Relating size and functionality in human social

networks through complexity

B. J. Westa,1, G. F. Massarib, G. Culbrethb, R. Faillab, M. Bolognac, R. I. M. Dunbard, and P. Grigolinib
a
Office of the Director, Army Research Office, Research Triangle Park, NC 27709; bCenter for Nonlinear Science, University of North Texas, Denton, TX
76203-1427; cDepartamento de Ingeniería Eléctrica-Electrónica, Universidad de Tarapacá, 6-D Arica, Chile; and dDepartment of Experimental Psychology,
University of Oxford, OX2 6GG Oxford, United Kingdom

Edited by Douglas S. Massey, Princeton University, Princeton, NJ, and approved June 30, 2020 (received for review April 10, 2020)

Extensive empirical evidence suggests that there is a maximal Another way to view the relation between network complexity
number of people with whom an individual can maintain stable and network size is by relating the functionality of the network of
social relationships (the Dunbar number). We argue that this arises interest to its complexity. West (5) established that the more
as a consequence of a natural phase transition in the dynamic self- sophisticated the functionality, the greater the complexity nec-
organization among N individuals within a social system. We pre- essary to support that function. For example, the degree of
sent the calculated size dependence of the scaling properties of complexity necessary to sustain the functionality of a modern city
complex social network models to argue that this collective behav- is, proportionately, significantly greater than that necessary to
ior is an enhanced form of collective intelligence. Direct calculation sustain the functionality of a primitive village. Consequently, we
establishes that the complexity of social networks as measured by interpret the many empirical relations, between functionality and
their scaling behavior is nonmonotonic, peaking around 150, size as being the result of an implicit relation between size and
thereby providing a theoretical basis for the value of the Dunbar complexity, with complexity being manifest through the system’s
number. Thus, we establish a theory-based bridge spanning the
functionality. This subtle, yet ubiquitous, driving of complexity by

ANTHROPOLOGY
gap between sociology and psychology.
size and in turn functionality being driven by complexity, has long
been known from the study of allometry (5): Average network
|
Dunbar number allometry relation | network calculations | complexity | functionality is typically a noninteger power of average network
functionality/size
size, without an explicit dependence on complexity.
In the recent past it has been argued that biological systems
D unbar hypothesized, on the basis of archeological, evolu-
tionary, and neurophysiological evidence, that 150 is the
limit on the number of people with whom a typical person can
function best when their dynamics are close to criticality (6). This
hypothesis is in keeping with the more general observations of
Anderson (7) regarding the disconnect between microdynamics

MATHEMATICS
maintain stable social relationships (1, 2). We suggest that this is and emergent macrodynamics in complex dynamic systems that
a consequence of internal dynamics producing self-organized

APPLIED
undergo phase transitions. For example, when a liquid is boiled it
criticality within a social network consisting of N people. We becomes a gas and the corresponding volume increases discon-
use two distinct complex network models of social group dy- tinuously as a manifestation of criticality. This universal behavior
namics that lead to phase transitions (3), termed criticality in the is manifest in the scaling behavior of certain system parameters
physics literature, to determine the optimal size of networks and called critical exponents, on which there is now a vast literature.
compare this with the Dunbar number. Such criticality generates
intermittent events, with time intervals between successive events Significance
being independent (renewal) and having an inverse power law
distribution. The inverse power law index in both network
Dunbar hypothesized, on the basis of empirical evidence, that a
models is shown herein by direct calculation to increase rapidly typical individual can have a stable relation with at most 150
in magnitude from 0.5, reach a maximum of ∼0.67, and then other people. We establish that this number results from the
decrease slowly back to 0.5, as the size of the network increases. internal dynamics of a complex network. Two network models
This nonmonotonic dependence of the scaling index on network having phase transitions are used to determine the optimal
size is a signature of complexity (4) and is used to argue that the size for the most efficient information exchange. Such criti-
collective social behavior at criticality supports optimal infor- cality generates intermittent events, with time intervals be-
mation transmission within the group. Consequently, the time to tween successive events being independent (renewal) and
transmit information between generic complex networks is scaling. The scaling index depends nonmonotonically on net-
minimal when both system sizes coincide with the predicted work size and direct calculations show that the index is maxi-
Dunbar number. Thus, the calculations presented herein yield a mum for networks the size of the Dunbar number and provides
theory-predicted value of the maximum group size that closely maximal information exchange efficiency. This result provides
agrees with the empirical Dunbar number, as well as showing a theory-based bridge to span the conceptual gap between
that networks of this size have optimal information transmission psychology and sociology.
properties. These results provide a theory-based bridge that uses
Author contributions: B.J.W. designed research; B.J.W., M.B., R.I.M.D., and P.G. performed
network science model calculations to span the current concep- research; G.F.M., G.C., R.F., and R.I.M.D. analyzed data; G.F.M., G.C., and R.F. did calcula-
tual gap between psychology and sociology. tions; and B.J.W., R.I.M.D., and P.G. wrote the paper.
In order for a group, organization, or living network to The authors declare no competing interest.
maintain its functionality as its size increases, macroscopic dy- This article is a PNAS Direct Submission.
namic modes must emerge to replace those that no longer sup- This open access article is distributed under Creative Commons Attribution License 4.0
port the system’s evolving purpose as driven by the increase in (CC BY).
complexity. The network’s size and functionality increase and 1
To whom correspondence may be addressed. Email: brucejwest213@gmail.com.
decrease together as determined by their separate relationships This article contains supporting information online at https://www.pnas.org/lookup/suppl/
to changing complexity, but not in direct proportion to
Downloaded by guest on May 4, 2021

