Relationship with dynamic multipliers not deeply analyzed

Impact Assessment

Short-term Impact (1-2 years):

• Provides optimal analysis tool for small-scale control systems
• Promotes adoption of more precise methods in neural network verification
• Stimulates further application of copositive theory in control

Medium-term Impact (3-5 years):

• Likely becomes standard content in robust control textbooks
• Inspires new approximation methods for high-dimensional problems
• Promotes application of incremental analysis paradigm to other problems

Long-term Impact (5+ years):

• If high-dimensional extensions succeed, may fundamentally change nonlinear system analysis
• Theoretical framework may extend to broader constraint types
• Becomes classical work at intersection of control theory and optimization

Potential Citation Scenarios:

1. Research on complete multiplier sets for other nonlinearity types
2. Development of new copositive approximation algorithms
3. Neural network robustness certification
4. Stability analysis of adaptive and learning control

Applicable Scenarios

Most Suitable Applications:

1. Low-Dimensional Control Systems (m ≤ 4):
   • Attitude control in aerospace (typically 3-dimensional)
   • Robot joint control (single joint or small robots)
   • Power electronic converters (few switching elements)
2. Small Neural Network Verification:
   • Small RNNs in embedded controllers
   • Simple neural networks in safety-critical systems
   • Incremental stability and robustness certification
3. Theoretical Research:
   • Benchmark for testing conservatism of new algorithms
   • Fundamental research on multiplier theory
   • Teaching and demonstrating optimal analysis methods

Less Suitable Scenarios:

1. High-Dimensional Systems (m > 4): Requires approximation methods
2. Large-Scale Neural Networks: Computationally infeasible
3. Real-Time Applications: Computation time may be excessive
4. Dynamic Uncertainty: Requires theory extension

Reproducibility Assessment

Code Availability: Paper does not provide code, but implementation is relatively straightforward

Reproduction Difficulty (★★☆☆☆):

• SDP framework is standard and mature (CVX + MOSEK)
• Copositive relaxation has established theory (Berman & Shaked-Monderer book)
• Main challenge is correct implementation of 4^m constraint generation

Recommendations:

• Authors should provide open-source implementation
• Detailed copositive decomposition algorithm description needed
• More test cases should be provided

Selected Key References

1. Megretski & Rantzer (2002): "System analysis via integral quadratic constraints" - Foundational IQC theory
2. Veenman et al. (2016): "Robust stability and performance analysis based on integral quadratic constraints" - Comprehensive survey
3. Fetzer & Scherer (2017): "Full-block multipliers for repeated, slope-restricted scalar nonlinearities" - Full-block multiplier theory
4. Berman & Shaked-Monderer (2003): "Completely Positive Matrices" - Classical copositive theory
5. Boyd et al. (1994): "Linear Matrix Inequalities in System and Control Theory" - LMI methods foundation

Overall Rating: ★★★★☆ (4.5/5)

This is a high-quality theoretical control paper that solves an important open problem. The theoretical innovation is significant, the method is elegant, and the proofs are rigorous. The main limitations are dimension scalability and experimental breadth. For small-scale problems, this is currently the optimal method; for large-scale problems, it provides important theoretical benchmarks and new research directions. Recommended for publication in top-tier control journals (such as Automatica or IEEE TAC).