⚡ electrical engineering

State-space fading memory

This paper introduces a state-space definition of the fading-memory property as an extension of incremental input-to-output stability, demonstrating that incremental input-to-state stability implies this property for time-invariant systems and validating its application to current-driven memristors.

Original authors: Gustave Bainier, Antoine Chaillet, Rodolphe Sepulchre, Alessio Franci

Published 2026-03-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Gustave Bainier, Antoine Chaillet, Rodolphe Sepulchre, Alessio Franci

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: How Systems "Forget"

Imagine you are talking to a friend. If you tell them a secret, they remember it. But if you tell them a secret from 20 years ago, its influence on your current conversation is tiny. You might still remember the fact, but it doesn't change how you act right now as much as something you heard five minutes ago.

In the world of engineering and math, this concept is called Fading Memory (FM). It describes systems where the past matters, but its influence slowly fades away like a dying echo.

For a long time, mathematicians (specifically Boyd and Chua in 1985) had a great way to describe this for simple, linear systems (like a basic radio or a filter). They proved that if a system has "Fading Memory," you can build a very simple machine to mimic it: a linear machine that processes the input, followed by a simple "readout" knob that adjusts the output.

The Problem: Real life isn't linear. It's messy, complex, and non-linear (like a car engine, a brain, or a new type of electronic component called a memristor). The old math didn't work well for these complex systems.

The Solution: This paper introduces a new, modern definition of "Fading Memory" that works for these complex, non-linear systems. It connects the old "echo" idea to modern "stability" math.

The Core Concepts (With Analogies)

1. The "Echo Chamber" vs. The "Black Box"

Old Way (Operator Theory): Imagine a black box. You put a sound in, and a sound comes out. The old math said, "If the box forgets old sounds quickly enough, we can model it easily." But this didn't tell us what was happening inside the box.
New Way (State-Space): This paper looks inside the box. It looks at the "state" (the internal gears, the voltage levels, the brain cells). It asks: "How does the internal state react when the input changes?"

2. The "Memory Kernel" (The Fading Filter)

Think of a Memory Kernel as a special pair of sunglasses that you wear when looking at your past.

When you look at something that happened 1 second ago, the glasses are clear.
When you look at something from 1 hour ago, the glasses are slightly tinted.
When you look at something from 10 years ago, the glasses are so dark you can barely see it.

The paper defines a mathematical rule for these "glasses." It says: "The difference between two inputs in the past is weighted by how dark the glasses are at that time." If the glasses get dark fast enough, the system has Fading Memory.

3. The "Twin Test" (Incremental Stability)

To prove a system has Fading Memory, the authors use a "Twin Test."
Imagine you have two identical systems (System A and System B).

Scenario: You feed them the same input starting at time $t=0$ . But before that, they had different inputs.
The Question: Will System A and System B eventually behave exactly the same?
The Answer: If they have Fading Memory, YES. Even though they started with different histories, the "echo" of that difference fades away, and they converge to the same behavior.

The paper proves that if a system is Incrementally Input-to-State Stable (δISS)—which is a fancy way of saying "the system is robust and doesn't go crazy when inputs change"—then it automatically has this Fading Memory property (at least for a while and within certain limits).

The Real-World Example: The "Smart Resistor"

The paper uses a Memristor as a perfect example.

What is it? Imagine a resistor (a component that slows down electricity) that has a memory. Its resistance changes based on how much current has flowed through it in the past. It's like a door that gets harder to open the more people have pushed through it recently, but eventually, it forgets and resets.
The Application: The authors show that if this "Smart Resistor" is stable (doesn't explode or go wild), it has Fading Memory.
Why does this matter? Because if it has Fading Memory, we can use the old, simple math (the "Linear Machine + Readout Knob" model) to approximate how this complex, non-linear Smart Resistor behaves. This makes it much easier to design circuits and AI that use these components.

The "So What?" (Why should you care?)

Simplifying the Complex: It gives engineers a bridge. They can take a messy, complex, non-linear system (like a neural network or a biological cell) and prove that it behaves like a simple, predictable machine if it has this "Fading Memory" property.
Better AI: Many modern AI models (like Recurrent Neural Networks) rely on the idea of remembering the past. This paper provides the mathematical "safety check" to ensure these AI models won't get confused by old data and will focus on what's happening now.
Universal Approximation: It confirms that if a system is stable and forgets the past correctly, you can build a simple machine to copy its behavior almost perfectly.

Summary in One Sentence

This paper creates a new mathematical rulebook that proves if a complex, non-linear system is stable and "forgets" its past inputs at a steady rate, we can model it using simple, linear tools, making it easier to design and understand everything from smart electronics to artificial intelligence.

1. Problem Statement

The Fading Memory (FM) property, originally formalized by Boyd and Chua (1985) within an operator-theoretic framework, describes systems where the influence of past inputs on current outputs gradually diminishes over time. This property is fundamental to:

System approximation (e.g., Volterra series).
Reservoir computing and Recurrent Neural Networks (RNNs).
Stability analysis of nonlinear systems.

