A Note on BIBO Stability

Michael Unser

Abstract

In undergraduate signals and systems education, a linear shift-invariant (LSI) system is traditionally defined as Bounded-Input Bounded-Output (BIBO) stable if and only if its impulse response is absolutely integrable ($L_1$). However, this standard definition contains a subtle mathematical inconsistency: it technically excludes the identity operator, as the Dirac delta distribution is not a measurable function in the Lebesgue sense. This paper addresses this "mathematical hygiene" issue by providing a rigorous characterization of BIBO stability using the framework of bounded Radon measures. By refining the definitions of the underlying function spaces, we reconcile the intuitive stability of impulsive systems with formal distribution theory.

Introduction

The concept of Bounded-Input Bounded-Output (BIBO) stability is foundational to the study of linear shift-invariant (LSI) systems. The prevailing wisdom in engineering textbooks asserts that a continuous-time system is BIBO stable if and only if its impulse response $h(t)$ satisfies the condition: $$\int_{-\infty}^{\infty} |h(t)| dt < \infty$$

While this condition ($h \in L_1(\mathbb{R})$) is sufficient for a wide class of systems, it fails to account for the most fundamental stable system: the identity operator. The impulse response of the identity operator is the Dirac delta distribution $\delta(t)$. Because $\delta(t)$ is not a "measurable function" in the Lebesgue sense, it does not belong to $L_1(\mathbb{R})$. Consequently, a literal interpretation of the standard textbook definition leads to the paradoxical conclusion that the identity operator is unstable. This paper seeks to resolve this discrepancy by employing tools from functional analysis and distribution theory.

The Mathematical Framework

To provide a rigorous correction, the analysis moves beyond simple function spaces and utilizes the framework of bounded Radon measures, building on the foundational work of Laurent Schwartz and Lars Hörmander.

Key Mathematical Tools

  • Schwartz’s Kernel Theorem: This is used to define the most general class of LSI operators that map test functions to distributions.
  • Banach Space Theory: The paper analyzes the mappings between different function spaces, specifically $L_\infty$ (bounded functions), $C_b$ (continuous bounded functions), and $C_0$ (functions vanishing at infinity).
  • Radon Measures: The space $\mathcal{M}(\mathbb{R})$ of bounded Radon measures is introduced as the appropriate domain for the impulse responses of stable systems. This space is the topological dual of $C_0(\mathbb{R})$ and naturally encompasses both standard integrable functions and impulsive distributions like the Dirac delta.

Key Insights and Results

The Continuity Trade-off

A significant contribution of this work is the distinction between systems that merely produce bounded outputs and those that produce continuous bounded outputs.

Theorem 2 demonstrates that if an impulse response is a measurable function, the system maps $L_\infty$ to $C_b$ (continuous bounded functions) if and only if $h \in L_1(\mathbb{R})$. This justifies the traditional $L_1$ condition in contexts where "regularization" or smoothing of the input signal is a requirement.

Theorem 3 (The Main Correction) provides the general fix: An LSI system is BIBO stable (mapping $L_\infty$ to $L_\infty$) if and only if its impulse response $h$ is a bounded Radon measure ($h \in \mathcal{M}(\mathbb{R})$). This definition is inclusive of the identity operator and digital filters (which consist of sums of delayed impulses), providing a unified framework for both analog and digital contexts.

Refining the "Worst-Case" Argument

Traditional proofs of the necessity of the $L_1$ condition often rely on a "worst-case" input $f(t) = \text{sign}(h(-t))$. However, this argument is often mathematically flawed in continuous time because it ignores sets of measure zero and does not guarantee that the resulting output is well-defined or continuous. This paper rectifies these proofs using truncated signals and limit processes to ensure mathematical rigor.

Sharp Bounds and Operator Norms

The paper proves that the operator norm of these stable systems is exactly equal to the total-variation norm of the measure (or the $L_1$ norm in the case of functions). This provides the sharpest possible bound for stability, ensuring that the theoretical limits match the physical behavior of the systems.

Conclusion

By shifting the definition of BIBO stability from absolute integrability in $L_1$ to membership in the space of bounded Radon measures $\mathcal{M}(\mathbb{R})$, we resolve a long-standing logical gap in engineering literature. This correction does not merely satisfy a need for mathematical rigor; it provides a consistent bridge between continuous-time theory and discrete-time applications, where impulsive responses are the norm rather than the exception. While this framework still correctly identifies truly unstable operators—such as the Hilbert transform or the derivative operator—it ensures that the identity operator and its derivatives are treated with the mathematical respect their stability warrants.

Summary updated

Abstract

The statements on the BIBO stability of continuous-time convolution systems found in engineering textbooks are often either too vague (because of lack of hypotheses) or mathematically incorrect. What is more troubling is that they usually exclude the identity operator. The purpose of this note is to clarify the issue while presenting some fixes. In particular, we show that a linear shift-invariant system is BIBO-stable in the L ∞ -sense if and only if its impulse response is included in the space of bounded Radon measures, which is a superset of L 1 (R) (Lebesgue's space of absolutely integrable functions). As we restrict the scope of this characterization to the convolution operators whose impulse response is a measurable function, we recover the classical statement.
BibTeX
@article{unser2020note,
  author = {Michael Unser},
  title = {{A Note on BIBO Stability}},
  journal = {IEEE Trans. Signal Process.},
  year = {2020},
  doi = {10.1109/tsp.2020.3025029},
}
Figure