http://orcid.org/0000-0002-0580-0598
Figures
Abstract
Across scientific disciplines, thresholded pairwise measures of statistical dependence between time series are taken as proxies for the interactions between the dynamical units of a network. Yet such correlation measures often fail to reflect the underlying physical interactions accurately. Here we systematically study the problem of reconstructing direct physical interaction networks from thresholding correlations. We explicate how local common cause and relay structures, heterogeneous in-degrees and non-local structural properties of the network generally hinder reconstructibility. However, in the limit of weak coupling strengths we prove that stationary systems with dynamics close to a given operating point transition to universal reconstructiblity across all network topologies.
Citation: Lünsmann BJ, Kirst C, Timme M (2017) Transition to reconstructibility in weakly coupled networks. PLoS ONE 12(10): e0186624. https://doi.org/10.1371/journal.pone.0186624
Editor: Michal Zochowski, University of Michigan, UNITED STATES
Received: April 28, 2017; Accepted: October 4, 2017; Published: October 20, 2017
This is an open access article, free of all copyright, and may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose. The work is made available under the Creative Commons CC0 public domain dedication.
Data Availability: All relevant data are within the paper and its Supporting Information files.
Funding: This work was supported by the Max Planck Society (BL, MT), the Germany Ministry for Education and Research (BMBF) under grant no. 01GQ1005B (CK, MT), and an independent research fellowship by the Rockefeller University, New York, USA (CK).
Competing interests: The authors have declared that no competing interests exist.
Introduction
Complex networked systems generate dynamics and thus functions that fundamentally depend on how their units interact [1–3]. As a consequence, knowing the interaction topology of such systems is a key towards understanding them [4–12]. Yet, direct access to the topology of physical interactions is largely limited for many natural systems and across scales, ranging from metabolic and gene regulatory networks on the subcellular level to neural circuits of millions of cells, to food webs among organisms and planetary climate networks [10, 13–21]. Thus, measures of pairwise statistical dependencies between time series of the dynamics of their units are often employed as proxies for physical interactions [15–17, 21–27]. Assuming sufficiently many and sufficiently accurate data, each such method provides useful information about how the considered statistical dependency measures vary across pairs of units. The value of such a statistical measure, thresholded as desired, e.g. for significance against coincident correlations, may be taken to quantify the interactions among these units. Yet, such measures themselves do not necessarily provide immediate insights into how the units are directly influencing each other via physical interactions. In particular, what do correlations generally tell us about direct physical interactions in network dynamical systems? And is it possible to detect direct physical interactions among units by thresholding these measures to reconstruct the topology of the network?
Here, we systematically address this question on a conceptual level and identify limits of network reconstructibility based on thresholding pairwise measures of statistical dependence. In general, non-linearities of intrinsic and coupling dynamics, correlated noise sources, heterogeneities in time scales and coupling strengths as well as nontrivial network topology jointly create complex statistical correlation patterns. To reveal principal limits of reconstructibility originating from network interactions (toplogy and strength), we here focus on systems with dynamics around a given operating point. More specifically, we analyze the idealized setting of linearly coupled systems with homogeneous dynamical parameters receiving independent additive noise inputs and evaluate network reconstruction from thresholding linear correlations obtained from sufficiently long time series. Reconstruction of physical interactions generally is at least as hard in any more complex setting, e.g., involving non-linear dynamics and adequate measures of statistical dependence such as mutual information. We explicate limits of reconstructibility due to local common cause structures, local relay structures, topological in-degree heterogeneities as well as non-local structural elements. Despite these limitations our analysis interestingly also reveals that, stationary systems close to operating points exhibit a transition to universal reconstructibility for sufficiently weak coupling, independent of the interaction topology.
Model and methods
Consider the dynamics
(1)
of network dynamical systems characterized by variables x = (x1, …, xN) that interact diffusively with generic coupling strength α > 0 on a network topology given by an adjacency matrix A. The units are driven by independent white noise ηi(t) of strength γ and relax on a time scale τgl > 0. The entries of the weighted adjacency matrix are Aij > 0 if unit j physically acts on i, with all other elements, including the diagonal being Aij = 0. Without loss of generality, we rescale time such that τgl = 1.
The diffusive coupling considered here emerges in approximations of coupled oscillator networks [28, 29], in population dynamics [30, 31] and in stochastic processes as, e.g., epidemic models [32].