doi:10.1073/pnas.2006875117/-/DCSupplemental.
one another. First published July 20, 2020.

www.pnas.org/cgi/doi/10.1073/pnas.2006875117 PNAS | August 4, 2020 | vol. 117 | no. 31 | 18355–18358

 In the complex networks used herein the universal properties are science decision-making model (DMM) used in our calculation. The DMM is
criticality, a dependence of critical parameters on network size based on a two-state master equation for each individual in the network (3).
N, and fluctuations due to the finite size of the network whose In the DMM an isolated individual randomly switches between two states,
intensity varies as as N−1/2 (3). and the random switching mechanism is identified as innovation. An indi-
vidual also imitates those with whom she or he interacts so that the imita-
The social organizations of primates provide a well-studied tion parameter determines the influence of others on the switching time of
example of biological social networks. A quarter century ago the individual. We demonstrate by network science model calculations that
Dunbar (1, 2) related measures of the average neocortical ratio N cooperatively interacting units generate criticality, namely, long-range
C to mean group size N of a sample of 36 primate genera and correlations enabling the system to make cooperative decisions and prop-
constructed a version of the social brain allometry relation erly respond to environmental perturbations. The efficiency of this response
(SBAR): requires a balance between the fluctuation intensity (proportional to N−1/2)
and the inverse power law regime in the limit of N very large. We prove that
this balance occurs in the neighborhood of N = 150.
N = aCb ,
We use a statistical analysis of time series generated by criticality-induced
intelligence, based on a method recently proposed to detect crucial events
where a and b are the empirical constants: a = 1.239 and b = by Culbreth et al. (20) to find the criticality point in the complex network
3.389. Here, C is the ratio of neocortical volume to that of the dynamical models. This method is based on converting empirical time-series
total volume of the brain minus that of the neocortex. The av- data into a diffusion process from which the probability density function (SI
erage neocortex ratio for humans was measured to be C = 4.1 Appendix) is calculated and the entropy determined. The procedure is called
(8), which when inserted into the SBAR model yields an average diffusion entropy analysis (DEA) and is used to determine the scaling be-
group size of N = 147.8. This value was rounded off to 150 in the havior of the empirical process driving the diffusion. When criticality-
subsequent literature and is now called the Dunbar number. induced intelligence (collective intelligence) becomes active, the driven
The social brain hypothesis (2) from which the Dunbar num- process is expected to depart from ordinary diffusion signified by having a
scaling index different from δ = 0.5. The modified DEA (MDEA) illustrated in
ber results, heuristically bridges the gap between psychology and
Culbreth et al. (20) filters out the scaling behavior of infinite stationary
sociology, in that it relates an empirical measure of cognition to memory, when it exists (21), and the remaining deviation of the scaling in-
the average size of a social group consisting of individuals dex from δ = 0.5 is solely due to crucial events.
sharing that average cognition measure through the SBAR. Al- The MDEA applied to the signal generated by the criticality-induced in-
though the Dunbar number is widely known, no first principles telligence implements the original DEA in conjunction with the method of
explanation as yet exists for it. Consequently, the search for a stripes. In the method of stripes the vertical axis is divided into many bins of
theory to predict the Dunbar number has shifted from explaining equal size and an event, either crucial or not, is recorded when the signal
the social brain hypothesis to establishing a rationale as to why moves from a given stripe to an adjacent stripe. A random walk (RW) step is
the number turns out to be ∼150. triggered by such an event and makes a step of constant length forward
each time an event occurs, thereby generating a diffusion-like trajectory
The back and forth exchange of information between complex
X(t). This trajectory contains information on the persistence of opinion
networks has been predicted theoretically with the principle of contained in the empirical time series, which can be detected by applying
complexity matching (4) and is observed experimentally in, for the MDEA method to X(t) determined from the RW rate equation (SI
