Signal Decay Model

• Signal decay model is a mathematical framework describing the exponential decrease in signal amplitude and energy over time due to linear dissipation.
• It applies to diverse systems including mechanical resonators, MRI sequences, and deep neural networks, enabling parameter extraction and system design.
• Limitations in non-linear regimes have led to advanced optimization strategies that improve artifact correction and signal propagation in complex systems.

A signal decay model is a quantitative framework describing the attenuation of signal amplitude, energy, or information content over time due to intrinsic mechanisms such as dissipation, decoherence, or loss channels. Signal decay models arise in diverse domains, including classical mechanics, magnetic resonance imaging (MRI), and neural information propagation. They provide precise mathematical laws for the time dependence of the signal, often enabling parameter extraction, system design, or artifact correction.

1. Fundamentals of Signal Decay in Physical Systems

Signal decay in classical resonators is exemplified by the weakly damped harmonic oscillator with viscous friction. The canonical system consists of a mass $m$ on a spring of constant $k$, with a linear damping force $F_d(t) = -b v(t)$, $b > 0$. The undamped frequency is $\omega_0 = \sqrt{k/m}$, and the damping rate is $\beta = b/(2m)$. The weak-damping regime is defined by $\beta \ll \omega_0$.

The instantaneous total energy is $E(t) = \frac{1}{2} k x^2(t) + \frac{1}{2} m v^2(t)$. The dissipation rate is given by $dE/dt = -b v^2(t)$, leading to monotonically decreasing energy. Direct solution of the equations of motion yields the exact exponential amplitude decay, but alternative methods—such as the energy dissipation rate approach—yield equivalent results without explicit integration (Lelas et al., 2024).

2. Derivation of the Exponential Decay Model

The exponential signal decay law emerges from ansätze and energy bookkeeping. For a weakly damped oscillator, consider initial conditions with the same total energy $E_0$ but different phase:

• (1) $k$0, $k$1 (purely potential)
• (2) $k$2, $k$3 (purely kinetic)

Assume trial solutions

$k$4

with a slowly decaying envelope $k$5 ($k$6), $k$7.

By computing the sum of dissipation rates for both cases and using: $k$8 the governing ODE is: $k$9 yielding the decay solutions: $F_d(t) = -b v(t)$0

$F_d(t) = -b v(t)$1

$F_d(t) = -b v(t)$2

These exponential laws accurately describe amplitude and energy attenuation for $F_d(t) = -b v(t)$3 (errors $F_d(t) = -b v(t)$4; up to $F_d(t) = -b v(t)$5 for $F_d(t) = -b v(t)$6) (Lelas et al., 2024).

3. Generalizations and Extensions of Decay Laws

Signal decay models extend to any lightly damped resonator (mechanical, RLC circuit, or optical), provided the loss is linear in the generalized velocity. The generic procedure is:

• Construct the undamped normal mode,
• Introduce an envelope $F_d(t) = -b v(t)$7,
• Write the differential relation $F_d(t) = -b v(t)$8, and
• Extract the envelope law $F_d(t) = -b v(t)$9.

Non-linear dissipation mechanisms (Coulomb friction, quadratic drag) do not, in general, lead to simple exponential decay. Periodic averaging or special-case ansätze may be required, but the result is rarely an exponential (Lelas et al., 2024).

4. Signal Decay Models in Magnetic Resonance Imaging

In FLASH MRI and related spoiled gradient-echo sequences, signal decay is governed by a combination of $b > 0$0 (longitudinal recovery), $b > 0$1 (transverse decay), and the sequence of applied RF pulses and gradients. The idealized "Look-Locker" monoexponential decay,

$b > 0$2

with effective time constant

$b > 0$3

models the approach to steady-state. However, real signals are modulated by a substructure due to imperfect spoiling, higher-order echoes, and interference between FID and stimulated pathways (Weinmüller et al., 2024).

Optimization frameworks such as MR-zero leverage differentiable modeling to adjust flip angles and RF phases, restoring exponential decay and suppressing artifacts. Joint optimization of $b > 0$4 outperforms conventional quadratic spoiling, stabilizing both magnitude and phase decay and improving image quality (Weinmüller et al., 2024).

5. Exponential Signal Decay in Neural Information Processing

In the context of deep predictive coding (PC) networks, signal decay manifests as exponential attenuation of prediction error signals with network depth. If each layer’s error signal $b > 0$5 satisfies

$b > 0$6

then, under bounded weight and activation derivatives, the norm of $b > 0$7 satisfies $b > 0$8 for some $b > 0$9.

This exponential decay hampers the propagation of learning signals and causes vanishing updates in deep networks. The Error Optimization (EO) reparameterization eliminates this bottleneck by directly optimizing over errors $\omega_0 = \sqrt{k/m}$0 and collapsing the inference dynamics into an objective with global (nonlocal) Jacobian-vector products. EO thus enables the propagation of $\omega_0 = \sqrt{k/m}$1 magnitude signals across all layers in a single step, achieving convergence rates independent of depth and empirical performance competitive with standard backpropagation (Goemaere et al., 26 May 2025).

6. Limitations and Domain-Specific Extensions

Exponential signal decay models are exact in the linear, weakly damped regime and under idealized conditions (e.g., perfect spoiling in MRI; linear loss in mechanics). For systems with nonlinear dissipation, non-Gaussian echo pathways, or strongly depth-dependent spectral properties, deviations from exponential behavior occur. While optimization and reparameterization strategies can extend the regime of accurate exponential modeling, new phenomena such as oscillatory modulations or universal plateaus may emerge. The universality of the Look-Locker target in MRI, for example, contrasts with tissue-specific responses achieved by alternative target functions, whose optimal RF trains lack the cross-tissue invariance (Weinmüller et al., 2024).

7. Applications and Implications Across Fields

Signal decay models underpin experimental design, quantitative parameter extraction, and the development of artifact-free measurement protocols. In mechanics, they are essential for accurate modeling of resonator decay without recourse to full analytical solutions. In MR imaging, signal decay models guide sequence design and artifact minimization. In neural computation, eliminating exponential signal decay enables the practical training of deep biologically inspired systems with local rules. The mathematical transparency and extensibility of signal decay models ensure their continued relevance in both foundational theory and applied engineering (Lelas et al., 2024, Weinmüller et al., 2024, Goemaere et al., 26 May 2025).