Other types of linear coupling (e.g., in [33]) can to the same extent be treated using diffusive coupling if individual self-coupling terms are introduced (see S1 Supplementary Material). This study thus also covers networks of the form . For the detailed analysis of factors that hinder reconstructibility we omit individual self-coupling terms to avoid unequal scaling of correlations to establish ideal conditions for correlation thresholding.
The dynamics generated by (1) characterizes linear systems as well as stationary systems sufficiently close to given operating points.
Can we infer the physical topology from optimally thresholding the matrix C of pairwise correlations (Fig 1)? The covariance matrix σ defined by the elements
(2)
computed using an unbiased time-average 〈⋅〉, yields the correlations
(3)
by normalization.
Reconstructing interaction networks from correlation thresholding may or may not yield correct connectivity pattern. (a)-(d) Successful reconstruction of a network (N = 15, average degree , γ = 1, α = 2, Aij ∈ {0, 1} for absent and present interactions, resp.). (e)-(h) Reconstruction of statistically similar network is unsuccessful for any threshold. (a,e) Adjacency matrix of original network (black indicated directed interaction, gray undirected network aimed for). (b,f) Dynamics of the units yielding (c,g) correlation matrices. Thresholding yields (d) correct or (h) incorrect reconstruction, depending on the exact topology.
Reconstructing the physical topology implies detecting non-zero elements in the coupling matrix A. Also, as correlation matrices are symmetric by construction, Cij = Cji, we relax the problem to the reconstruction of the undirected representation of the physical interaction network. Thus, we aim for the correct reconstruction of the matrix A′ the elements of which are given by
(4)
Correlations (3) may be thresholded using a (possibly optimized) threshold θ to yield an estimate
with elements
if Cij > θ and
otherwise. Below we focus on the question whether there is any threshold of the correlation matrix (3) that yields a correct estimate of A′. If there is no such threshold, we call the network non-reconstructible (in this sense).
The theory of Ornstein-Uhlenbeck processes [34] yields an analytical expression for the covariance matrix
(5)
Here, the matrix J is given by its elements
(6)
The integral (5) can be used to compute the covariance matrix σ of specific network topologies with special symmetries (see Fig 2 and S1 Supplementary Material). However, numerical computation of (5) for a random network is computationally not practical.
(Unspecified parameters as in Fig 1) (a,b) Regions of reconstructible (shaded gray) and non-reconstructible networks (white shading) are non-linearly separated for (a) common cause structures and (b) relay structures (Regimes computed by interpolating analytic results using (3) and (5), details see S1 Supplementary Material). (c) Non-local effect renders larger networks non-reconstructible: Each circle would be reconstructible alone, but the joint network is not (α = 2).
Partial integration of (5) yields the Lyapunov equation
(7)
which we solve numerically [35] to obtain the covariance matrix σ for explicit network topologies (α, γ, A). Again, (7) can be solved analytically if needed [36]. Via the relation (7) and (3), we thus semi-analytically obtain all the real-valued elements Cij of the correlation matrix without any sampling error.
We order those to determine whether there is a threshold θ separating all existing from all non-existing links.
Results
Topology-induced limits of reconstructibility
Even under these idealized conditions, physical interactions are in general not reconstructible from thresholding the correlation matrix C. Whereas some topologies can be reconstructed via a threshold that separates existing from absent links (Fig 1a–1d), many attempted reconstructions yield false positive and false negative predictions of links, independent of the threshold (Fig 1e and 1f) and are thus intrinsically non-reconstructible by correlation thresholding.
Topologically induced errors and ultimately the limits in reconstructibility can be of local or of non-local nature (Fig 2): For instance, common input might cause unconnected units to be more correlated than connected units, a dilemma known as the common cause effect (Fig 2a inset). Likewise, two units may be strongly correlated if the network provides connectivity between them across a set of intermediate units, thereby forming a relay structure (Fig 2b inset). For both settings, reconstructibility non-linearly depends on a combination of overall coupling strength and the number of interfering units in a systematic way (Fig 2a and 2b, main panels).
In larger networks with diameter d ≥ 3, additional non-local effects limit reconstructibility (illustrated in Fig 2c). Differences in the correlation strength may, for instance, be caused by different link densities in different parts of the network, and imply incorrect link classification.
