A Turbulence-Inspired Dynamical Systems Hypothesis of Ictogenesis - Nonlinear Amplification of Perturbations Near Critical States
ENGLISH
A Turbulence-Inspired Dynamical Systems Hypothesis of Ictogenesis
Nonlinear Amplification of Perturbations Near Critical States
Wanhong Huang · huangwanhong.g.official@gmail.com
Abstract
Seizure onset is conventionally modelled as a bifurcation: a slow variable drives the brain across a stability boundary, and the fast subsystem falls into a pathological attractor. This account predicts critical slowing before onset, yet empirical slowing signatures are inconsistent across patients. We propose an alternative framing borrowed from hydrodynamic transition to turbulence. In subcritical shear flows the laminar state remains linearly stable over a wide parameter range while becoming increasingly vulnerable to finite-amplitude perturbations of specific structure. We hypothesise that the interictal brain occupies an analogous regime: linearly stable, but subject to transient amplification that converts small structured perturbations into escape from the physiological basin. On this account, ictogenesis is not the crossing of a bifurcation point but the shrinking of a basin of attraction, and the relevant predictor is not slowing but the amplitude and structure of the minimal seizure-triggering perturbation. We derive four falsifiable consequences concerning perturbation structure, threshold scaling, episode statistics, and marker dissociation.
Keywords: ictogenesis; subcritical transition; non-normal transient growth; basin of attraction; critical slowing down; seizure prediction; dynamical systems.
§1 The Problem
The dominant dynamical account of seizure onset is bifurcation-theoretic. In the Epileptor model and its descendants, a slow permittivity variable drifts until the fast subsystem loses stability, typically through a saddle-node bifurcation at onset and a homoclinic bifurcation at offset. The framework is productive: it reproduces canonical ictal waveforms, supports virtual-brain parameterisation of the epileptogenic zone, and yields a clean prediction. If the system approaches a bifurcation, it should exhibit critical slowing down: lengthening autocorrelation, rising variance, slower recovery from probing stimuli.
That prediction has not cleanly survived contact with intracranial recordings. Slowing signatures are robust in some patients and absent in others; preictal windows vary by orders of magnitude; and seizure prediction algorithms built on variance-based markers generalise poorly across subjects. The standard responses are that noise obscures the signature, that the relevant slow variable is unobserved, or that patients differ in bifurcation class. Each is plausible. None explains why the failures are patterned rather than random.
We take the inconsistency as informative rather than as measurement noise, and ask what class of transition would produce onset without slowing.
§2 Existing Accounts
Three lineages bear on the hypothesis. First, bifurcation models of seizure dynamics, from early lumped-mass formulations through the Epileptor taxonomy of onset and offset bifurcations, which supply the vocabulary we adopt. Second, the self-organised criticality literature, which reads neural avalanche statistics as evidence that cortex operates near a critical point; this is a distinct sense of “critical” from the bifurcation sense and the two are frequently conflated, a conflation we avoid throughout by reserving critical state for proximity to a stability boundary. Third, the hydrodynamic stability literature on subcritical transition, in which pipe and Couette flow become turbulent despite the laminar solution remaining linearly stable at all Reynolds numbers; here the operative mechanism is non-normal transient growth, and the operative quantity is the threshold perturbation amplitude, which decays as a power law in Reynolds number. The present hypothesis is an attempt to import the third body of theory into the first.
§3 The Hypothesis
Statement. The interictal brain is a linearly stable state whose basin of attraction contracts as epileptogenic parameters drift. Seizure onset occurs when a perturbation of sufficient amplitude and appropriate structure is transiently amplified beyond the basin boundary. Ictogenesis is basin erosion, not bifurcation crossing.
Two mechanisms are separable here, and the separation matters.
Transient amplification is linear. The linearised dynamics around the interictal state are governed by a non-normal operator: the connectivity of cortex is neither symmetric nor normal, and excitatory–inhibitory circuits are the canonical example of non-normality in neuroscience. Non-normal operators permit large transient energy growth even when every eigenvalue has negative real part. A perturbation may therefore grow by orders of magnitude before decaying, without any eigenvalue approaching the imaginary axis. Nothing slows.
Escape is nonlinear. Transient growth alone returns the system to rest. Escape requires that the amplified perturbation reach a region where nonlinear terms redirect the trajectory into a competing basin. The nonlinearity does not amplify; it commits.
The two-step structure has an immediate consequence for the empirical puzzle of §1. Critical slowing indexes eigenvalue approach to marginal stability. If onset is driven by non-normal growth rather than eigenvalue crossing, slowing need not occur at all. Patients in whom slowing is robust and patients in whom it is absent may then be undergoing genuinely different transitions rather than the same transition observed with different noise floors.
The framing also reinterprets the slow epileptogenic process. Under the bifurcation account, slow drift moves a parameter toward a critical value. Under the present account, slow drift shrinks the basin, lowering the minimal perturbation energy required for escape, while the interictal state remains formally stable throughout. Structural changes following injury, synaptic reorganisation, gliosis, would then be read not as pushing the brain toward instability but as narrowing its margin against perturbation.
