Research analysis · Closed-loop stimulation

The safety rule that forbids holding a brain circuit still

Every clinical neurostimulator obeys charge balance: zero net injected charge per stimulation cycle, because residual charge corrodes electrodes and damages tissue. Ruan, Liu, Stursberg, Fan, van Gerven and Qin treat that rule not as an engineering footnote but as a constraint on what control can do, and prove it forbids something most closed-loop schemes implicitly assume: holding a neural system at a nonzero steady state. Their alternative, stabilizing a periodic orbit instead of a fixed point, works on a Hodgkin-Huxley neuron, including one already locked into tonic spiking.

Source: Charge-Balanced Control of Linear Systems: Controllability, Stabilizability, and Periodic Regulation, arXiv:2611.02288, 2026-11-01. Primary source. Read the full HTML text of the preprint, including proofs and the Hodgkin-Huxley application. The work was supported by the Dutch Brain Interface Initiative through NWO; that funding appears here as context, not as the subject.

What the work claims

This is a control-theory paper motivated by electrical neuromodulation, and its claims are theorems about discrete-time linear systems whose input must satisfy a charge-balance constraint: the sum of the control inputs over each cycle of h steps must be zero. The authors define charge-balanced controllability and charge-balanced stabilizability and give necessary and sufficient conditions for both.1

Two results carry the weight. First, a system under charge-balanced inputs cannot be held at any nonzero steady state: if you need the state to converge to a constant other than the origin, no charge-balanced input sequence will do it. The authors therefore recast nonzero regulation as tracking a charge-balanced periodic orbit, and characterize when convergence to such an orbit is achievable, with optimal feedback controllers for both stabilizing the origin and tracking the orbit.1 Second, the framework is validated on a Hodgkin-Huxley neuron: a cycle-wise controller with charge-balanced pulse pairs suppresses tonic spiking and returns the neuron to its unstable equilibrium, even when stimulation begins after spiking is fully established.1

How it works

Charge balance changes the controllability problem in a specific way. An input sequence whose entries sum to zero over each h-step cycle is not an arbitrary input: it lives in a subspace. The authors lift the dynamics to the cycle level, and the charge-balance subspace selects which lifted modes can be reached. Their stabilizability theorem for cycles of length h at least 2 has three parts: the unconstrained system must be stabilizable; the eigenvalue 1 must not belong to the system matrix; and for any complete set of eigenvalues forming a spectral h-cycle on or outside the unit circle, the intersection of the corresponding control input subspaces must be trivial.1

The eigenvalue condition deserves plain language. A mode at eigenvalue 1 is conserved at every cycle boundary under any charge-balanced input: the charge constraint averages exactly over the mode that sums, so that direction can never be steered. The spectral-cycle condition guards against aliasing: when you lift the dynamics to h-step cycles, distinct eigenvalues can collide into one, and if their input paths overlap, controllability is lost in a way the unconstrained test would never detect.1

The Hodgkin-Huxley demonstration uses the standard model with membrane capacitance 1 microfarad per square centimetre, maximal conductances 120, 36 and 0.3 millisiemens per square centimetre for sodium, potassium and leak, and reversal potentials of 50, minus 77 and minus 54.4 millivolts. An applied current of 20 microamperes per square centimetre drives tonic spiking around an unstable equilibrium near minus 56.59 millivolts. The authors linearize at that equilibrium, discretize with a zero-order hold at 0.5 milliseconds, and choose a cycle length of h equals 2. Because the input is constant between samples, forcing the two samples in each cycle to sum to zero makes the continuous-time injected charge exactly zero over every cycle. The cycle-wise optimal controller, designed in scaled coordinates and then applied to the full nonlinear neuron, silences the spiking. When stimulation starts at 20 milliseconds, before spiking establishes, and when it starts at 100 milliseconds, after spiking is established with the input clamped to plus or minus 20 microamperes per square centimetre, the neuron returns to equilibrium in both cases.1

Where a skeptic should push

The theorems are about linear systems; the neuron is the only nonlinear test, and it is one cell. Everything in the Hodgkin-Huxley demonstration is measured and controlled: the full four-dimensional state, membrane voltage plus all three gating variables, is assumed available to the controller. No neural interface, on tissue or on an organoid, observes gating variables, and state estimation under charge-balanced actuation is exactly the kind of problem this framework does not yet address. The step from a linearization-valid controller silencing one neuron to a network whose dynamics are high-dimensional, stochastic and state-partially-hidden is where the real difficulty lives, and the paper does not cross it.

