Correlated-noise-induced long lifetimes of BCM synapses: robustness of the spectral mechanism beyond the linearized, uniform-correlation, thermodynamic-limit analysis, and its mapping to monocular-deprivation experiments

Bazzani, Castellani and Cooper (Physical Review E 81, 051917, 2010) proposed that correlated input noise can preserve synaptic strength in Bienenstock–Cooper–Munro (BCM) plasticity: linearizing the BCM rule with a frozen sliding threshold, each eigenmode of the synaptic weight vector decays as $e^{-\lambda t}$ with $\lambda$ an eigenvalue of the input covariance matrix, and the eigenvalue distribution of a random covariance ensemble $C=C^{0}+mV$ (uniform correlation $q$ in $C^{0}$, Wigner-type fluctuation $V$) develops an accumulation of eigenvalues near zero once the rescaled fluctuation amplitude exceeds $m_{0}=1/2$ — producing very long-lived synaptic states and a candidate explanation for why retinal inactivation (which raises LGN noise correlations) causes less deprived-eye response loss than eyelid suture. The derivation, however, holds in a narrow regime: linearized dynamics with constant threshold, uniform correlations, the thermodynamic limit $N\to\infty$, neglected $O(m^{2})$ covariance terms, and weak symmetric static lateral connections with scalar inputs. The open problem is to determine whether the noise-induced long-lifetime mechanism and its critical threshold survive when these assumptions are lifted — the nonlinear regime with dynamical sliding threshold, finite networks with analytic finite-size corrections, structured (nonuniform) correlations, the $O(m^{2})$ terms, and general lateral coupling — and to map the model parameters onto measurable thalamic covariance statistics so that the mechanism becomes experimentally testable, including its relation to spike-timing-dependent plasticity as the authors explicitly flagged for future work. No published work is known to extend, refute, or experimentally test this spectral analysis, or to treat the general BCM–STDP relation in the correlated-noise setting, so the problem remains open.

Background

Raw Hebbian synaptic plasticity — strengthening synapses in proportion to the correlation between pre- and postsynaptic activity — is unstable, since weights grow without bound. The Bienenstock–Cooper–Munro (BCM) rule (Bienenstock, Cooper and Munro, 1982) stabilizes Hebbian learning for rate-coded neurons (neurons whose activity is described by a time-averaged firing rate rather than by individual spikes) by introducing a modification function $\phi(y,\theta_{m})$ of the postsynaptic activity $y$ with two zero crossings, at $y=0$ and at a sliding threshold $\theta_{m}$ that grows super-linearly with the time-averaged history of the cell's own activity (in the simplest quadratic form, $\phi(y)=y^{2}-y\,\theta_{m}$ with $\theta_{m}=\langle y^{2}\rangle$). Synapses potentiate (long-term potentiation, LTP) when activity exceeds $\theta_{m}$ and depress (long-term depression, LTD) below it; the moving threshold makes $y=\theta_{m}$ an attractive fixed point for stationary inputs. BCM was developed to describe experience-dependent development of visual cortex, in particular monocular deprivation (MD): closing one eyelid (lid suture) during the critical period (the early postnatal window of heightened experience-dependent plasticity) causes cortical neurons to lose responsiveness to the deprived eye — a shift in ocular dominance, the balance of cortical responsiveness between the two eyes — whereas pharmacologically silencing retinal output with tetrodotoxin (TTX, a voltage-gated sodium-channel blocker) produces much less loss — a classic discrepancy for plasticity theory. The cortical input arrives from the lateral geniculate nucleus (LGN), the thalamic relay between retina and visual cortex. Linden, Heynen, Haslinger and Bear (2009) recorded LGN activity in awake animals under normal vision, lid suture, and TTX, and found that inactivation does not reduce mean LGN firing but strongly increases correlations between LGN cells — undermining the standard assumption that the deprived eye simply delivers less input drive.

Blais, Cooper and Shouval (2009) showed by theory and simulation that the change in input correlations, not mean activity, can explain the MD protocols: LGN noise correlations significantly slow the loss of response to the deprived eye for plasticity rules that depress only the stimulated synapse (homosynaptic depression, as in BCM), while having the opposite effect for heterosynaptic rules of principal-component-analysis (PCA) type. Bazzani, Remondini, Intrator and Castellani (2003) had earlier treated uncorrelated (white) noise in BCM-type rules.