§4 Falsifiable Consequences
(i) Structure dependence of triggering. If transient growth is non-normal, the perturbations that trigger seizures are not the largest but those aligned with the optimal growth mode of the linearised operator. Prediction: stimulation patterns matched to the leading transient-growth direction of a patient-specific connectome will lower the seizure-triggering threshold substantially relative to energy-matched random or unstructured stimulation. This is directly testable in responsive-neurostimulation cohorts where stimulation parameters are already varied.
(ii) Threshold scaling. In pipe flow, the critical perturbation amplitude decays as a power law in Reynolds number. The analogue prediction is that the minimal triggering stimulation amplitude decays as a power law in whatever slow epileptogenic parameter is operative, rather than vanishing at a well-defined critical point. Repeated threshold-probing across a documented epileptogenic course would distinguish power-law decay from bifurcation-style collapse.
(iii) Memoryless episode statistics. Turbulent puffs in transitional pipe flow decay and split with exponentially distributed lifetimes; the process is memoryless. If ictal episodes are the neural analogue of localised turbulent structures, seizure durations within a patient should be closer to exponential than to peaked, and the hazard of termination should be approximately flat in time. Deviations from flatness would bound the analogy.
(iv) Dissociation of markers. The strongest test is joint. In patients where slowing markers are absent, threshold-probing should nevertheless reveal a declining escape threshold. Absence of slowing together with a falling threshold supports the subcritical account; absence of both suggests neither mechanism is operative and the onset is exogenously driven.
§5 Scope and Limits
This is a hypothesis about the class of transition, not a model. We have written down no equations and fitted no data, and the analogy to hydrodynamics is structural rather than physical: cortex is not a fluid, the relevant state space is not velocity fields, and no Reynolds number is claimed to exist. What transfers is a formal possibility that the bifurcation framing has left underexplored, that a system can be stable to every infinitesimal disturbance and still fail reliably, and that when it does, the observable precursors are properties of the basin rather than of the spectrum.
The hypothesis is most likely to fail in the following way. If patient-specific connectomes turn out to have low non-normality in the relevant frequency band, transient growth factors will be small, and the mechanism loses its force regardless of how well the analogy reads. That quantity is computable from existing diffusion and electrocorticographic data and should be established before the clinical predictions above are pursued.
