To process information efficiently, neural circuits must perform a delicate balancing act. On one hand, they need high precision: whether they are locking onto the exact phase of a sensory stimulus, suppressing chaotic background noise, or maintaining a tight balance between excitation and inhibition, accuracy requires strong, responsive feedback.
However, a fundamental physical constraint thwarts this pursuit of precision: transmission delays. Because signals take time to travel across axons and synapses, a network's feedback loop is always acting on slightly outdated information. When you combine high feedback sensitivity (gain) with even a tiny temporal lag, the system can easily overcorrect, transforming stabilizing negative feedback into destructive, runaway oscillations.
This tension defines a universal trade-off in neural dynamics. Pushing a network toward maximum precision inevitably drives it toward the brink of dynamical instability.
The Intuition: The Danger of Overcorrection
To understand why this happens, imagine adjusting the water temperature in a shower where there is a three-second delay between turning the knob and feeling the result. If the water is too cold, and you aggressively crank the knob toward hot (high gain), you won't realize you've overcorrected until three seconds later, when the water becomes scalding. In response, you panic and crank the knob all the way back, causing the shower to freeze. You are trapped in an unstable, oscillating loop.
In recurrent neural networks, a similar phenomenon occurs. As demonstrated by Kadmon et al. (2020), networks use strong negative feedback to suppress internal noise and chaotic fluctuations, allowing for high-precision predictive coding. However, as the feedback strength increases to minimize coding errors, it interacts with inherent synaptic transmission delays.
Beyond a certain threshold, this interaction causes the network to break out into parasitic, system-wide "ringing" oscillations. The authors of a recent study on efficient coding with chaotic neural networks note that while feedback successfully dampens chaotic fluctuations in a target subspace, excessive gain combined with axonal lags induces resonant oscillations that ultimately degrade signal quality.
To avoid this, Kadmon et al. (2020) derived a mathematical heuristic for the optimal operating point: to maximize accuracy without risking signal degradation, the feedback gain should be set to exactly half the critical value that triggers these unstable oscillations.
Phase-Locking and the Collapse of Oscillations
When we look at individual, coupled neuronal oscillators, this trade-off manifests as a direct conflict between timing accuracy and dynamical robustness.
Neurons often synchronize their firing to communicate or encode information, acting much like biological Phase-Locked Loops (PLLs). If two non-identical neurons need to lock onto a specific reference phase with high precision, their synaptic plasticity rules must be highly sensitive to timing errors.
[ High Feedback Gain ] ---> [ Overcorrection / Lag ] ---> [ Bifurcation / Chaos ]
|
v
[ Complete Activity Cessation ]
As explored by Tyukin et al. (2012), attempting to force higher precision by ramping up the sensitivity or gain of these synchronization rules destabilizes the system. Instead of achieving perfect alignment, the system undergoes a series of dynamical bifurcations.
In their companion paper, Tyukin et al. (2012) detail the exact mathematical routes to this instability. When the feedback gain crosses a critical threshold, the system can undergo a Neimark-Sacker bifurcation, causing the steady synchronized state to dissolve into wild phase oscillations. Alternatively, it may follow a period-doubling route into chaotic firing patterns.
Most drastically, excessive sensitivity can induce a state where the aggressive regulatory feedback suppresses the neuron's internal excitability so severely that its rhythmic behavior ceases entirely. The drive for perfect regulation accidentally kills the very oscillations it was trying to steer.
Network-Scale Consequences
This fundamental limit on feedback precision ripples outward, constraining how neural populations code information, generate movement, and learn.
The Collapse of Balanced Coding
In "tightly balanced" spiking networks, high precision is achieved because excitatory and inhibitory inputs continuously cancel each other out. However, Timcheck et al. (2022) showed that when inhibitory signals are delayed by axonal or synaptic transmission, this cancellation fails over short windows. The network suffers from a brief blindness where it cannot prevent redundant firing, leading to a collapse in computational efficiency. Interestingly, the authors found that an optimal level of background noise can actually disrupt this pathological, delay-induced synchronization, acting as a buffer that restores coding efficiency.
Synchronization vs. Pattern Generation
In motor systems, the goal is often not global synchronization, but rather carefully timed, phase-shifted sequences (like the traveling waves needed for locomotion). Patel et al. (1999) discovered this constraint when building neuromorphic VLSI hardware for axial locomotion: if connection strengths (gains) are turned up too high, the system defaults to global synchronization, wiping out the smooth traveling waves necessary for movement.
Similarly, Ali et al. (2018) demonstrated that while strong feedback loops are excellent for stabilizing a network against external noise, they severely limit temporal flexibility. In highly connected recurrent circuits, a local change meant to adjust the timing of one specific motor element propagates through the feedback loops, causing widespread interference across the entire sequence.
Instability in Recurrent Learning
The precision-stability trade-off also plagues artificial and brain-inspired algorithms that rely on recurrent dynamics to compute error signals. For instance, Equilibrium Propagation requires networks to settle into a stable steady state before calculating gradients. As Liu et al. (2026) point out, the simultaneous interaction of feedforward and feedback signals in these architectures frequently causes numerical instability or interference. Resolving this requires carefully scaling down feedback signals to balance the quality of the error signal against the speed and stability of the network's convergence.
Summary
Whether analyzing a single pair of phase-locked neurons, a neuromorphic robot, or a deep recurrent network, the underlying physics remains identical. Negative feedback is a powerful tool for enforcing precision and suppressing noise, but it is fundamentally bounded by time. Recognizing the hard boundaries imposed by transmission delays allows us to better understand both the architectural constraints of the biology and the design principles required for stable, brain-inspired computing.
Key Foundational Work
Adaptive and Phase Selective Spike Timing Dependent Plasticity in Synaptically Coupled Neuronal Oscillators (2012) — contributed 5 concept elements to this direction.
Precise Self-tuning of Spiking Patterns in Coupled Neuronal Oscillators (2012) — contributed 3 concept elements to this direction.
A Neuromorphic VLSI System for Modeling the Neural Control of Axial Locomotion (1999) — contributed 1 concept elements to this direction.