Downloaded by guest on September 30, 2021 a Denton K. Kaleda cultural in transmission bias conformity changing randomly On PNAS biases. 14) (13, other and/or frequency-dependent name 12) with baby (8, However, consistent 11). more (10, appear names distributions motifs, baby pottery and Neolithic citations, (9), breeds suggested patent dog been of choices has for (8), account performing frequency-dependent copying to random children Unbiased as known young (7). transmission, by discrimination exhibited numerical been Anti- has (7). tasks conformity discrimination numerical and (6), discrimination frequency, population its is than variant less or common (4). greater more respectively rate a a and (3). at when of population conformity adopted occur terms the include which in in biases anticonformity, variants couched char- cultural “frequency-dependent” is and the (2). These biases of quality, prestige transmission frequencies the and the of success as char- class including such Another incorporated anal- transmitters, variation, Other have of groups. trait acteristics biases between of transmission vari- value acteristics cultural increasing trait of to that average lead yses of the could constrained value in but transmission” mean ance variability “group the This within-group on population. the depended the in next trait the to generation C selection random these conformity random to realism added contribute would population models. finite migration incorporating and research sizes differences Future on migration, difficult. by conformity connected rather of those example effect to for populations, promises the properties between This of group-level biases. interpretation of (anti)conformist realis- make range under more emerge wide a can the produce that of may representation stochastically selec- vary tic weak models, role to changing of selection number randomly the and conformity, of of level effect simultaneously. the Allowing stable the tion. coefficients, locally study determinis- conformity stochastically we in be random Finally, can those with equilibria from example, all different previ- For be models. in may tic seen equilibria is these equilibria result of in three this variation Moreover, stochastic the conformity. to of robust of copying there models one expected, deterministic random ous is to or anticonformity) convergence nor conformity is conformity if neither occur, (i.e., strong if may Even (n and values. anticonformity time parameter deterministic different over in with vary models occur stochastic to can models num- role nonconvergence fixed of anticonformity, number a the assumes allow which We model, time. conformity ber into used over coefficients widely change conformity a can time-varying populations stochastic extent incorporate in the that (anti)conformity suggests con- of research many constant empirical Whereas incorporated coefficients, as have formity transmission. well models McElreath cultural as Richard mathematical and conformist Kendal, previous in R. Akc¸ay, Jeremy display Erol biases by animals reviewed anticonformist 2021); nonhuman 15, April and review for Humans (sent 2021 6, July Feldman, W. Marcus by Contributed eateto ilg,Safr nvriy tnod A935 and 94305; CA Stanford, University, Stanford Biology, of Department uashv xiie ofriyi etlrtto 5,line (5), rotation mental in conformity exhibited have Humans aindnmc fatatwoetasiso rmone from popu- transmission whose finite trait the a studied of (1) dynamics Feldman lation and avalli-Sforza n 01Vl 1 o 4e2107204118 34 No. 118 Vol. 2021 f“oemdl”smldb ahidvda.W also We individual. each by sampled models” “role of | tcatclclstability local stochastic a r Liberman Uri , n t oee,dnmcproperties dynamic However, . b | convergence n acsW Feldman W. Marcus and , | t .Under ). b colo ahmtclSine,TlAi nvriy e vv sal69978 Israel Aviv, Tel University, Aviv Tel Sciences, Mathematical of School ulse uut2,2021. 20, August Published doi:10.1073/pnas.2107204118/-/DCSupplemental.y at online information supporting contains article This 1 BY-NC-ND) (CC .y Attribution-NonCommercial- 4.0 Commons NoDerivativesCreative License distributed under is article access open This interest.y competing no declare Max authors The R.M., and Durham; Anthropology.y of Evolutionary for University Institute J.R.K., Planck Pennsylvania; of University E.A., research, Reviewers: performed research, paper. designed y M.W.F. the wrote and and U.L., K.K.D., contributions: Author vr otpeiu ahmtclmdl fcnomt have conformity conformity time-dependent, of coefficients. than turnover models rather convex constant, mathematical How- incorporated a previous bias. indicative show direct most names or 1930) ever, function frequency-dependent baby to positive turnover male of (1880 indicative but concave from decades names earlier bias, a baby from frequency-dependent show US the negative popular 2010 to of 12, to ref. 1950s In 1960 the Aschian. than from other declined and, also has humans has of States, It studies (16). (29). United 133 1990s tits least see the great at (but observed in and in 26) been tested (28), (25, empirically has monkeys monkeys been vervet (22) capuchin 27), conformity 24), ref. Aschian (23, chimpanzees 34). others” in p. in behav- observed 22, or options knowledge (ref. countervailing personal by of dispositions overriding ioral “the