Universal transition to reconstructibility
The coupling strength α controls the impact of both, local and non-local influences on reconstructibility. For instance, analytic treatment of a small common cause structure (Fig 3) reveals that the system becomes reconstructible for all sufficiently small coupling strengths α while it is non-reconstructibility if α is too large. This systematic transition prevails for any number of common input units in common cause structures as well as for any number of relay units in relay structures (See S1 Supplementary Material for detailed derivations).
(a) Correlation thresholding yields different estimators (shaded areas with graphs as insets) for a given topology (adjacency matrix on bottom right) depending on coupling strength and threshold. For sufficiently small coupling strength α (left of gray dashed line), there are ideal thresholds yielding perfect reconstruction (green shading). (Analytic results obtained using (3) and (5).) (b) Fraction of reconstructible networks exhibits transition to full reconstructibility at positive coupling strength α (inset) and αN (main panel), illustrated for random networks of N ∈ {16, 64, 256} units and link probability p = 0.5. Every arbitrary network exhibits such a transition individually (see text).
Interestingly, all topology-induced limits disappear for sufficiently weak coupling, as seen from the following analytic argument: Rewriting the matrix
(8)
in terms of the graph Laplacian L with elements
(9)
(where δij = 1 if i = j and zero otherwise is the Kronecker-delta) and expanding (5) for α ≪ 1 yields
(10)
The term σ(1) on the r.h.s. of (10) does only contribute to entries σij that reflect existing links because otherwise Lij = Aij = 0. Thus, the covariance of coupled units scales linearly with α whereas for uncoupled units it scales quadratically. So for sufficiently small coupling strength α, covariances of coupled units will be larger than those of uncoupled units. This result transfers to the elements of the correlation matrix C in (3) because diagonal elements of the covariance matrix σ are of order
(11)
Hence, every network topology is reconstructible for sufficiently small coupling strengths.
Our analysis reveals, that the expansion of the covariance matrix σ in the coupling strength α is an effective separation into contributions of paths through the network with increasing lengths. The nth summand in the series represents the contribution of paths up to length n.
Similar results have been obtained in mean field models of spiking neuronal networks if the covariance matrix is self-consistently expanded in the mean neuronal input [37]. The two expansions of both approaches however differ in the order of the terms. Only expansions in network coupling strength ensure full reconstructibility of the network connectivity in the weak coupling limit.
As shown in the supplementary material, this transition to reconstructibility in the weak coupling limit is not limited to the considered network model and can essentially be found in all generic linear networks (for more details see S1 Supplementary Material).
Illustrative example of reconstructibility transition
Furthermore, specific families of networks with homogeneous connectivity are reconstructible via correlation thresholding for all coupling strengths, weak and strong. As we demonstrate for illustration, this is the case for directed ring like topologies with neighbors. In these networks the correlation matrix C is strictly proportional to the covariance matrix σ so that it is sufficient to show reconstructibility with respect to the covariance matrix. Also, since the covariance matrix σ is a circulant, it is sufficient to show reconstructibility only for the connections of one unit. The reconstructibility conditions is identical for all units. For simplicity of presentation, we take the number N of units to be even.
We order the units in such a way that it reflects the network topology, i.e.
(12)
and replace
in (7) to obtain
(13)
as a self-consistency equation for the entries σij of the covariance matrix σ.
Here, the index i indicates the number of the unit in the circle and the integer n refers to the distance from the diagonal n = j − i. This nomenclature reflects the symmetry of the circulant matrix σ. All indices have been taken modulo the number of units N for simplicity.
Transforming this equation into Fourier space yields
(14)
with solution
(15)
in Fourier coordinates. An inverse Fourier transformation yields the analytic solution
(16)
where the sequences
are repeated convolutions of the step sequence
(17)
i.e.,
(18)
For more detailed derivations, please see the supplementary material.
Since the sequences are monotonically decreasing in the interval n ∈ [−N/2, N/2] covariance only decreases with distance in the circular graph. Because for any given unit i, connected units are closer than non-connected units, for every such network with
-regular topology, a threshold exists that separates existing from absent links, making these networks reconstructible for arbitrary coupling strengths, for any network size N and for any number of neighbors
. For
the undirected representation of the network is fully connected and reconstruction is trivial.
Which heterogeneities hinder reconstruction?