example, turn-taking in dyadic conversations (9), the therapeutic Appendix).
influence of arm-in-arm walking (10), and the influence of In order to explore the generality of our results, we have selected two
zealots on group behavior (3). Human social networks, in par- network models generating criticality-induced intelligence that have totally
ticular, are best viewed as being designed to solve coordination different microdynamics, but both have transitions to criticality. The first is
problems (11), and these lend themselves naturally to the format the DMM, where N individuals each choose between two conflicting states,
sometimes known as opinion dynamics models (12, 13). The which they do under the influence of their nearest neighbors (SI Appendix).
greater the complexity of a network the more information the This model falls into the Ising universality class, thereby making it possible to
compare the results obtained herein to the predictions of Chialvo (22). The
network contains and just as an entropy gradient provides an
second model is that of swarm intelligence (SI) proposed by Vicsek et al. (23)
entropic force in a physical network, a complexity gradient and is also a member of the Ising universality class. These models are ap-
provides an information force between living networks. Conse- propriate for describing human social networks since human (and primate)
quently, with the brain of the individual modeled as one complex social groups exist to enable behavioral (and informational) coordination
network and the social group as a second complex network, the (24). They are also appropriate for modeling the way in which an individual
tools of network science are used herein to provide a theoretical models their own social network in the virtual world they construct within
explanation for the value of the Dunbar number as the value of their brain. These social and neurologic networks correlate incompletely
network size that optimizes information transport. with one another, with information from the one influencing the way the
In keeping with the criticality hypothesis, it is reasonable to other organizes itself (we adjust our model network as formulated in our
mind/brain as a function of information received about observed changes in
implement the SBAR association of functionality and size with
the state of the social network).
the emergence of complexity from criticality. This is achieved
To show that network size is optimal with regard to the transfer of in-
dynamically by critical dynamics generating crucial events. formation, we consider two complex networks, A and B, which interact with
Complexity, as measured by functionality, is manifest in the one another at criticality. We seek the size at which they are identical, having
collective behavior of nonlinear dynamic networks to model the same size N. For a time L, the global field ξA(t) of network A and the field
cognition, using the concept of collective intelligence (14). Long- ξB(t) of network B are calculated and the cross-correlation between the two
range correlations are amplified at the onset of phase transition time series determined.
and are often studied by means of dynamic networks that are This cross-correlation experiment is done in two ways. In the first, a small
members of the Ising universality class (15), which provides the percentage (5% of units of A, randomly chosen) make their choice on the
mathematical rationale for the complex network models used in basis of the choices made by their nearest neighbors and one randomly
chosen unit of network B. Network B is influenced by network A through the
the numerical calculations presented herein. These network
same interaction process. As a result of this back and forth interaction the
models at criticality generate intermittent (16) and crucial events cross-correlation time is expected to be symmetric around τ = 0. This trivial
(3, 17), which according to Alligrini et al. (18) is a manifestation observation turns out to be the case and therefore its calculation is not
of consciousness. shown herein.
In the second case, the interaction is restricted to one direction with
Materials and Methods network A perturbing B, but no return perturbation of B on A. As a con-
Wisdom and Goldstone (19) conducted and interpreted a sequence of ex- sequence of the unidirectional nature of the information flow, we expect
Downloaded by guest on May 4, 2021

periments on problem-solving tasks using the social concepts of innovation that the cross-correlation function ought to shift to a positive time delay as
and imitation. These same two concepts are fundamental in the network the network size increases beyond the critical size. The time delay is a

