Spike budgets depend on controller wiring, not gain
A spiking neural controller that implements exactly the same closed-loop dynamics can need 200 times more spikes depending only on how its neurons are wired to the plant's inputs and outputs. The authors prove the stability guarantees survive the rewiring, then show that the familiar engineering instinct, lower the gain to save energy, is the wrong lever.
Source: Towards Design of Efficient Neuromorphic Controllers, arXiv:2611.13874, 10 November 2026. Primary source. Read: full text (PDF, 7 pages, retrieved 2026-11-13).
What the work claims
This is a theory-plus-simulation paper from Koen Scheres (KU Leuven) with Erik Steur and Maurice Heemels (TU Eindhoven), and it is a follow-up to the group's earlier result, described within the paper, that a single layer of integrate-and-fire neurons can emulate a classical static output feedback controller with formal, global practical-asymptotic stability guarantees. That first result established that spiking control can be provably correct. This paper asks the question the first one deferred: how many spikes does the provably correct controller actually burn, and can that number be reduced without surrendering the guarantees?1
The bold claim is twofold. First, spike sparsity is not a property of the controller being emulated but of the network topology and synaptic weights used to emulate it: the same controller gain, acting on the same plant, can be realized with radically different spike counts. Second, the standard synthesis strategy of penalizing control effort to obtain a lower-gain, and supposedly cheaper, controller does not reliably produce a sparser implementation. Direction, not magnitude, is what matters.
How it works
The setup is an emulation architecture. Take a linear time-invariant plant that can be stabilized by a static output feedback gain K, meaning a constant matrix that maps measured outputs directly to control inputs with no internal dynamics. The earlier construction implements K with a single layer of integrate-and-fire neurons. For every input-output pair, two neurons integrate the positive and negative parts of the measured signal; whenever a neuron's internal state crosses a threshold, it emits a weighted spike and resets. The synaptic weight divided by the threshold equals the desired gain entry, so the spike train's average effect reconstructs the control signal, up to an error bound the earlier theorem makes explicit. Practical stability means the state converges to a neighborhood of the origin whose size is set by that bound.
The paper's diagnosis of waste is geometric. When the plant's output sits near the hyperplane where the true control input should be zero, the neurons implementing the individual gain entries are all spiking vigorously in opposite directions; the spikes are supposed to cancel, but they only cancel if they arrive at exactly the same instant, which essentially never happens. The plant's own state dynamics perform an imperfect summation of opposing spike trains, and the imperfection is paid for in spikes.
The fix is a change of basis buried in the wiring. Any invertible linear transformations applied to the controller's inputs and outputs can be folded into the synaptic weights and thresholds, so the group implements a transformed gain while the closed-loop dynamics stay identical. Three choices are tested: orthogonalizing the plant's input directions with a Gram-Schmidt projection, projecting onto the dominant eigenvectors of the closed-loop matrix, and a singular value decomposition of K, which is the sparse option because only the nonzero singular directions need neurons. A proposition in the paper shows the error bound can grow by at most the norm of the input transformation, and an orthonormal transformation leaves it exactly unchanged: you are guaranteed to preserve accuracy while changing only the direction in which spikes act.
The numerical demonstration uses a standard benchmark, the vertical take-off and landing helicopter model, with a two-input, two-output stabilizing gain. Over a 25 second simulation, the original wiring needs 1,103,666 spikes. Orthogonalizing the input directions cuts this to 644,690. Projecting onto dominant eigenvectors cuts it to 463,427. The singular value decomposition wiring, which halves the neuron count for this system, needs 5,122 spikes: 0.46 percent of the baseline, and the trajectory error actually improves rather than degrades. The paper is careful to also test the synthesis route: redesigning K for low gain yields a controller needing 41,317 spikes with the original wiring, and the structured wirings still beat or match it. Low gain is not the lever.
Where a skeptic should push
The single most load-bearing assumption is that spike count is a meaningful proxy for energy. The paper equates efficiency with sparse spiking, which is defensible for neuromorphic silicon, where energy per spike is the dominant cost, but the simulations count spikes in a mathematical model, not joules on a chip. No hardware measurement appears anywhere. A second caveat the authors themselves state: the minimum inter-spike interval can worsen even when the average improves, and the average spike count is notoriously hard to predict, a difficulty they note is shared with the event-triggered control literature. The 200-fold figure is one benchmark system, four wiring choices, and a single 25 second horizon. It is illustrative, not a scaling law.
There is also an honesty worth noting in the paper's own conclusion. The authors observe that two mechanisms the real brain uses to achieve sparse activity are absent from their model entirely: inhibition, which lets biological neurons actively suppress each other and raise the distance to threshold, and temporal or correlation coding, which they speculate could be far more efficient than the rate coding their emulation uses. In other words, the architecture that provably works is missing the very features that make biological controllers sparse, and the paper says so plainly. That is a map of the open problem, not a defect to hide.
Organoid intelligence and the spike budget
For organoid intelligence, the currency of closed-loop computing on living tissue is not joules of electricity but biological events. Every stimulation pulse in a closed-loop organoid experiment costs something the silicon version does not: metabolic load on the tissue, calcium influx, evoked synaptic plasticity, and a nudge along the excitability axis that, pushed hard enough, ends in hyperexcitable, seizure-like states that destroy the preparation as a computing substrate. If you run a feedback controller through a multielectrode or optogenetic interface, this paper's central result applies almost verbatim, with harsher economics: the wiring of your interface, which electrodes you read, which you stimulate, and how you encode the control signal in stimulation pulses, can change the biological spike tax by orders of magnitude while implementing the identical computation.