Given the insights from the ring-like networks, we hypothesized that if topological irregularities increase, they decrease and ultimately hinder network reconstructibility. To analyze the overall impact of topology on reconstruction quality, we investigated ensembles of directed networks in the regime between regular and random, employing a modified Watts-Strogatz small world model [38]: Starting with a regular ring of N units with each unit receiving directed links from preceding nodes (and thus a mean in- and out-degree of
) the source and the target of each link are detached with probability qout and probability qin respectively. The resulting loose ends are randomly redistributed in the network while avoiding self-loops and multiple links. This creates networks of mean degree
whose in-degree distribution
and out-degree distribution
are altered separately from their original values
by varying qin and qout. This random graph ensemble contains networks with unimodal degree distributions (binomial for qin = qout = 1,
and 1 ≪ N) so that the variances of the distributions serve as indicators for the inhomogeneities in the network.
Considering a fixed coupling strength (e.g., α = 1), we quantify reconstructibility by measuring the AUC, the area under the ROC (receiver operating characteristic) curve, generated by a variable correlation threshold θ. AUC ranges from AUC = 0.5 for random guessing to AUC = 1 for perfect reconstructibility (see S1 Supplementary Material for an introduction to ROC curves). For networks that are not densely connected (), we find that reconstruction quality systematically decreases with in-degree heterogeneity, with the AUC exhibiting a functional dependency on the variance of the in-degree distribution, yet is almost independent of the variance of the out-degree distribution (compare Fig 4a with Fig 4b). Thus, the reconstruction error is mainly explained by the in-degree heterogeneity. We obtain qualitatively similar results across different average connectivities
(inset of Fig 4a).
(a) AUC exhibits functional dependency on the variance of the in-degree distribution varin, regardless of the variance of the out-degree varout. Inset: Qualitative behavior is the same for differnet mean degrees. (b) No significant dependency of reconstruction quality on out-degree heterogeneity (network size N = 150 throughout, α = 1, Aij ∈ {0, 1}).
This finding coincides with intuition: Since the influence of a source unit on its target decreases with the number of additional signals its target receives (see common cause structure, S1 Supplementary Material), large differences in the in-degree directly correspond to high variability in pairwise correlations. Correlations in complex networks with inhomogeneous in-degree thus strongly depend on the local link density and can therefore not be faithfully reconstructed using a global threshold. In contrast, the number of outgoing connections does not directly impact pairwise correlations.
Conclusions
In summary, we have systematically investigated reconstructibility of physical interaction networks from thresholding statistical correlations. Beyond valuable previous studies which targeted the impact of correlated noise and estimation errors [39, 40], we revealed intrinsic limits of reconstructibility induced by the strengths of network interactions and their topology. In particular, a number of distinct topological factors contribute in a systematic way: local common cause structures, local relay structures, in-degree heterogeneities as well as non-local structural elements of a network resulting from different link densities in different network parts. Intriguingly, for stationary dynamics and arbitrary network topologies we uncovered a transition to full reconstructibility when decreasing the coupling strengths. Whereas the exact critical coupling strength to transition to reconstructibility depends on the topology, it is guaranteed to occur for all topologies.
Given the limitations of correlation thresholding, alternate methods of reconstruction from time series data are required (e.g, [6, 7, 24, 41]). In systems of linearly coupled spiking neurons, coupling strengths may for example be reconstructed using sparse reconstruction of connections [42] if connections are sparse or covariance inversion [36, 41] if temporal information is available.
For systems that are strongly non-linear and non-stationary, the range of inference methods is currently largely limited to systems with models known a priori. Such non-linear systems in general pose a number of additional challenges, including that there typically is no well-defined, temporally constant coupling strength between the units. Future studies would need to investigate model-independent methods to obtain physical interaction structure from recorded non-linear dynamics [4–11, 24].
Our main result on full reconstructiblity in the weak coupling limit might provide a useful initial step towards the reconstruction of non-linear and non-stationary networks: By systematically combining localized but faithful reconstructions obtained from an entire set of dynamics around different operation points in weakly coupled networks a global picture of the underlying interactions and their network state-dependencies could be obtained. Our finding, that the transition to reconstructibility is observed in all generic linear networks (see S1 Supplementary Material) yields promising perspectives for future investigations.
Our results on topology-induced limits of network reconstructibilty not only further our theoretical insights about the relations between statistical correlation and physical interaction networks [23, 24, 43] but also indicate where principal care has to be taken in applications when analyzing statistical correlation data to reveal aspects of direct physical interactions.