namely not 4, did ref. therefore and anticon- use, anticonformist). we of be to which definition behaviors 4, these different ref. consider of a that used choice from both authors mate formity displayed to these flies respect (but Fruit with 18). (19) bias ref. anticonformist see solving (but and tits 17) conformist great (16, and box (15) puzzle feeder a a choosing sticklebacks spined owo orsodnemyb drse.Eal [email protected]. y Email: addressed. be may correspondence whom To a,1 oe rvdsaueu prxmto otestochastic the to approximation deterministic useful the a model. however, provides straightfor- conditions, be model some not Under therefore may ward. poly- in (anti)conformity and differences of monomorphisms population simultane- terms of real-world including Interpreting stability possible, morphism. local are stochastic conformity established dynamics ous weak an Novel and in conformity stochastically model. models,” allow vary “role to we sampled selection of systems, included numbers real-world not coefficients, improve have to vary potentially can models To applicability coefficients. conformity mathematical conformity of most time-varying level time, population’s over a evidence fads, empirical that including Although suggests behavior. groups collective of and properties norms, emergent (anti)conformity explain under may dynamics Population anticonformity. to humans from Animals, Significance eprlvrainmyas cu nfrso conformity of forms in occur also may variation Temporal from conformity of definition different a used 21) (20, Asch nine- in observed been has conformity animals, nonhuman In https://doi.org/10.1073/pnas.2107204118 ipa ofriyand conformity display Drosophila, https://www.pnas.org/lookup/suppl/ | f12 of 1
ANTHROPOLOGY EVOLUTION Cavalli-Sforza and Feldman (ref. 3, chap. 3) and Boyd “non-” or “anticonformist” states over time (40, 42, 47, 50, 51), and Richerson (ref. 4, chap. 7) studied models of frequency- to our knowledge, random temporal variation in the confor- dependent transmission of a cultural trait with two variants. Boyd mity coefficients themselves has not been modeled previously. and Richerson (4) incorporated conformist and anticonformist In reality, the degree to which groups of individuals conform bias through a conformity coefficient denoted by D. In their may change over time, as illustrated by the finding that young simplest model, if the frequency of variant A is p and that of vari- children anti-conformed while older children conformed in a ant B is 1 − p, then the frequency of variant A in the offspring discrimination task (7); thus, it seems reasonable to expect generation, p0, is that different generations may also exhibit different levels of conformity. Indeed, generational changes have occurred for p0 = p + Dp(1 − p)(2p − 1), [1] Aschian conformity (29) and possibly in frequency-dependent copying of baby names (12). Our stochastic model may there- where D > 0 entails conformity (A increases if its frequency is fore produce more realistic population dynamics than previ- 1 p > 2 ), D < 0 entails anticonformity, D = 0 entails random copy- ous deterministic models, and comparisons between the two ing, and −2 < D < 1. In this model, each offspring samples the can suggest when the latter is a reasonable approximation to cultural variants of n = 3 members of the parental generation the former. (hereafter, role models). Sampling n > 3 role models requires We also allow the number of role models, nt , to vary over different constraints and, if n > 4, there are multiple conformity time. Agent-based conformity models have incorporated tem- coefficients (Eq. 19). poral (43) and individual (43, 45, 46) variation in the number Many subsequent models have built upon Boyd and Rich- of sampled individuals, whereas here, all members of genera- erson’s (4) simplest model (Eq. 1). These have incorpo- tion t sample the same number nt of role models. Causes of rated individual learning, information inaccuracy due to envi- variation in nt are not explored here, but there could be sev- ronmental change (30–34), group selection (35), and other eral. For instance, different generations of animals may sample transmission biases, including payoff bias (36), direct bias, different numbers of role models due to variation in popula- and prestige bias (37). Other models, which include a sin- tion density. In humans, changes in the use of social media gle conformity coefficient and preserve the essential features platforms or their features may cause temporal changes in of Eq. 1, incorporate individual learning, environmental vari- the number of observed individuals. For example, when Face- ability (32, 38), group selection (39), and multiple cultural book added the feature “People You May Know,” the rate of variants (38). new Facebook connections in a New Orleans dataset nearly In agent-based statistical physics models, the up and down doubled (52). spins of an electron are analogous to cultural variants A and In the stochastic model without selection, regardless of the B (40, 41). Individuals are nodes in a network and choose fluctuation in the conformity coefficient(s), if there is confor- among a series of actions with specified probabilities, such as mity on average, the population converges to one of the three independently acquiring a spin, or sampling neighboring