18356 | www.pnas.org/cgi/doi/10.1073/pnas.2006875117 West et al.

 consequence of the information about A transmitted by 5% of them to the
B network and does not instantaneously change the behavior of all of the B
units. It requires a finite time to transmit the information from the one B
unit receiving the information about the motion of the A units to all of the
other B units. Attanasi et al. (25) made the conjecture that the transmission
of information from one peripheral individual in a network who acquires
some important item of environmental information (e.g., an approaching
predator) to all of the other members of the network occurs through a
diffusion process. Lukovic et al. (26) assigned an important role in infor-
mation transmission to the visible crucial events and argued that the phase
state of the network requires a sufficiently large number of dynamic
switches.

Material and Data Availability. The original sources for all data used herein
have been cited in the main text, SI Appendix, or the cited literature. The
research codes used in the calculations of the figures have been cited and
are available upon request from the University of North Texas authors:
G.F.M., G.C., R.F., and P.G.

Results
We evaluate first the mean field of a DMM network to produce
the signal and Fig. 1 illustrates the results of that analysis. Here
the calculation on the network is done in two different ways. One
way is with every individual interacting with every other indi-
vidual in the network (all-to-all, ATA). The other is that each Fig. 2. Scaling detection of the Dunbar number is obtained by calculating
individual interacts only with its nearest neighbor on a 2D lattice, the nonmonotonic dependence of the scaling index δ on a network of size N,

ANTHROPOLOGY
using the SI model of Vicsek et al. (23).
with periodic boundary conditions (2D-lattice). Both calculations
yield criticality at the appropriate theoretical values of the con-
trol parameter, whose critical values depend on the size of the and SI in Fig. 2, with respect to the nonmonotonic dependence
network. Identifying the calculated value of the time rate of of the scaling index on network size, is remarkable. Most note-
change of the mean field variable with the empirical time series
worthy for our purposes here is that both networks display
ξ(t), we generate the RW and obtain the trajectory X(t) to which
dominant peaks in the vicinity of the Dunbar number, which
we apply the MDEA to obtain the scaling index as a function of
network size (SI Appendix). The resulting scaling parameter δ is given the empirical value of 148.7 from the SBAR is truly

MATHEMATICS
shown in Fig. 1 and varies nonmonotonically with the size of the astonishing.
network. The parameter peaks, achieving a maximum value close Of course, theoretically predicting the Dunbar number does

APPLIED
to δ = 0.67, when N is in the vicinity of the Dunbar number 150. not establish that this size of the network influences the trans-
The scaling index falls quickly to δ = 0.5 to the left of the peak, mission of information, much less that the Dunbar size of a
for N < 150 and more slowly to the same value to the peak’s network optimizes the exchange of information between net-
right, for N > 150. works. We can, however, determine network efficiency from the
The same calculation is carried out using the SI model pro- cross-correlation of the time series for the perturbing network A
posed by Vicsek et al. (23) and the results are displayed in Fig. 2. and the perturbed network B. Fig. 3 shows that the time delay
The qualitative agreement observed between the DMM in Fig. 1 between the driven and the driving networks is extremely small
when N = 150 (the delay time is τ ≈ 0), whereas larger networks
require a finite nonzero time to reorganize and maximize their
correlation. The larger the deviation in network size from the
Dunbar number, the greater the delay in transmitting the in-
formation throughout the perturbed network. This is an evident
signal that the Dunbar effect facilitates the transport of infor-
mation from the individual who first acquires the information to
all of the other individuals in the network.
As the interaction strength (K) between individuals in a net-
work approaches the critical value (KC), the dynamics of indi-
viduals change from virtually independent behavior (K < KC) to
highly organized behavior at the critical point (K = KC). The two-
time correlation function changes from a rapid exponential re-
laxation of perturbations when the interaction strength is sub-
critical (K < KC), to a much slower inverse power law relaxation
of perturbations at criticality (K = KC), and then returns to the
rapid exponential relaxation when the interaction strength
becomes supercritical (K > KC). The substantially slower relax-
ation perturbations at criticality entails long-range correlations,
whose persistence facilitates the information transfer at criti-
cality. The measure of this persistence is manifest in the degree
Fig. 1. Scaling detection of the Dunbar number is obtained by calculating
to which the scaling index δ exceeds 0.5, but note that even at δ =
the nonmonotonic dependence of the scaling index δ on a network of size N.
The two calculation are depicted using a DMM (3): The red circles with an 0.5 (the value obtained in an unbiased RW process), the network
index is μ = 1.5, denoting that the network is still at criticality (SI
Downloaded by guest on May 4, 2021