Bazzani, Castellani and Cooper (2010) — the source of this problem — gave the mechanism a spectral form. The deprived-eye input is modeled as noise $x(t)$ with covariance matrix $C$. Linearizing the BCM rule around $y\simeq 0$ with the threshold held constant gives $\dot{w}=-\epsilon\,C\,w$, where $\epsilon$ is the learning-rate constant (the magnitude of the slope of the linearized modification function $\phi\simeq-\epsilon y$), so in the eigenbasis of $C$ each weight component decays as $w_{\lambda}(t)=w_{\lambda}(0)\,e^{-\epsilon\lambda t}$ — the source paper writes this decay as $e^{-\lambda t}$, absorbing the learning rate into the time unit. Very small eigenvalues $\lambda$ of the covariance matrix mean very slow decay, i.e. very long lifetimes of synaptic strength. To determine when correlated noise produces such small eigenvalues, the authors model $C$ as a random positive-definite symmetric matrix $C=C^{0}+mV$ up to neglected $O(m^{2})$ terms, where $C^{0}$ is deterministic with unit diagonal and uniform off-diagonal correlation $q\in[0,1]$, and $V$ is a random symmetric fluctuation of amplitude $m$. Writing $C=(\sqrt{C^{0}}+mW)^{2}$ with $W$ a Wigner matrix — a symmetric random matrix with independent, identically distributed, zero-mean entries whose rescaled eigenvalues follow the semicircle law, the foundational result of random matrix theory (Wigner, 1958) — they compute the eigenvalue density of $C$ in the thermodynamic limit $N\to\infty$ at fixed rescaled fluctuation $m_{0}=m\sqrt{N/(1-q)}$. For $m_{0}<1/2$ the density is a shifted semicircle-type law supported on $(1-q)(1-2m_{0})^{2}<\lambda<(1-q)(1+2m_{0})^{2}$; for $m_{0}>1/2$ a second branch appears, the density becomes singular at $\lambda\to 0$, and eigenvalues accumulate near zero, with critical fluctuation variance $m^{2}_{\mathrm{crit}}=(1-q)/(4N)$. Physically: once covariance fluctuations exceed a threshold (at fixed dimensionality $N$ and correlation $q$), the ensemble possesses eigenvectors with extremely long decay times, so synapses aligned with them persist — a candidate explanation for why TTX (which raises LGN correlations) produces weaker deprivation effects than lid suture. The analysis is extended to a network of BCM neurons coupled by lateral connections $L$, assumed symmetric, time-independent, and weak ($\|L\|\ll 1$), with the external input reduced to one dimension (a scalar synapse) per neuron; the averaged decay dynamics is generated by a combined matrix $\hat{L}$ whose entries, in the source's notation, are $\hat{L}_{kl}=(I-L)^{-1}_{kl}C_{kl}$, and when $\hat{L}$ takes the same deformed-ensemble form the single-neuron method applies, so sufficiently large or fluctuating networks develop modes of extremely long lifetime. Finite networks ($N=10^{3}$) are checked against the thermodynamic laws by simulation.

The source analysis is thus valid in a well-defined but narrow regime: the plasticity dynamics is linearized around $y\simeq 0$, the sliding threshold is treated as constant (the slow-learning approximation — replacing the threshold by its expectation value $\theta_{m}\to\langle y^{2}\rangle$, justified when learning is slow compared with the input fluctuations — is invoked only as motivation), correlations are uniform across the input population, the covariance expansion drops $O(m^{2})$ terms, the eigenvalue distributions are computed at $N\to\infty$ with finite-size behavior left to simulation, and the network variant assumes static symmetric weak coupling with scalar synapses. The relation of these rate-based, time-averaged results to spike-timing-dependent plasticity (STDP — plasticity rules driven by the relative timing of pre- and postsynaptic action potentials rather than by firing rates) is deferred by the authors to future work. The paper closes by noting that the predicted long-lifetime states may be of physiological significance and can be tested experimentally.