Acknowledgments
We thank E. Ching, A.-L. Barabasi and G. Yan for valuable discussions. This work was supported by the Max Planck Society (BL, MT), the Germany Ministry for Education and Research (BMBF) under grant no. 01GQ1005B (CK, MT), and an independent research fellowship by the Rockefeller University, New York, USA (CK). This article is part of the research activity of the Advanced Study Group 2017 From Microscopic to Collective Dynamics in Neural Circuits held at Max Planck Institute for the Physics of Complex Systems in Dresden (Germany).
References
- 1. Strogatz SH. Exploring complex networks. Nature. 2001;410(6825):268–76. pmid:11258382
- 2.
Newman M. Networks: An Introduction. NY, USA: Oxford University Press, Inc. New York; 2010. Available from: https://books.google.com/books?hl=de{&}lr={&}id=LrFaU4XCsUoC{&}pgis=1.
- 3. Kirst C, Timme M, Battaglia D. Dynamic information routing in complex networks. Nature Communications. 2016;7(6):11061. pmid:27067257
- 4. Yeung MKS, Tegnér J, Collins JJ. Reverse engineering gene networks using singular value decomposition and robust regression. Proc Natl Acad Sci U S A. 2002;99(9):6163–8. pmid:11983907
- 5. Gardner TS, di Bernardo D, Lorenz D, Collins JJ. Inferring genetic networks and identifying compound mode of action via expression profiling. Science. 2003;301(5629):102–5. pmid:12843395
- 6. Yu D, Righero M, Kocarev L. Estimating topology of networks. Phys Rev Lett. 2006;97(18):1–4.
- 7. Timme M. Revealing network connectivity from response dynamics. Phys Rev Lett. 2007;98(22):1–4.
- 8. Yu D, Parlitz U. Driving a network to steady states reveals its cooperative architecture. Europhys Lett. 2008;81(4):48007.
- 9. Yu D, Parlitz U. Inferring network connectivity by delayed feedback control. PLoS One. 2011;6(9):e24333. pmid:21969856
- 10. Shandilya SG, Timme M. Inferring network topology from complex dynamics. New J Phys. 2011;13(1):013004.
- 11. Ciofani M, Madar A, Galan C, Sellars M, Mace K, Pauli F, et al. A validated regulatory network for Th17 cell specification. Cell. 2012;151(2):289–303. pmid:23021777
- 12.
Casadiego J, Nitzan M, Hallerberg S, Timme M. No Title. submitted, 2017;.
- 13.
Drossel B, McKane AJ. Modelling Food Webs. In: Bornholdt S, Schuster HG, editors. Handb. Graphs Networks From Genome to Internet; 2002. p. 218–247.
- 14. Van Bussel F, Kriener B, Timme M. Inferring synaptic connectivity from spatio-temporal spike patterns. Front Comput Neurosci. 2011;5(February):3. pmid:21344004
- 15. Friston KJ. Functional and effective connectivity: a review. Brain Connect. 2011;1(1):13–36. pmid:22432952
- 16. Tkačik G, Walczak AM. Information transmission in genetic regulatory networks: a review. J Phys Condens Matter. 2011;23(15):153102. pmid:21460423
- 17. Rubido N, Martí AC, Bianco-Martínez E, Grebogi C, Baptista MS, Masoller C. Exact detection of direct links in networks of interacting dynamical units. New J Phys. 2014;16(9):093010.
- 18. Molkenthin N, Rehfeld K, Marwan N, Kurths J. Networks from flows–from dynamics to topology. Sci Rep. 2014;4:4119. pmid:24535026
- 19. Molkenthin N, Rehfeld K, Stolbova V, Tupikina L, Kurths J. On the influence of spatial sampling on climate networks. Nonlinear Process Geophys. 2014;21(3):651–657.
- 20. Guez OC, Gozolchiani A, Havlin S. Influence of autocorrelation on the topology of the climate network. Phys Rev E. 2014;90(6):062814.
- 21. Gilson M, Moreno-Bote R, Ponce-Alvarez A, Ritter P, Deco G. Estimation of Directed Effective Connectivity from fMRI Functional Connectivity Hints at Asymmetries of Cortical Connectome. PLoS Comput Biol. 2016;12(3):e1004762. pmid:26982185
- 22.
