Can Memory Replace Sensors? Observability and History Length in Fractional-Order Brain Networks

Picture estimating the activity of every region in a brain network from a few electrodes. Graph-based theory, which looks only at which regions influence which, gives a startlingly low answer. On a 64-channel EEG network modeled with fractional-order dynamics, it says two electrodes suffice. A fractional-order system is one whose present rate of change depends on its whole past through a slowly decaying power-law weighting. A careful identification study on the same kind of data instead needed 35 of 64 channels to recover the states and the hidden inputs. We argue that much of this gap comes from a variable neither line of work treats as a design choice: how much past signal the estimator integrates. In a fractional system, information about a distant, unmeasured region keeps arriving at a sensor for a long time instead of fading exponentially. A longer history window should therefore partly substitute for extra sensors. Practical fractional solvers, however, cut the history short to stay tractable, and that may silently remove the information the sensor count relied on.

We propose to measure this sensor-versus-memory trade-off and to build a cheap predictor of it from the network's graph. We then use that predictor as prior knowledge inside a cost-aware Bayesian optimizer, a sequential search that fits a probabilistic model of performance to past trials and picks the next trial to try. The optimizer searches jointly over which channels to measure and how long a history to keep. The output is a set of best-available trade-offs between channels, compute, and reconstruction error for unmeasured channels. It tests the claim that, for fractional networks, memory length and sensor count are partly exchangeable resources.

Core Claim

Structural observability asks whether the zero/non-zero pattern of the system's interaction matrix lets every hidden state reach a sensor, which makes the system reconstructible for almost all actual coupling values. It certifies reconstructibility, but says nothing about how strongly each hidden state's signal arrives. In a discrete fractional system $\Delta^\alpha x[k+1] = A x[k]$ observed through channels $S$ over a history window of $K$ steps, a hidden node's influence travels along the directed paths that the structural certificate identifies. That certificate is the matching and root-component cover from Minimum Number of Probes for Brain Dynamics Observability. Along those paths the influence is weighted by the power-law memory coefficients rather than by exponentially decaying ones.

We hypothesize three things. First, at a fixed noise level the error in reconstructing unmeasured channels falls with $K$ for longer than an integer-order model would predict, and the effect is stronger the further $\alpha$ is from 1. Second, a path-based "observability margin" can rank sensor sets by their noisy reconstruction error far better than structural rank alone. The margin is computed from the certificate, the estimated $\alpha$, and $K$, in time comparable to the matching itself. Third, giving this margin to a Bayesian optimizer finds sensor/window configurations that match the 35-channel identification quality with fewer channels or less compute than greedy rank-based selection.

The novelty is not tuning a fractional order with Bayesian optimization; that would be a minor engineering contribution. It is treating history length as a sensing resource, and connecting the structural certificate to numerical conditioning, which the file identifies as an open gap.

Prior Work and Its Role

Two papers motivate the gap by disagreeing. Minimum Number of Probes for Brain Dynamics Observability proves polynomial-time minimal dedicated-sensor placement for linear fractional systems by finding strongly connected components that need monitoring and solving a maximum-weight matching. It reports that two sensors (FC5, F8) suffice at 70–80% sparsity and seven or more at 10% connectivity. Dealing with Unknown Unknowns adds hidden inputs and estimates $A$, $\alpha$, and the inputs with an expectation-maximization procedure. It proves the observability objective is submodular and selects sensors greedily with a $(1-1/e)$ guarantee in $O(n^5K^3)$ time, finding 35 of 64 channels necessary and a 66% error reduction from modeling inputs. We reuse both as machinery and baselines. Neither treats $K$ as a variable to trade against sensors, and neither ranks rank-complete sensor sets by how well-conditioned they are.

