# Does Axonal Delay Dispersion Determine What a Neuron Computes?

A single measurable anatomical parameter — the dispersion of axonal conduction delays converging on a dendritic branch — determines whether a cortical neuron functions as an event detector or an order-selective sequence detector, and the same parameter predicts cortical column diameter with a timing tolerance of approximately one millisecond. That is the central claim of a computational study published today on arXiv (2609.04195) by Cheng Bi and Jipeng Sun.

For engineers building intracortical [brain-computer interfaces](https://bciintel.com/glossary/brain-computer-interface) and the neuroscientists informing their decoder designs, this matters immediately: if temporal coding is structured by anatomy in the way Bi and Sun propose, then the myelination profile of the cortical region being recorded from — not just firing rate — constrains what information is theoretically recoverable from a neural population, and how spike-sorting algorithms should be tuned to capture it.

**The core result in plain terms:** narrow delay dispersion across a neuron's inputs produces event detectors that respond to coincident volleys of spikes. Wide dispersion produces sequence detectors that respond only when spikes arrive in a specific temporal order. The transition between these regimes is not a design choice — the authors report it emerges from random delays and connectivity, with the crossover tracking the inter-event interval at a slope statistically indistinguishable from one.

---

## The Delay-Signature Framework

Bi and Sun propose what they call a "delay-signature framework." The axonal conduction delays arriving at a single dendritic branch act as a physical key: only input sequences whose spike-time differences are compensated by those delays arrive synchronously at the dendrite. Coincidence detection — implemented biophysically via calcium plateau thresholds — then converts that synchrony into an all-or-none output.

This framing reframes myelination in functionally significant terms. The authors argue myelination is not merely a regulator of conduction speed, but functions as a computational switch: heavily myelinated projections compress delay variance into the narrow-dispersion regime (event detection), while unmyelinated or lightly myelinated projections operate in the wide-dispersion regime (sequence detection). The distinction maps directly onto known anatomical differences between cortical areas and layers.

For BCI hardware teams, this is not an abstract point. [Electrode arrays](https://bciintel.com/glossary/electrode-array) implanted in motor cortex — a region with a well-characterized mix of myelinated corticospinal projections and local unmyelinated interneurons — are sampling from a heterogeneous population where different neurons are performing qualitatively different computations. A decoder optimized purely for rate coding or mean firing rate will be leaving structured temporal information on the table.

---

## The Millisecond Timing Tolerance and Its Implications for Decoding

The second result from the simulations is equally specific: the delay-dispersion parameter sets an absolute timing tolerance of approximately one millisecond, with the authors noting that slowing (increased delay) is better tolerated than speeding (decreased delay). This asymmetry has a practical analogue in stimulation: intracortical microstimulation (ICMS) delivered slightly late to a sequence-detecting neuron is less disruptive than stimulation that arrives early relative to the expected volley.

The approximately one-millisecond window also has direct implications for spike-sorting in intracortical recordings. Current commercial and research-grade spike-sorting pipelines — used by groups running BrainGate-derived protocols and by implant developers including [Blackrock Neurotech](https://bciintel.com/companies/blackrock-neurotech) — typically apply refractory period constraints and waveform clustering at sub-millisecond precision. If sequence-detecting neurons are firing in response to ordered input patterns at millisecond-scale intervals, then clustering approaches that collapse temporally close spikes into single units may be systematically misattributing sequence-detection events as noise or multi-unit activity.

This is a hypothesis, not an experimentally confirmed finding — and that distinction matters. The Bi and Sun paper is a computational modeling study. The integrator-neuron simulations they report are compelling mechanistic arguments, but they have not been validated against simultaneous intracortical recordings with known myelination profiles in the same tissue.

---

## The Cortical Column Diameter Prediction

The third result is the most testable and arguably the most provocative: combining the approximately one-millisecond timing tolerance with horizontal conduction velocity produces a predicted cortical column diameter. The authors state that two cortical areas with direct empirical measurements fall where the relation places them, but the paper does not specify which areas or what those measured diameters are beyond this description. Independent replication with a broader anatomical dataset is the obvious next step, and it is the kind of falsifiable prediction that distinguishes a useful theoretical framework from a post-hoc rationalization.

If the prediction holds across a wider range of areas, it would offer BCI hardware designers a new tool: anatomically informed estimates of the spatial scale at which electrode contacts should be spaced to capture computationally distinct populations. For high-density ECoG arrays — the technology [Synchron](https://bciintel.com/companies/synchron) and [Precision Neuroscience](https://bciintel.com/companies/precision-neuroscience) are each advancing through different implant strategies — knowing the functional column diameter relevant to a specific cortical region could inform contact pitch in ways that current empirical trial-and-error does not.