Aertsen A. Information Processing in the Cortex: Experiments and Theory. Springer; 2012. Available from: https://books.google.com/books?hl=de{&}lr={&}id=VO7uCAAAQBAJ{&}pgis=1http://material.brainworks.uni-freiburg.de/publications-brainworks/1992/books/braintheory-springer-1992.htm.
- 23. Timme M, Casadiego J. Revealing networks from dynamics: an introduction. J Phys A. 2014;47(34):343001.
- 24. Nitzan M, Casadiego J, Timme M. Revealing physical network interactions from statistics of collective dynamics. Sci Adv. 2017;3(February):e1600396. pmid:28246630
- 25. Opper M, Roudi Y, Sollich P. Special issue on modelling and inference in the dynamics of complex interaction networks: advanced approximation techniques. J Phys A. 2015;48(23):230201.
- 26. van Albada SJ, Helias M, Diesmann M. Scalability of Asynchronous Networks Is Limited by One-to-One Mapping between Effective Connectivity and Correlations. PLoS Comput Biol. 2015;11(9):e1004490. pmid:26325661
- 27. Wang WX, Lai YC, Grebogi C. Data based identification and prediction of nonlinear and complex dynamical systems. Physics Reports. 2016;644:1–76.
- 28. Hale JK. Diffusive coupling, dissipation, and synchronization. Journal of Dynamics and Differential Equations. 1997;9(1):1–52.
- 29. Pecora LM, Carroll TL. Master Stability Functions for Synchronized Coupled Systems. Physical Review Letters. 1998;80(10):2109–2112.
- 30.
Takeuchi Y. Global Dynamical Properties of Lotka-Volterra Systems. WORLD SCIENTIFIC; 1996. Available from: http://www.worldscientific.com/worldscibooks/10.1142/2942.
- 31.
Cantrell R, Cosner C, Cantrall RS, Cosner C. Spatial ecology via reaction-diffusion equations. J. Wiley; 2003. Available from: https://books.google.de/books?id=Cl8e4-4oWBwC{&}printsec=frontcover{&}dq=lotka+volterra+diffusion+pdf{&}hl=en{&}sa=X{&}ved=0ahUKEwiM8-PzrevVAhUBmBoKHd9NDUo4ChDoAQglMAA{#}v=onepage{&}q=lotkavolterradiffusionpdf{&}f=false.
- 32. Belik V, Geisel T, Brockmann D. Natural Human Mobility Patterns and Spatial Spread of Infectious Diseases. Physical Review X. 2011;1(1):1–5.
- 33. Grytskyy D, Tetzlaff T, Diesmann M, Helias M. A unified view on weakly correlated recurrent networks. Frontiers in computational neuroscience. 2013;7:131. pmid:24151463
- 34.
Gardiner C. Stochastic Methods: A Handbook for the Natural and Social Sciences. Springer Berlin Heidelberg; 2009. Available from: https://books.google.com/books?id=otg3PQAACAAJ{&}pgis=1.
- 35. Bartels RH, Stewart GW. Solution of the matrix equation AX + XB = C [F4]. Commun ACM. 1972;15(9):820–826.
- 36. Dahmen D, Bos H, Helias M. Correlated fluctuations in strongly coupled binary networks beyond equilibrium. Physical Review X. 2016;6(3).
- 37. Trousdale J, Hu Y, Shea-Brown E, Josić K. Impact of network structure and cellular response on spike time correlations. PLoS Computational Biology. 2012;8(3). pmid:22457608
- 38. Watts DJ, Strogatz SH. Collective dynamics of ‘small-world’ networks. Nature. 1998;393(6684):440–2. pmid:9623998
- 39. Bialonski S, Wendler M, Lehnertz K. Unraveling spurious properties of interaction networks with tailored random networks. PLoS One. 2011;6(8). pmid:21850239
- 40. Bialonski S, Lehnertz K. Assortative mixing in functional brain networks during epileptic seizures. Chaos. 2013;23(3). pmid:24089975
- 41. Ching ESC, Tam HC. Reconstructing links in directed networks from noisy dynamics. Physical Review E. 2017;95(1). pmid:28208378
- 42. Pernice V, Rotter S. Reconstruction of sparse connectivity in neural networks from spike train covariances. Journal of Statistical Mechanics: Theory and Experiment. 2013;2013(03):P03008.
- 43. Witt A, Kersting M, Geisel T, Nagler J. Backbones of network correlates. New J Phys. 2016 (submitted);.