A second group constrains the design. FROND, in Unleashing the Potential of Fractional Calculus in Graph Neural Networks with FROND, and the adjoint method in Efficient Training of Neural Fractional-Order Differential Equation via Adjoint Backpropagation show that fractional dynamics need a truncated "short memory" window to scale, because history kernels cost quadratic time. They choose the order and window by small manual grids. Their truncation is exactly the operation whose effect on observability we want to quantify. The EEG fitting study Re-thinking EEG-based non-invasive brain interfaces supports fractional models as better trackers of EEG than integer-order linear ones. It places sensors by anatomy, with no observability criterion.

A third group supplies the search method. Practical Bayesian Optimization of Machine Learning Algorithms introduces Gaussian-process search with Matérn 5/2 kernels, averaging over the model's own settings, and Expected Improvement per Second, which divides the expected gain of a trial by its predicted cost. We need the cost-aware version because evaluation cost grows with $K$. Grey-Box Bayesian Optimization for Sensor Placement in Assisted Living Environments shows that domain structure can be injected into the optimizer for placement problems, there for spatial coverage. We inject dynamical observability structure instead. Do Deep Convolutional Nets Really Need to be Deep and Convolutional? provides the methodological rule we follow: every compared method gets the same automated tuning budget.

Research Program

The first component is analysis. For linear fractional networks, we write the windowed observability matrix over $K$ steps using the Grünwald–Letnikov memory coefficients. These coefficients are the weights the discrete fractional operator places on past states; for orders away from 1 they decay as a power of the lag. From this matrix we derive the margin. For each unmeasured node, it is the attenuated signal energy delivered along its certificate path to a chosen sensor, accumulated over the window. The set's margin is the minimum over nodes. We will prove, or else empirically bound, how this margin relates to the smallest singular value of the windowed matrix. We also characterize the window length at which it saturates, as a function of $\alpha$ and path length.

The second component is the optimizer. The search space is the channel subset, the window $K$, and the sparsity weight on hidden inputs, which the hidden-input study tuned heuristically. Each evaluation runs the hidden-input identification on training trials, then reconstructs held-out unmeasured channels from the selected ones. The Gaussian-process model of performance takes the margin, the structural feasibility bit, and $K$ as input features alongside channel indicators. Sensor sets that are not even structurally observable are rejected by the free polynomial-time test before any expensive evaluation. Candidates are proposed by locally editing the incumbent sets with swaps. Cost is modeled jointly with error, since both depend on $K$. Treating them as independent is a stated limitation of the original cost-aware method.

The third component is extending the analysis to the nonlinear case. Using FROND-style fractional diffusion on graphs, we ask whether the margin still predicts which observed-node subsets allow the unobserved nodes' features to be reconstructed. We treat this as exploratory, because no structural guarantee exists once nonlinear activations are involved.

Baselines

The naive reference is reconstruction with random channel subsets of each size, using the full available history and default hidden-input regularization. It shows how much any selection principle buys. The purely structural baseline is the minimum dedicated-sensor placement of Minimum Number of Probes for Brain Dynamics Observability, run on the same estimated graph. We expect it to be rank-feasible but to reconstruct poorly, which would confirm the conditioning gap.

The strongest existing method is the greedy submodular selection with hidden-input identification of Dealing with Unknown Unknowns. Its reported results are 35 of 64 sensors needed and a more-than-66% reduction in relative prediction error from modeling hidden inputs on 109 PhysioNet subjects. It is the strongest competitor because it targets the same reconstruction question, on the same data, with a guarantee. We run it with its window $K$ and regularization tuned by the same Bayesian optimizer and budget we give ourselves, so that tuning effort is not a confound. Beating it means one of two outcomes on held-out subjects, with confidence intervals over subjects: equal reconstruction error with at least several fewer channels, or equal channels with lower error at no greater compute.

We add a greedy log-determinant selection on the estimated windowed observability matrix. It is a standard numerical sensor-design criterion, and the file reports no Lacuna result for it, but it targets conditioning directly and is the most dangerous simple competitor. For learned channel selection, we include the aggregation-selection of Graph Convolution Neural Network Based End-to-End Channel Selection. It reports 87.9% motor-imagery accuracy on BCI Competition IV 2a with 10 channels (92.2% with all) and 82.0% on PhysioNet.

