The motor cortex generates a versatile set of network dynamics that support rhythmic and goal-directed movements. Computational studies have sought mechanistic explanations for this repertoire, but a unified account that explains both transient and self-sustained dynamics has been lacking. The authors propose that a small set of biologically motivated constraints can account for these diverse dynamical regimes.
Three simple constraints form the foundation of the approach reported in this preprint: adherence to Dale’s law (neurons are exclusively excitatory or inhibitory), an explicit constraint enforcing network stability, and nonlinear neuronal response functions. Together these ingredients are used to construct entire families of connectivity matrices that do not require iterative learning or manual fine-tuning of synaptic weights.
The work highlights that these constraints are sufficient, in combination, to generate network-level phenomena spanning transient responses, steady-state activity, and self-sustained periodic dynamics. The claim is made on the basis of computational reverse-engineering rather than fitting parameters to recorded datasets.
A single dynamical mechanism is identified as central to producing the observed repertoire: the interaction between non-normal amplification and neuronal nonlinearity. Non-normal amplification is described as inherent to Dalean networks; when coupled with nonlinear input–output relationships of neurons, this interaction can "ignite" activity and sustain multi-stable dynamics. In this picture, non-normality provides a substrate for transient amplification, while nonlinearity enables the translation of amplified perturbations into persistent or periodic network states.
The authors argue that this mechanism yields controllable dynamics that can be multi-stable, enabling transitions among transient, steady, and self-sustained regimes depending on inputs and network configuration.
Rather than training networks to match specific datasets, the approach reverse-engineers connectivity matrices that satisfy the three constraints. The resulting families of matrices can be generated without hand-tuning or learning of weights. Although the abstract summarizes the conceptual approach, details on the precise algorithms, parameter ranges, network sizes, or forms of neuronal nonlinearity are described in the main text and supplementary material of the preprint.
Because the networks are not fitted to empirical recordings, the method tests whether simple, biologically grounded constraints alone can produce dynamics that resemble those seen in motor cortex population activity.
Networks constructed with the three constraints reproduce a wide repertoire of dynamics: transient responses that amplify and decay, steady-state activity levels, and self-sustained periodic oscillations. These regimes emerge across entire families of Dalean connectivity matrices, indicating that the phenomena do not depend on precise weight patterns.
Importantly, without any data-driven fitting, the networks generate population-level signatures similar to those reported for motor cortex. This suggests that fundamental network structure and single-neuron nonlinearity can produce many features of cortical population dynamics observed during motor tasks.
The results imply that the complexity and flexibility of cortical population dynamics—at least those signatures compared in the study—may not require extensive sculpting by synaptic learning. Instead, basic biological constraints such as Dale’s law, enforced stability, and neuronal nonlinearity may be sufficient to produce transient and self-sustained activity patterns relevant for movement control. This shifts emphasis from learning-based explanations toward how inherent circuit properties can enable diverse functional regimes.
This report is a preprint and has not undergone peer review. The abstract provides the conceptual findings and central claims, but specific methodological details (for example, exact network models, nonlinearity forms, quantitative validation, and comparisons to experimental data) are contained in the full manuscript and supplementary materials. Those details were not enumerated in the abstract.
Readers should consult the full preprint and supplementary material for implementation specifics, parameter choices, and extended analyses that support the summarized conclusions.