Declaration of generative AI use. The author used Claude (Anthropic) in the preparation of this manuscript for structural discussion, terminological refinement, and drafting assistance. The hypothesis, its conceptual content, and all scientific claims are the author’s own. The author reviewed and edited all output and takes full responsibility for the content of the publication.
中文
一个受湍流启发的癫痫发生的动力系统假说
临界状态附近扰动的非线性放大
黄万宏 · huangwanhong.g.official@gmail.com
摘要
癫痫发作的起始通常被建模为一次分岔：一个慢变量驱动大脑越过一个稳定性边界，而快子系统落入一个病理性吸引子。这一说明预测起始之前的临界慢化，然而经验上的慢化特征在患者之间不一致。我们提出一个借自流体动力学向湍流转捩的替代定框。在亚临界剪切流中，层流状态在一个宽泛的参数范围内保持线性稳定，同时对特定结构的有限振幅扰动变得日益脆弱。我们假设发作间期的大脑占据一个类似的区制：线性稳定，但受制于把小的有结构扰动转化为从生理吸引域逃逸的瞬态放大。在这一说明上，癫痫发生不是一个分岔点的越过、而是一个吸引域的收缩，而相关的预测指标不是慢化、而是最小的引发发作之扰动的振幅与结构。我们推导出四个关于扰动结构、阈值标度、发作事件统计与标志物解离的可证伪后果。
关键词： 癫痫发生；亚临界转捩；非正规瞬态增长；吸引域；临界慢化；发作预测；动力系统。
§1 问题
关于癫痫发作起始的主导动力学说明是分岔理论的。在 Epileptor 模型及其后继者中，一个慢的电容率变量漂移，直到快子系统失去稳定性，典型地在起始处通过一个鞍结分岔、在终止处通过一个同宿分岔。这一框架是富有成效的：它复现出典型的发作波形、支持对致痫区的虚拟脑参数化，并得出一个干净的预测。倘若该系统趋近一个分岔，它应展现临界慢化：延长的自相关、上升的方差、从探测性刺激中更慢的恢复。
那一预测未曾干净地经受住同颅内记录的接触。慢化特征在一些患者中稳健、在另一些中缺席；发作前窗口相差数个数量级；而建立在基于方差之标志物上的发作预测算法在受试者之间泛化得很差。那些标准的回应是噪声遮蔽了那一特征、相关的慢变量未被观察，或患者在分岔类别上有别。每一个都可信。没有一个解释为什么这些失败是有模式的、而非随机的。
我们把这一不一致当作有信息量的、而非当作测量噪声，并问哪一类转捩会产生没有慢化的起始。
§2 现存的说明
三条谱系关乎这一假说。第一，癫痫发作动力学的分岔模型，从早期的集总质量表述到 Epileptor 关于起始与终止分岔的分类学，它们供出我们所采纳的语汇。第二，自组织临界性文献，它把神经雪崩统计读作皮层在一个临界点附近运作的证据；这是一个不同于分岔意义上的”临界”意义，而两者频繁地被混同，一种我们通篇借把临界状态保留给对一个稳定性边界的邻近而避免的混同。第三，关于亚临界转捩的流体动力学稳定性文献，其中管流与库埃特流变得湍流，尽管层流解在所有雷诺数下保持线性稳定；此处起作用的机制是非正规瞬态增长，而起作用的量是阈值扰动振幅，它随雷诺数以一个幂律衰减。当下的假说是一个把第三批理论导入第一批的尝试。
§3 假说
陈述。 发作间期的大脑是一个线性稳定的状态，其吸引域随致痫参数漂移而收缩。发作起始发生在一个具有足够振幅且恰当结构的扰动被瞬态地放大到超出吸引域边界之时。癫痫发生是吸引域侵蚀，而非分岔越过。
有两个机制在此可分离，而这一分离要紧。
瞬态放大是线性的。 围绕发作间期状态的线性化动力学由一个非正规算子支配：皮层的连接性既非对称也非正规，而兴奋–抑制回路是神经科学中非正规性的典型例子。非正规算子容许大的瞬态能量增长，即便每一个特征值都有负实部。因而一个扰动可能在衰减之前增长数个数量级，而没有任何特征值趋近虚轴。没有任何东西慢化。
逃逸是非线性的。 单靠瞬态增长使系统返回静息。逃逸要求那个被放大的扰动抵达一个区域，在其中非线性项把轨迹重定向进一个竞争性的吸引域。那非线性不放大；它作出承付。
这一两步结构对§1的经验难题有一个直接后果。临界慢化标示特征值向边缘稳定性的趋近。倘若起始由非正规增长、而非由特征值越过所驱动，慢化根本无需发生。慢化在其中稳健的患者与慢化在其中缺席的患者，于是可能正在经历真正不同的转捩、而非在不同噪声底限下被观察的同一转捩。
这一定框也重新诠释那个慢的致痫过程。在分岔说明之下，慢漂移把一个参数移向一个临界值。在当下的说明之下，慢漂移收缩那个吸引域，降低逃逸所需的最小扰动能量，而发作间期状态自始至终保持形式上稳定。损伤之后的结构改变，即突触重组、胶质增生，于是会被读作不是把大脑推向不稳定、而是收窄它对抗扰动的裕度。
§4 可证伪的后果
（一）引发的结构依赖性。 倘若瞬态增长是非正规的，那么引发发作的扰动不是最大的、而是那些同该线性化算子的最优增长模式对齐的。预测：与一个患者特异性连接组的领先瞬态增长方向相匹配的刺激模式，将相对于能量匹配的随机或无结构刺激大幅降低引发发作的阈值。这在刺激参数已然被改变的响应性神经刺激队列中是可直接检验的。
（二）阈值标度。 在管流中，临界扰动振幅随雷诺数以一个幂律衰减。类比的预测是最小的引发性刺激振幅随无论什么慢的致痫参数在起作用而以一个幂律衰减，而非在一个明确定义的临界点消失。跨一个有记录的致痫进程反复地探测阈值，会把幂律衰减同分岔式的坍缩区分开来。
（三）无记忆的发作事件统计。 转捩管流中的湍流团以指数分布的寿命衰减并分裂；这一过程是无记忆的。倘若发作事件是局域化湍流结构的神经类似物，那么一个患者之内的发作时长应更接近指数分布、而非峰值分布，而终止的风险率应在时间上近似平坦。对平坦性的偏离会给这一类比划界。
（四）标志物的解离。 最强的检验是联合的。在慢化标志物缺席的患者中，阈值探测应尽管如此揭示一个下降的逃逸阈值。慢化的缺席连同一个下降的阈值支持那个亚临界说明；两者都缺席则提示两个机制都不在起作用，而那个起始是外源地被驱动的。
§5 范围与界限
这是一个关于转捩之类别的假说、而非一个模型。我们未曾写下任何方程、未曾拟合任何数据，而向流体动力学的类比是结构性的、而非物理性的：皮层不是一种流体，相关的状态空间不是速度场，而没有雷诺数被声称存在。所迁移的，是一个分岔定框留下未充分探索的形式可能性，即一个系统能对每一个无穷小扰动都稳定、而仍可靠地失败，而当它如此时，那些可观察的前兆是那个吸引域、而非那个谱的诸性质。
这一假说最可能以下面的方式失败。倘若患者特异性连接组结果证明在相关频段中有低的非正规性，瞬态增长因子将是小的，而该机制失去它的力量，无论那类比读起来多么好。那个量从现存的扩散与皮层电图数据是可计算的，且应在上面的临床预测被追求之前被确立。
关于生成式 AI 使用的声明。 作者在本文稿的准备中使用了 Claude（Anthropic）进行结构性讨论、术语精炼与起草协助。该假说、它的概念内容，以及所有科学主张均为作者本人所有。作者审阅并编辑了所有输出，并对本出版物的内容承担全部责任。