When one electrode suffices and when half the network is not enough

What the work claims

This is a pure theory paper: theorems, constructive proofs, and a numerical verification, no biological or hardware experiment. The authors ask a class-level question about recurrent networks with linear-threshold or ReLU update functions: under a common notion of global finite-horizon observability, meaning any two distinct initial states must produce distinguishable output sequences within a finite window, what is the minimum number of nodes you must measure, in the best and worst case¹?

The answer, established rigorously, is that the minimum is set not by network size alone but by the combination of update rule and state domain. Over the binary domain, they construct K-linear-threshold networks, for K at least 4 and node counts n equal to h times K with h odd and coprime to K, that are observable from a single node, and they exhibit the explicit 4-node case. Binary-valued K-ReLU networks behave very differently: they are dynamically equivalent to K-AND Boolean networks, which inherit class-level bounds requiring a nontrivial fraction of all nodes in the worst case. For nonnegative-valued ReLU networks the problem collapses to classical linear observability, governed by the Kalman rank condition, with one node sufficient in the best case. For general real-valued ReLU networks, they prove that observability requires at least n/2 observation nodes when node updates may depend on arbitrarily many state variables, and they construct a network attaining exactly K = n/2 nodes, so the bound is tight. Numerically, across 100 randomly generated networks for each of three sizes, every observation set smaller than n/2 admitted a counterexample pair of indistinguishable initial states.

How it works

The engine behind the results is what the authors call an affine-region rank obstruction. A ReLU network is piecewise affine: on any region where the activation pattern, the pattern of which units are switched on, is fixed, the finite-horizon map from initial state to output sequence is affine, and its linear part can lose rank. If the rank of that linear part is smaller than the number of freely varying initial-state components, two distinct initial states map to identical outputs, permanently indistinguishable no matter how long you watch. State-dependent ReLU deactivation is what creates the rank loss, and over general real-valued domains the deactivation pattern varies with the initial condition, so no fixed linear analysis can rescue you. Summing the rank contributions across a finite horizon yields the n/2 lower bound: with fewer than n/2 sensors, the observation map cannot be injective somewhere in state space.

The binary-side results come from a different mechanism entirely: information concentration through discrete dynamics. The single-node-observable K-linear-threshold construction works like a shift register made of threshold blocks. A chosen block's state is rotated through a transformation that preserves enough structure that, by watching one node over time, the initial configuration of the block that seeded the sequence can be decoded; the coprimality condition between block count h and block size K prevents short cycles from aliasing distinct initial conditions. Because the state space is finite, time itself multiplexes the information of many nodes into one measured trajectory. The reconstruction procedure needs only on the order of log base 2 of the block size plus one time steps, so observability is not just possible but fast.

The equivalence between binary K-ReLU networks and K-AND Boolean networks transfers existing class-level bounds: the worst-case minimum for K-AND networks is bounded below by [(1-K) + ((2^K - 1)/2^K) log2(2^K - 1)] times n and above by ((2^K - K - 1)/(2^K - 1)) times n, and those bounds now apply verbatim to binary K-ReLU networks. Whenever the lower bound exceeds one, every binary K-ReLU network of that class needs more observation nodes than the constructed K-linear-threshold networks do, isolating the update rule, not the state domain, as the culprit. The nonnegative ReLU case is the boundary where the nonlinearity never fires: if all relevant pre-activations stay nonnegative, ReLU is the identity, the network is a positive affine system, and the hundred-year-old linear theory applies unchanged.

Where a skeptic should push

The n/2 bound is a worst-case statement over all update rules with unrestricted fan-in, and the authors are honest that it is a lower bound, not a typical-case one: random networks in their own simulations never violated it, but nothing here guarantees that a specific, nicely structured network needs anything close to n/2 sensors. Conversely, the single-node result is a best-case existence proof over adversarially convenient constructions; a generic threshold network will not be observable from one node. The practical truth for any given system lies between the bounds, and this paper deliberately does not chart that middle ground.