Two search baselines get equal evaluation budgets and wall-clock limits: plain cost-aware Gaussian-process search without the margin feature, following Practical Bayesian Optimization of Machine Learning Algorithms, and random search. The key ablation recomputes the margin with $\alpha$ forced to 1, which gives integer-order, exponentially decaying memory. If this version performs as well, the fractional-memory ingredient is not doing the work.

Why not just…?

Why not just run greedy submodular selection and then grid-search the window? This is a real threat. The window is one integer, and a grid over it is cheap. The experiment settles it directly. If greedy selection plus a tuned window reaches the same point on the channel–error trade-off as our method, the optimizer contributes nothing. The surviving contributions would then be the measured sensor–memory substitution and the margin as a predictor, and we would report it that way. Our reason to expect a difference is that greedy rank-based selection stops once the rank is full, so it cannot prefer between two rank-complete sets that differ in conditioning, while reconstruction error is driven by conditioning.

Why not just select channels by greedy log-determinant of the estimated observability matrix? It optimizes conditioning without any graph structure, and it may match us. Its weakness is that it trusts the estimated $A$, and on EEG that estimate is biased by hidden inputs, as the 66% gap in the hidden-input study shows. Our optimizer scores held-out reconstruction directly, and the margin uses only the graph pattern, the order, and the window. The head-to-head comparison, and its sensitivity to deliberately corrupted estimates of $A$ on synthetic networks, will show which assumption holds.

Why not just learn channel selection end to end, as EEG-ARNN does? Its selection is judged by classification accuracy, which can stay high when unmeasured channels are unrecoverable. We report both accuracy and held-out-channel reconstruction for its 10-channel sets. If those sets also reconstruct well, structural reasoning adds little for this use case.

Why not just use an integer-order model with a longer window? The $\alpha=1$ ablation is exactly this test. A real risk is that PhysioNet orders were estimated at only 0.97–1.28 (Minimum Number of Probes), close to integer order, so memory effects may be weak on real EEG. That is why we also test synthetic networks with orders spread across $(0,1)$.

Proposed Experiments

Synthetic studies use fractional linear networks with known couplings built on the 34-region structural brain graph of Decentralized Observability with Limited Communication between Sensors. We vary $\alpha$, sparsity, noise, and hidden-input strength. Here we measure the full error surface over sensor count and $K$, and test whether the margin's rank correlation with true reconstruction error clearly exceeds that of structural rank and of the integer-order margin.

Real-data studies use PhysioNet EEG, splitting by subject, with error measured on reconstructing withheld channels. Downstream motor-imagery accuracy is measured with the selected channels on BCI Competition IV 2a, compared against the 87.9% at 10 channels from EEG-ARNN, and on BCI Competition III IVa/IVb, compared against the 87.6% and 85.7% of the fractional-feature classifier in Re-thinking EEG-based non-invasive brain interfaces. For graph learning, the exploratory nonlinear test uses Cora and Airport with FROND/Adj-F-GRAND backbones, against reported accuracies of 85.0% on Cora (Adj-F-GRAND-l) and 98.1% on Airport (F-GRAND-l). We mask node subsets and measure feature reconstruction as a function of the memory window.

Three outcomes are decisive. The first is whether the error surface shows substitution, meaning fewer channels are needed at equal error as $K$ grows, beyond what the $\alpha=1$ control shows. The second is whether the margin predicts held-out error well enough to beat structural rank as a selection signal. The third is whether margin-informed search reaches the greedy-submodular error level with fewer channels or less compute at equal search budget. If substitution appears only on synthetic data with small $\alpha$ and not on EEG, that is an informative negative result: real EEG near integer order would gain little from long fractional memory.