Problem Statement

Determine whether the correlated-noise long-lifetime mechanism of Bazzani, Castellani and Cooper (2010) — the accumulation of covariance-matrix eigenvalues near zero above the critical rescaled fluctuation level $m_{0}=1/2$ (with $m_{0}=m\sqrt{N/(1-q)}$ the covariance-fluctuation amplitude $m$ rescaled by the input dimension $N$ and the uniform correlation $q$), and the consequent very slow decay $w_{\lambda}(t)=w_{\lambda}(0)e^{-\lambda t}$ of BCM synaptic weights (learning rate absorbed into the time unit) — is robust and experimentally testable when the assumptions of the source analysis are lifted. The single overarching objective is to establish whether correlated input noise is a genuine, robust, and biologically grounded mechanism for synaptic persistence in BCM plasticity (as required to explain the lid-suture-versus-retinal-inactivation discrepancy in monocular deprivation); the following requested results jointly serve that objective:

1. Nonlinear threshold-dynamics regime: analyze the full BCM rule $\dot{w}=x\,\phi(y,\theta_{m})$ (input $x$, postsynaptic activity $y$, synaptic weights $w$, modification function $\phi$) with the sliding threshold $\theta_{m}$ treated as a dynamical variable evolving with the neuron's activity history, beyond the linearization $\phi\simeq-\epsilon y$ around $y\simeq 0$ and beyond the slow-learning approximation (replacing the sliding threshold by its expectation value) — do the long-lived synaptic states and the persistence of small-eigenvalue modes survive the nonlinearity and the moving threshold?

2. Finite size and structured correlations: compute or characterize finite-$N$ corrections to the eigenvalue distributions (the analogues of the thermodynamic-limit laws for $m_{0}<1/2$ and $m_{0}>1/2$), and extend the analysis to heterogeneous networks with structured, nonuniform pairwise correlations beyond the uniform-$q$ baseline $C^{0}$; determine how the spectral transition at the critical fluctuation level is shifted, broadened, or destroyed in these settings, and assess whether the neglected $O(m^{2})$ terms change the qualitative picture.

3. Network generality: extend the network analysis beyond lateral connections $L$ that are symmetric, time-independent, and weak ($\|L\|\ll 1$), and beyond the reduction of each neuron's external input to a single scalar dimension.

4. Biological mapping and experimental test: map the model parameters — the correlation $q$, the fluctuation amplitude $m$, and the rescaled $m_{0}$ — onto experimentally measurable quantities (LGN covariance statistics recorded under lid suture versus retinal inactivation, TTX concentration, degree of deprivation), and derive quantitative, falsifiable predictions for how synaptic persistence or ocular-dominance dynamics should change with the measured noise-correlation level.

5. Relation to STDP: as the authors explicitly flag for future work, analyze how the rate-based BCM results under correlated noise relate to spike-timing-dependent plasticity, including whether spike-based plasticity rules preserve the same noise-correlation dependence of synaptic lifetimes.

A rigorous negative result on any one of components 1-5 resolves the task, since it already refutes the claim that the mechanism is robust under the lifted assumptions; the affirmative answer requires all five. Components 1-5 are scientifically linked facets of the same viability question: whether the mechanism survives the lifted assumptions and is testable against the deprivation experiments.