ATA interaction, the blue circles with a nearest-neighbor interaction on a

2D lattice. Appendix).

West et al. PNAS | August 4, 2020 | vol. 117 | no. 31 | 18357

 allometry relation by eliminating time from two rate equations
describing organs growing at different rates within the same
body. Assuming that time and complexity are directly propor-
tional (28, 29), the elimination of one is equivalent to the
elimination of the other in the allometry relation between the
average number of people in a social group and the average
cognitive measure in the SBAR. From this we establish by nu-
merical calculations that the Dunbar number is the optimal
group size. This conclusion is here entailed by a network science
theoretical model, but was reached by Dunbar using the archived
evolutionary and neurophysiological data for multiple species
over long time periods (1, 2).
The scaling index was used as a measure of the network’s
dynamic complexity as previously done in a broad range of ap-
plications (3). Consequently, determining the dependence of the
magnitude of the scaling parameter on the size of the network
enabled us to ignore the microdynamics of the two models used
in the calculations and determine that their complexities share
the same functional dependence on network size. The remark-
able result is that the two model calculations suggest that the
complexity that manifests itself in the control of information
Fig. 3. The cross-correlation function of two interacting networks, A and B, transport in the dynamics of complex networks also determines
at criticality are depicted for three sizes of networks. Here 5% of A units the empirical SBAR and consequently the Dunbar number. In
determine their behavior by selecting the average behavior of their six
addition, the cross-correlation calculation indicates that the
nearest neighbors and one unit of B. The unit of B is influenced by the 5% of
A, but otherwise interacts normally with the other members of its group. Dunbar number is a consequence of the information transfer
The cross-correlation function is calculated using a z-variable obtained by being optimal for complex dynamic networks of this size. The
centering each time series on its time averaged value and normalizing the significance of this fact in terms of the flexibility and stability of a
difference variable to the SD to obtain: C(τ) = 〈ZA (t)ZB (t + τ)〉, where the social group cannot be overemphasized, since it provides an
brackets denote a time average over the interval L. evolutionary advantage in which collections of weaker individ-
uals find added strength in groups of a preferred size, whereby
they can respond to a predatory attack or other ecological
Discussion
threats as a single collective entity and thus survive.
We have established that complexity can be a hidden variable
responsible for an empirical relation being directly observed ACKNOWLEDGMENTS. G.F.M., G.C., R.F., and P.G. thank the US Army
between a network’s size and functionality. Recall that this was Research Office for supporting this research work through Grant
how Huxley (27) was able to “prove” the theoretical form of the W911NF1910104.