However, a significant gap exists in the literature:

Operator vs. State-Space: Boyd and Chua's definition relies on operators defined over the entire time axis (including negative time) and requires continuity with respect to a "fading norm." This does not translate directly to modern nonlinear state-space theory, where stability is often analyzed via incremental properties like Incremental Input-to-State Stability (δISS) and Incremental Input-to-Output Stability (δIOS).
Lack of Canonical Representation: While δISS guarantees strong stability properties (BIBO, CICO, PIPO), it does not inherently provide a canonical representation theorem (like the LTI + static nonlinearity structure) that Boyd and Chua established for FM operators.
The Core Question: Can a state-space definition of FM be formulated that bridges the gap between operator-theoretic FM and incremental stability (δISS/δIOS), thereby allowing the application of Boyd and Chua's approximation theorems to nonlinear state-space models?

2. Methodology

The authors develop a rigorous state-space definition of FM and analyze its relationship with existing stability concepts through the following steps:

Definition of State-Space FM: The authors propose a new definition (Definition 2) that extends δIOS. It introduces a memory kernel $w(\Delta t)$ $w (Δ t)$ (a continuous, non-increasing function decaying to zero) that explicitly bounds the decay of past input differences.
- The output mismatch is bounded by a term decaying with initial condition differences ( $\beta$ ) and a term weighted by the memory kernel acting on past input mismatches ( $\gamma$ ).
Connection to Operator Theory: The paper establishes a formal link between the new state-space definition and Boyd and Chua's original operator definition.
- It proves that for "well-behaved" inputs (Lipschitz continuous and equibounded), the state-space FM is equivalent to Uniform FM in the operator sense (uniform continuity of the memory functional).
- It introduces the concept of $(x_0, u_0)$ -realization to map state-space trajectories to operator inputs, allowing the transfer of approximation theorems.
Implication Analysis (δISS $\to$ FM): The authors investigate whether the well-known δISS property implies the new state-space FM property.
- They utilize Lyapunov converse theorems and Grönwall's inequality to construct a memory kernel from the δISS Lyapunov function.
- They demonstrate that while δISS does not imply FM globally without restrictions, it implies semi-global FM (on compact sets of initial conditions) under the assumption of equibounded inputs.
Application to Memristors: The theory is applied to current-driven memristors, a class of devices with inherent memory, to derive sufficient conditions for them to possess the FM property.

3. Key Contributions

State-Space Definition of FM: The paper introduces Definition 2, a state-space formulation of FM that incorporates a memory kernel. This definition is essentially a δIOS property with an explicit input-fading mechanism.
Equivalence Theorem: The authors prove that for time-invariant systems with equibounded, Lipschitz-continuous inputs, δISS implies semi-global state-space FM (Theorem 4). This bridges the gap between incremental stability and fading memory.
Extension of Approximation Theorems: By establishing the link between state-space FM and Boyd and Chua's Uniform FM, the paper shows that δISS systems can be approximated arbitrarily well by a Linear Time-Invariant (LTI) state-space model followed by a static nonlinear readout (Theorem 5). This generalizes Boyd and Chua's result to nonlinear state-space systems.
Memristor Analysis: The paper provides practical conditions (bounded current, δISS internal dynamics, Lipschitz memristance) under which current-driven memristors satisfy the FM property, validating the framework on a physical device.

4. Key Results

Theorem 1 & 9 (Equivalence): There is a rigorous equivalence between the state-space FM property and the operator-theoretic Uniform FM property, provided the system admits a specific realization mapping initial conditions to past inputs.
Theorem 4 (δISS $\implies$ FM): If a system is globally δISS and inputs are equibounded, the system possesses the state-space FM property on any compact set of initial conditions. This is proven by constructing a specific memory kernel $w(t)$ derived from the system's Lyapunov function.
Proposition 2 & 3 (System Properties): Systems with FM exhibit:
- CICO (Converging-Input Converging-Output): If inputs converge, outputs converge.
- PIPO (Periodic-Input Periodic-Output): Periodic inputs lead to periodic outputs, independent of initial conditions.
Memristor Conditions: A current-driven memristor has FM if its internal dynamics are δISS and the memristance function is Lipschitz continuous.

5. Significance and Impact

Unification of Stability Concepts: The paper unifies the operator-theoretic view of fading memory (crucial for machine learning and approximation theory) with the state-space view of incremental stability (crucial for control theory).
Practical Design for RNNs and Reservoir Computing: Since RNNs and reservoir computers are often modeled as state-space systems, this work provides a theoretical justification for why systems with δISS properties are suitable for these architectures. It guarantees that such systems can be approximated by simpler LTI+Nonlinear structures, aiding in model reduction and analysis.
New Stability Tool: The introduction of the memory kernel into state-space analysis offers a more nuanced tool than standard δISS. It allows for the derivation of time-varying bounds on input perturbations to achieve specific output tolerances (Proposition 4), which is more precise than static bounds.
Future Directions: The work opens avenues for exploring FM in interconnected systems (small-gain theorems), dissipativity theory, and partial differential equations (spatial "myopia"). It also suggests generalizing the fading norm from $L_\infty$ to $L_p$ spaces.

In summary, this paper successfully formalizes Fading Memory within the state-space framework, proving that δISS is a sufficient condition for semi-global FM. This result validates the use of powerful approximation theorems for a broad class of nonlinear systems, including those used in modern machine learning and physical devices like memristors.