---

## What This Means for BCI Decoder Design

The broader implication for the field is a reinforcement of temporal coding as a first-class signal modality, not a secondary consideration. Current intracortical BCIs achieving the highest decoded bits per second — whether in speech decoding or cursor control — rely primarily on firing rate modulation over relatively long integration windows. The Bi and Sun framework suggests there is a parallel channel of information encoded in the precise ordering and timing of spikes across neurons, structured by axonal anatomy, that existing decoders are not systematically exploiting.

Whether that additional channel is practically accessible given electrode count, signal-to-noise constraints, and computational overhead at the implant level is a separate engineering question. But the theoretical case that it exists and is anatomically determined — rather than random or epiphenomenal — is now more formally specified than it was yesterday.

For closed-loop BCI systems that also deliver [bidirectional](https://bciintel.com/glossary/bidirectional-bci) somatosensory feedback via ICMS, the asymmetric tolerance result (slowing tolerated better than speeding) may have direct relevance to stimulation parameter optimization. If the receiving neurons are sequence detectors with delay-signature tuning, then stimulation timing errors in one direction disrupt the computation more than errors in the other — a constraint that closed-loop timing architectures should account for.

---

## Skeptical Assessment

Several caveats apply before this framework influences device development decisions:

1. **Simulation-only evidence.** The integrator-neuron model captures important biophysical features but is not a biophysically detailed compartmental model. The calcium plateau threshold mechanism is plausible but the exact parameters governing the narrow-to-wide dispersion transition may not transfer directly to in vivo recordings.

2. **The column diameter validation is thin.** Two data points do not establish a predictive relationship. This is the claim most in need of broader empirical support before it should influence array design.

3. **No identified company connection.** The arXiv metadata flagged a Synchron association in our automated relevance scoring, but the paper itself carries no corporate affiliation — the authors are listed as Cheng Bi and Jipeng Sun without institutional details in the abstract. The BCI relevance is real; the company connection appears to be a false positive from the relevance filter.

4. **Translation timeline.** Even if the framework is fully validated, integration into spike-sorting pipelines or closed-loop stimulation controllers is a multi-year engineering effort, not a near-term deliverable.

---

## Key Takeaways

- A single anatomical parameter — axonal delay dispersion — determines whether a cortical neuron detects coincident events or temporally ordered sequences, according to computational simulations by Bi and Sun (arXiv:2609.04195).
- The framework maps myelination onto a computational switch, not merely a speed regulator: narrow dispersion (myelinated) → event detection; wide dispersion (unmyelinated) → sequence detection.
- Timing tolerance is approximately one millisecond, with slowing better tolerated than speeding — a distinction relevant to ICMS-based somatosensory feedback in closed-loop BCIs.
- The same dispersion parameter predicts cortical column diameter; the authors report two empirical measurements align with the prediction, but broader validation is needed.
- Current intracortical BCI decoders optimized for rate coding may be systematically missing structured temporal information encoded in spike order.
- This is a modeling paper — no in vivo electrophysiology data is presented, and clinical translation implications are indirect but real.

---

## Frequently Asked Questions

**What is axonal delay dispersion and why does it matter for BCIs?**
Axonal delay dispersion refers to the spread of conduction delays across the axons converging on a single dendritic branch. According to Bi and Sun's modeling work, this parameter determines whether the target neuron functions as an event detector (coincident spikes) or a sequence detector (ordered spikes). For BCI decoder design, it implies that the myelination profile of a recorded cortical region shapes what temporal information is theoretically recoverable from that population.

**How does the ~1ms timing tolerance affect spike sorting?**
Spike-sorting pipelines that collapse temporally close events — particularly within approximately one millisecond — may misclassify sequence-detection firing patterns as noise or multi-unit activity. The Bi and Sun framework predicts this window is an intrinsic property of the delay-signature mechanism, not an artifact, suggesting that sub-millisecond temporal resolution should be preserved rather than smoothed in populations suspected of sequence detection.

**Does this research change how ICMS stimulation parameters should be set?**
Potentially, yes, for closed-loop systems delivering somatosensory feedback. The model predicts an asymmetric tolerance: neurons tolerate stimulation that arrives slightly late better than stimulation that arrives slightly early relative to their delay-signature tuning. This asymmetry has not been validated in vivo, but it offers a testable hypothesis for groups optimizing ICMS timing in bidirectional BCI trials.

**Which cortical areas are most relevant to BCI and most affected by this framework?**
Motor cortex and somatosensory cortex — the primary targets of current intracortical BCI implants — contain heterogeneous mixtures of myelinated and unmyelinated projections. The framework predicts these areas host both event-detecting and sequence-detecting populations simultaneously. Speech-decoding BCIs targeting areas such as Broca's area may encounter even larger proportions of sequence detectors given the temporal structure of phoneme sequences.

**Is this finding clinically applicable now?**
No. This is a computational modeling study with simulation-based results. Clinical applicability depends on in vivo electrophysiological validation, integration into spike-sorting and decoding software, and ultimately device-level implementation — a pathway measured in years, not months.