1. R. I. M. Dunbar, Neocortex size as a constraint on group size in primates. J. Hum. Evol. 16. Y. F. Contoyiannis, F. K. Diakonos, A. Malakis, Intermittent dynamics of critical fluc-
20, 469–493 (1991). tuations. Phys. Rev. Lett. 89, 35701 (2002).
2. R. I. M. Dunbar, Coevolution of neocortical size, group size and language in humans. 17. H. G. Schuster, Deterministic Chaos: An Introduction, (VCH Publisher, New York, ed. 2,
Behav. Brain Sci. 16, 681–735 (1993). 1988).
3. B. J. West, M. Turalska, P. Grigolini, Networks of Echoes, Imitation, Innovation and 18. P. Allegrini et al., Self-organized dynamical complexity in human wakefulness and
Invisible Leaders, (Springer, 2014). sleep: Different critical brain-activity feedback for conscious and unconscious states.
4. B. J. West, E. L. Geneston, P. Grigolini, Maximizing information exchange between Phys. Rev. E Stat. Nonlin. Soft Matter Phys. 92, 032808–032814 (2015).
complex networks. Phys. Rep. 468, 1–99 (2008). 19. T. N. Wisdom, R. L. Goldstone, Innovation, imitation, and problem-solving in a net-
5. B. J. West, Nature’s Patterns and the Fractional Calculus, (Walter De Gruyther, Berlin, worked group. Nonlinear Dyn. Psychol. Life Sci. 15, 229–252 (2011).
Boston, 2017). 20. G. Culbreth, B. J. West, P. Grigolini, Entropic approach to the detection of crucial
6. T. Mora, W. Bialek, Are biological systems poised at criticality. J. Stat. Phys. 144, events. Entropy 21, 178–189 (2019).
268–275 (2011). 21. B. B. Mandelbrot, J. W. Van Ness, Fractional Brownian motions, fractional noises and
7. P. W. Anderson, More is different. Science 177, 393–396 (1972). applications. SIAM Rev. 10, 422–431 (1968).
8. H. Stephan, H. Frahm, G. Baron, New and revised data on volumes of brain structures 22. D. R. Chialvo, S. A. Cannas, D. Plenz, T. S. Grigera, Controlling a complex system near
in insectivores and primates. Folia primatol. 35, 1–29 (1981). its critical point via temporal correlations. arXiv:1905.11758v1 (28 May 2019).
9. D. H. Abney, A. Paxton, R. Dale, C. T. Kello, Complexity matching in dyadic conver- 23. T. Vicsek, A. Czirók, E. Ben-Jacob, I. Cohen, O. Shochet, Novel type of phase transition
sation. J. Exp. Psychol. Gen. 143, 2304–2315 (2014). in a system of self-driven particles. Phys. Rev. Lett. 75, 1226–1229 (1995).
10. Z. M. H. Almurad, C. Roume, H. Blain, D. Delignières, Complexity matching: Restoring 24. T. Dávid-Barrett, R. I. M. Dunbar, Cooperation, behavioural synchrony and status in
the complexity of locomotion in older people through arm-in-arm walking. Front. social networks. J. Theor. Biol. 308, 88–95 (2012).
Physiol. 9, 1766 (2018). 25. A. Attanasi et al., Finite-size scaling as a way to probe near-criticality in natural
11. T. Dàvid-Barrett, R. I. M. Dunbar, Language as a coordination tool evolves slowly. R. swarms. Phys. Rev. Lett. 113, 238102–238106 (2014).
Soc. Open Sci. 3, 160259 (2016). 26. M. Lukovic, F. Vanni, A. Svenkeson, P. Grigolini, Transmission of information at crit-
12. C. Castellano, S. Fortunato, V. Loreto, Statistical physics of social dynamics. Rev. Mod. icality. Physica A 416, 430–438 (2014).
Phys. 81, 591–646 (2009). 27. J. S. Huxley, Problems of Relative Growth, (Dial Press, New York, 1931).
13. G. Iñiguez, T. Govezensky, R. Dunbar, K. Kaski, R. A. Barrio, Effects of deception in 28. B. J. West, K. Mahmoodi, P. Grigolini, Empirical Paradox, Complexity Thinking and
social networks. Proc. Biol. Sci. 281, 20141195 (2014). Generating New Kinds of Knowledge, (Cambridge Scholars Publishing, Cambridge,
14. I. Couzin, Collective minds. Nature 445, 715 (2007). UK, 2019).
15. J. Cardy, Scaling and Renormalization in Statistical Physics, Cambridge Lecture Notes 29. P. Grigolini, L. Palatella, G. Raffaelli, Asymmetric anomalous diffusion: An efficient
in Physics, (Cambridge University Press, Cambridge, UK, 1996). way to detect memory in time series. Fractals 9, 439–449 (2001).
Downloaded by guest on May 4, 2021

18358 | www.pnas.org/cgi/doi/10.1073/pnas.2006875117 West et al.