The single most load-bearing assumption is exact knowledge of the dynamics. Every reconstruction procedure here, the logarithmic-time decoder for the tight ReLU construction and the shift-register decoding for threshold networks alike, presumes the observer knows the update rules, the coupling structure, and the state domain. In any real substrate those are estimated, partially and noisily, and observability proofs are notoriously fragile to model mismatch: two states indistinguishable under the true dynamics can look distinguishable under a slightly wrong model, which is a false confidence, not a solution. The simulation section also tests sufficiency of small observation sets by searching for counterexamples within a bounded initial-state box (half-width 50 in each dimension) and a tolerance of 10^-2 over 40 time steps; the bound itself is proven, but the empirical claim that tested sets below n/2 always fail rests on that search finding counterexamples, which is evidence, not proof, of typical-case behavior. As a reviewer I accept the theorems as correct and well constructed, and treat the distance from these idealized network classes to measured biological tissue as the open question.

Observability limits for organoid readout design

The non-obvious implication for organoid intelligence is that the field's central instrumentation question, how many electrodes does a useful readout need, has a hidden dependence that almost nobody states: the answer is set by which dynamical regime the tissue occupies, not by tissue size. A neural culture whose effective units behave like rectified, continuous-valued elements with state-dependent dropout sit in the general real-valued ReLU class, and then the provable statement is uncomfortable: for worst-case recurrent wiring, every channel you do not record is a potential degree of freedom whose value no decoder can recover from the channels you do. Blind spots under this bound are not a signal-to-noise problem and cannot be fixed by longer recordings or better spike sorting; they are information-theoretically sealed. A 20,000-electrode array on a network of hundreds of thousands of relevant units may still be in the regime where provable full-state reconstruction is impossible in principle.

The opportunity is the regime dependency itself. The nonnegative case shows that when rectifying units stay active, the whole system reduces to a positive linear one and the classical observability toolbox applies, with one well-chosen channel sometimes sufficing. That gives an engineering lever no one is using deliberately: holding organoid tissue in a stimulation and modulatory regime where units remain above threshold converts the readout problem from a nonlinear nightmare into linear systems theory, where electrode placement is a solved design question and reconstruction is fast and principled. Conversely, if a culture's activity is dominated by sparse, all-or-none, burst-like events, it drifts toward the finite-state world where a single channel can, in the best case, carry the entire state through time multiplexing. The paper effectively hands organoid interfacing a design axis: tune the tissue's activation statistics and you move between one-sensor and half-the-network observability.

The threat is subtler and worth stating plainly. A system that is unobservable is also unauditable: if a computing substrate's internal state cannot in principle be reconstructed from its outputs, then claims about what computation it performed, whether it learned, and whether it is in a desired state rest on faith. For biological computing, where governance questions about monitoring and containment of living processors are already live, an n/2-style bound is a blueprint for substrates that resist inspection by construction. Anyone building organoid systems in the general real-valued regime with sparse electrodes is not just accepting noisy readout; they are accepting an unverifiable machine. The credible path runs through the regimes the paper proves are friendly: either keep the tissue active and linear-like, or embrace the binary regime and engineer for single-channel time-multiplexed observability. The middle, which is where most cultures sit by default, is exactly where the lower bound bites.

The bottom line

Established: global finite-horizon observability of binary linear-threshold networks can be achieved from a single observation node, with logarithmic-time reconstruction; binary-valued K-ReLU networks are equivalent to K-AND Boolean networks and inherit worst-case node requirements growing with network size; nonnegative ReLU networks reduce to classical linear observability; and general real-valued ReLU networks provably require at least n/2 observation nodes, a bound the authors' construction attains exactly. Hypothesis: these regime distinctions map onto organoid tissue, making the tissue's activation statistics, not the electrode count, the first-order determinant of whether a readout can in principle recover the network's state. What would confirm it: demonstrations on cultured networks that moving the same tissue between sparse-burst and tonic-active regimes changes how much of the state a fixed electrode set can reconstruct. What would break it: if biological recurrence is so structured, or so noisy in just the right way, that the worst-case bounds never bind, the theory would remain elegant and the practical guidance would reduce to folklore.