There is also a modeling assumption worth naming: charge balance is enforced per sampled cycle by pairing inputs, and under saturation the authors saturate a block coordinate so each pair stays balanced. Real stimulators deliver biphasic waveforms whose imperfect symmetry, and whose charge recovery phases, are lumped away here. Finally, stabilizing an unstable equilibrium of a neuron is a demonstration of suppression, not of computation: it shows charge-balanced control can reach and hold a biologically meaningful target, but it says nothing yet about steering a circuit through a trajectory of your choosing. The framework's periodic-orbit machinery is precisely the piece that would enable trajectory-like targets, and it is validated here only in its most conservative use.

What charge balance forbids in organoid training

Every serious proposal for training organoid computers is a closed-loop stimulation protocol: read the activity, compute an error or a reward, inject corrective current, repeat. This paper puts a hard constraint on that loop that the field mostly treats as an afterthought. If your stimulation is charge-balanced, and for the long-term health of the tissue it must be, then there is an entire class of control objectives you simply cannot program: you cannot hold the network at a constant elevated or suppressed activity setpoint. The mathematics forbids it. Any training scheme whose implicit goal is "drive the tissue to this steady firing level and keep it there" is asking for a nonzero steady state, and the correct response is not a better optimizer but a different target class, namely periodic orbits: stimulation patterns that nudge the network along a repeating trajectory and let its own dynamics do the holding.

The deeper warning concerns which modes are controllable at all. Biological neural networks are full of near-integrator behavior: synaptic integration, slow afterpotentials, adaptation currents. A mode sitting at or near eigenvalue 1 is, by this paper's own theorem, the mode a charge-balanced input cannot steer at cycle boundaries. Slow dynamics are the computationally interesting ones in reservoirs and in cortex, and they are precisely the ones a safe stimulator has the least authority over. That is the non-obvious implication: the safety constraint does not just limit amplitude, it limits reachability along the slowest, most computation-bearing directions. The opportunity is equally specific. The orbit-tracking machinery gives wetware training a legitimate design grammar: pick a target orbit with zero-mean drive, enforce balance per cycle, and guarantee convergence under the stated conditions. A protocol built this way would be the first with a tissue-safety proof attached to its training loop, and that is the kind of guarantee regulators and reviewers will eventually demand. The threat is that most current closed-loop organoid work runs charge-imbalanced or unreported stimulation; if this analysis is right, some fraction of what the field calls training is quietly damaging its substrate in exchange for dynamics it could not hold anyway.

The bottom line

Established, as mathematics with a simulation anchor: charge-balanced inputs on a discrete-time linear system admit necessary and sufficient controllability and stabilizability conditions; nonzero steady states are unreachable; periodic orbits are the correct nonzero targets; and the resulting controller suppresses tonic spiking in a Hodgkin-Huxley neuron, including from an established spiking state. Unestablished: performance with partial state observation, on networks, with realistic biphasic hardware and biological noise. What would confirm relevance to biological computing: a closed-loop protocol on a neural organoid that enforces per-cycle charge balance, targets a periodic orbit, and shows trained task performance at least matching an unconstrained-stimulation baseline. What would break the framing: evidence that biological tissue tolerates small chronic charge imbalance without damage, which would relax the constraint but trade it for an ethical problem of a different kind.

Frequently asked questions

What does charge balance mean in neural stimulation?

It means the net charge injected over each stimulation cycle is zero. The rule exists because residual charge drives irreversible electrochemical reactions, degrades electrodes and damages tissue.

What can charge-balanced control not do?

It cannot hold a system at any nonzero steady state. The paper proves this and replaces fixed-point targets with charge-balanced periodic orbits, for which it gives stabilizability conditions and optimal controllers.

How was it demonstrated?

On a Hodgkin-Huxley neuron driven into tonic spiking by 20 microamperes per square centimetre. A controller applying balanced two-step input pairs, sampled at 0.5 milliseconds, suppressed spiking and returned the neuron to its unstable equilibrium, even when started after spiking was established.

Why does the eigenvalue 1 matter?

A mode at eigenvalue 1 sums its inputs perfectly over a balanced cycle, so the charge constraint cancels exactly along that direction and the mode cannot be steered. Slow, integrator-like modes, common in neural tissue, are the least controllable under charge balance.

What does this change for organoid computers?

Closed-loop training protocols must abandon constant setpoints and design zero-mean periodic stimulation targets instead. It also suggests many current training stimulations may be chronically imbalanced, risking silent tissue damage for control authority that charge balance rules out anyway.

References

  1. J. Ruan, Z. Liu, O. Stursberg, Y. Fan, M. van Gerven, Y. Qin. Charge-Balanced Control of Linear Systems: Controllability, Stabilizability, and Periodic Regulation. arXiv:2611.02288. 2026. https://arxiv.org/abs/2611.02288. Accessed 2026-11-05.