The non-obvious implication cuts against a natural intuition. If a closed-loop protocol is stressing the tissue, the reflex is to turn the stimulation intensity down, which is exactly the low-gain redesign this paper shows is unreliable. The better lever, per this work, is to restructure the input-output basis of the loop: align stimulation channels with the dominant directions of the tissue's response, remove redundant opposing channels whose effects cancel in the tissue's state rather than in your pulse timing, and prefer a small number of structured drivers over many individually weak ones. Done well, this could let a training loop deliver the same information to a dish with two orders of magnitude less disturbance. Done naively, a misdirected controller does not merely waste energy; in living tissue it manufactures pathology, because the redundant, non-cancelling pulses keep arriving and each one induces plasticity. A controller that is 200-fold sparser is also 200-fold gentler, and gently stimulated tissue retains trainability far longer.
The genuine threat is the paper's own identified gap. The authors name inhibition as the missing mechanism that makes biological controllers sparse, and patterned, well-timed inhibitory control is precisely what organoid interfaces are worst at delivering: optogenetic inhibition is slow and requires genetic modification, and multielectrode arrays stimulate excitatory-dominated cultures far more easily than they suppress them. The field's most celebrated closed-loop demonstrations today run excitation-heavy loops of the kind this paper proves are the wasteful topology. The opportunity is a research program the paper effectively specifies: closed-loop organoid systems whose stimulation policies are synthesized with direction-awareness and an explicit biological spike budget, treating the tissue's metabolic and plasticity cost as the objective, not an afterthought.
The bottom line
Established: for emulation-based spiking control of linear systems, network topology and synaptic weight structure change spike count by one to two orders of magnitude at identical closed-loop dynamics and preserved stability guarantees, demonstrated in simulation on a benchmark plant. Established: penalizing control gain is not a reliable sparsity strategy. Open: everything about inhibition, temporal coding, and dynamic (recurrent) controllers, which the authors flag as the routes to biological-grade sparsity. For organoid intelligence the result is a design principle and a warning: measure the biological spike budget of your closed loop, restructure the interface basis before you lower the gain, and treat the field's inhibitory deficit as a first-order engineering problem, because the theory says that is where the efficiency actually lives. What would confirm it: a closed-loop organoid experiment showing that a direction-aware stimulation policy sustains training with measurably less metabolic stress and plasticity drift than an intensity-matched naive policy.
Frequently asked questions
What exactly did the authors change to get 200 times fewer spikes?
They changed how the spiking network is wired to the plant, not the plant or the control law. By folding invertible linear transformations of the controller inputs and outputs into the synaptic weights and neuron thresholds, they implemented the same closed-loop dynamics through a different basis. A singular value decomposition wiring, which needs only the nonzero singular directions of the gain matrix, reduced the spike count from 1,103,666 to 5,122 in the 25 second helicopter simulation while also reducing trajectory error.
Do the formal stability guarantees survive the rewiring?
Yes. The paper proves the emulation error bound can grow by at most the norm of the input transformation in general, and that orthonormal transformations leave the bound exactly unchanged. The stability guarantee from the group's earlier work therefore carries over; the simulations additionally show the realized error improving for every tested transformation.
Why is lowering the controller gain not the right way to save spikes?
Because spike count in this architecture tracks the direction of the controller's action relative to the plant's dynamics, not the magnitude of the gain. A redesigned low-gain controller still needed tens of thousands of spikes with the original wiring, while a high-gain controller with a well-chosen structured wiring needed only a few thousand. The paper concludes that synthesis procedures must account for directionality explicitly.
Why does this matter for computing on living neural tissue?
In organoid closed-loop systems the cost of a spike is biological: metabolic stress, calcium load, and stimulation-evoked plasticity, with a failure mode of hyperexcitability. A stimulation interface whose basis fights the tissue's natural dynamics pays that biological cost many times over for the same information, and the redundant pulses actively reshape the tissue. Restructuring the interface basis can in principle deliver identical training with far less disturbance.
What mechanism is missing from the spiking models in this paper?
Inhibition. The authors note that biological neurons can suppress each other and raise the distance to threshold, a mechanism their single-layer excitatory architecture lacks, and they suggest two-layer architectures with competing inhibitory goals as a route to further sparsity. They also flag temporal and correlation coding as unexplored alternatives to the rate coding their emulation uses.
How large is the demonstrated effect?
The headline numbers come from one benchmark system, the standard VTOL helicopter model: 1,103,666 spikes for the original wiring over 25 simulated seconds, versus 5,122 for the singular value decomposition wiring, 463,427 for an eigenvector projection, and 644,690 for orthogonalized inputs. The paper presents these as illustrative; the qualitative claim, that topology dominates sparsity, is supported by the proofs, but the size of the win on any other system is an open empirical question.
References
- K. Scheres, E. Steur, and M. Heemels. Towards Design of Efficient Neuromorphic Controllers. arXiv:2611.13874 [eess.SY]. 2026. https://arxiv.org/abs/2611.13874. Accessed 2026-11-13.