References

1. Armando Bazzani, Gastone C. Castellani, Leon N. Cooper, 'Eigenvalue distributions for a class of covariance matrices with application to Bienenstock-Cooper-Munro neurons under noisy conditions', Physical Review E 81, 051917 (2010). DOI: 10.1103/PhysRevE.81.051917, https://doi.org/10.1103/PhysRevE.81.051917 (open-access version: arXiv:1001.4708, https://arxiv.org/abs/1001.4708)
2. E. L. Bienenstock, L. N. Cooper, P. W. Munro, 'Theory for the development of neuron selectivity: orientation specificity and binocular interaction in visual cortex', The Journal of Neuroscience 2, 32-48 (1982). DOI: 10.1523/JNEUROSCI.02-01-00032.1982, https://doi.org/10.1523/JNEUROSCI.02-01-00032.1982
3. Monica L. Linden, Arnold J. Heynen, Robert H. Haslinger, Mark F. Bear, 'Thalamic activity that drives visual cortical plasticity', Nature Neuroscience 12, 390-392 (2009). DOI: 10.1038/nn.2284, https://doi.org/10.1038/nn.2284
4. Brian S. Blais, Leon N. Cooper, Harel Z. Shouval, 'Effect of correlated lateral geniculate nucleus firing rates on predictions for monocular eye closure versus monocular retinal inactivation', Physical Review E 80, 061915 (2009). DOI: 10.1103/PhysRevE.80.061915, https://doi.org/10.1103/PhysRevE.80.061915
5. A. Bazzani, D. Remondini, N. Intrator, G. C. Castellani, 'The Effect of Noise on a Class of Energy-Based Learning Rules', Neural Computation 15, 1621-1640 (2003). DOI: 10.1162/089976603321891837, https://doi.org/10.1162/089976603321891837
6. Eugene P. Wigner, 'On the Distribution of the Roots of Certain Symmetric Matrices', The Annals of Mathematics 67, 325-327 (1958). DOI: 10.2307/1970008, https://doi.org/10.2307/1970008
7. Leon N. Cooper, 'STDP: spiking, timing, rates and beyond', Frontiers in Synaptic Neuroscience 2, 14 (2010). DOI: 10.3389/fnsyn.2010.00014, https://doi.org/10.3389/fnsyn.2010.00014
8. Eugene M. Izhikevich, Niraj S. Desai, 'Relating STDP to BCM', Neural Computation 15, 1511-1523 (2003). DOI: 10.1162/089976603321891783, https://doi.org/10.1162/089976603321891783
9. Daniel Bush, Andrew Philippides, Phil Husbands, Michael O'Shea, 'Reconciling the STDP and BCM Models of Synaptic Plasticity in a Spiking Recurrent Neural Network', Neural Computation 22, 2059-2085 (2010). DOI: 10.1162/neco_a_00003-bush, https://doi.org/10.1162/neco_a_00003-bush
10. Leon N. Cooper, Mark F. Bear, 'The BCM theory of synapse modification at 30: interaction of theory with experiment', Nature Reviews Neuroscience 13, 798-810 (2012). DOI: 10.1038/nrn3353, https://doi.org/10.1038/nrn3353
11. Carlo Mengucci, Daniel Remondini, Gastone Castellani, Enrico Giampieri, 'WISDoM: Characterizing Neurological Time Series With the Wishart Distribution', Frontiers in Neuroinformatics 14, 611762 (2021). DOI: 10.3389/fninf.2020.611762, https://doi.org/10.3389/fninf.2020.611762
12. Xie Xiao-Ping, Chen Hong-Ping, Cao Zhi-Tong, He Guo-Guang, 'The normalized KLD coefficient and its application in detection of correlation and synchronization from multivariable series', Acta Physica Sinica 61, 130505 (2012). DOI: 10.7498/aps.61.130505, https://doi.org/10.7498/aps.61.130505
13. Gastone Castellani, Leon N. Cooper, Luciana Renata De Oliveira, Brian S. Blais, 'Energy Consumption and Entropy Production in a Stochastic Formulation of BCM Learning', Journal of Computational Biology 28, 257-268 (2021). DOI: 10.1089/cmb.2020.0118, https://doi.org/10.1089/cmb.2020.0118
14. Leonidas M. A. Richter, Julijana Gjorgjieva, 'A circuit mechanism for independent modulation of excitatory and inhibitory firing rates after sensory deprivation', Proceedings of the National Academy of Sciences 119, e2116895119 (2022). DOI: 10.1073/pnas.2116895119, https://doi.org/10.1073/pnas.2116895119
15. François X. P. Bourassa, Paul François, Gautam Reddy, Massimo Vergassola, 'Manifold learning for olfactory habituation to strongly fluctuating backgrounds', bioRxiv 2025.05.26.656161 (2025). DOI: 10.1101/2025.05.26.656161, https://doi.org/10.1101/2025.05.26.656161
16. Carlos Stein Naves de Brito, Wulfram Gerstner, 'Learning what matters: Synaptic plasticity with invariance to second-order input correlations', PLOS Computational Biology 20, e1011844 (2024). DOI: 10.1371/journal.pcbi.1011844, https://doi.org/10.1371/journal.pcbi.1011844
17. Kevin R. Duffy, Nathan A. Crowder, Arnold J. Heynen, Mark F. Bear, 'Comparative analysis of structural modifications induced by monocular retinal inactivation and monocular deprivation in the developing cat lateral geniculate nucleus', Journal of Comparative Neurology 531, 1244-1260 (2023). DOI: 10.1002/cne.25493, https://doi.org/10.1002/cne.25493