individ- equilibria present in Boyd and Richerson’s (4) model with con- uals and adopting the majority or minority spin in the sample. n ∗ formity (D(j ) > 0 for 2 < j < n in Eq. 19). These are p = 1 The number of sampled role models can be greater than three ∗ ∗ 1 (fixation on variant A), p = 0 (fixation on variant B), and p = 2 (42, 43). (Anti)conformity may occur if all (42–47), or if at (equal representation of A and B). However, their stability least r (40, 48), sampled individuals have the same variant. properties may differ from those in the deterministic case. In In contrast, Boyd and Richerson’s (4) general model (Eq. 19) Boyd and Richerson’s (4) model with random copying, every allows, for example, stronger conformity to a 60% majority initial frequency p0 is an equilibrium. Here, with random copy- of role models and weaker conformity or anticonformity to a ing expected and independent conformity coefficients, there is 95% majority (in humans, this might result from a perceived ∗ 1 convergence to p = 0, 2 , or 1. In this case, and in the case difference between “up-and-coming” and “overly popular” ∗ 1 with conformity expected, convergence to p = 0, 2 , or 1 also variants). holds with stochastic variation in the number of role models, In Boyd and Richerson’s (4) general model, individuals sam- nt . With either stochastic or constant weak selection in Boyd n n ple role models, which is more realistic than restricting and Richerson’s (4) simplest model (Eq. 1) and random copying 1 ∗ to 3 (as in Eq. ); individuals may be able to observe more expected, there is convergence to a fixation state (p = 0 or 1). n > 4 than three members of the previous generation. With , Finally, with anticonformity in the deterministic model or anti- different levels of (anti)conformity may occur for different sam- conformity expected in the stochastic model, nonconvergence j n ples of role models with one variant. In addition to the can occur. example above with 60 and 95% majorities, other relationships between the level of conformity and the sample j of n are pos- Generalizing the Deterministic Model sible. For example, the strength of conformity might increase Here we generalize Eq. 1 to allow the coefficient D to change as the number of role models with the more common variant randomly at each generation. If t denotes the generation number increases. In a recent exploration of Boyd and Richerson’s (4) (t = 0, 1, 2, ... ), then Eq. 1 becomes general model, we found dynamics that departed significantly from those of Eq. 1 (49). If conformity and anticonformity occur n for different majorities j of n role models (i.e., j > 2 ), poly- morphic equilibria may exist that were not possible with Eq. 1. pt+1 = pt + Dt pt (1 − pt )(2pt − 1). [2] In addition, strong enough anticonformity can produce noncon- vergence: With as few as 5 role models, stable cycles in variant frequencies may arise, and with as few as 10 role models, chaos In Eq. 2 we assume that the random variables Dt for t = is possible. Such complex dynamics may occur with or without 0, 1, 2, ... are identically distributed with −2 < Dt < 1 for n = 3 2 selection. role models, E(Dt ) = d, Var(Dt ) = σ , and Dt are independent Here, we extend both Boyd and Richerson’s (4) simplest of pt in each generation t. We focus on the possible convergence (Eq. 1) and general (Eq. 19) models to allow the conformity of pt to equilibrium as t → ∞. coefficient(s) to vary randomly across generations, by sampling Eq. 1 entails that σ2 = 0, and we now summarize the known them from probability distributions. Although some agent-based results for this model. If the conformity coefficients are constant models allow individuals to switch between “conformist” and over time, i.e., Dt ≡ d 6= 0 for all t, then
2 of 12 | PNAS Denton et al. https://doi.org/10.1073/pnas.2107204118 On randomly changing conformity bias in cultural transmission Downloaded by guest on September 30, 2021 Downloaded by guest on September 30, 2021 x If 2) ,2, 1, 1. variable probability ensures random theorem a is convergence there martingale that the Therefore, gale. all {p for 1 ability Proof: is, That 1. Result If 2) eut2. Result result: following the gives theorem convergence when only equality with Since 1. Observation of Proof If 1) 2. Observation If 3) nrnol hnigcnomt isi utrltransmission cultural in bias conformity changing randomly On al. et Denton If 1) 2. Observation of Proof and If 2) observations following the process on the based for is of analysis Our convergence distribution. global also process is the there generations, over randomly If 4) As 2) equilibria, possible three are There 1) eainbtenteecntn qiiraadterandom the and equilibria constant these variable between relation a process the of equilibria If 1) 1. Observation (1 steprocess the As result. following the have we observations, two the Using h osatequilibria constant The ngnrlzn Eq. generalizing In t E E 0 iia aclto hw that shows calculation similar p p p p both where equilibrium p B } 1 2 2 1 − (p t t t t 0 t ∞ . . . where , < (p (p < (1 ) =0 d d 0 d d 0 −−−→ −−−→ 2 1 < d x n type and t t t D t < From < p →∞ →∞ t t < = = +1 < p < = ) ∞ → . − lim > +1 +1 hnwt rbblt ehave we 1 probability with then 1, 0 t sabuddmrigl,sbatnae rsupermartin- or submartingale, martingale, bounded a is p 0 p o n starting any For ≤ p p E E p aeywt niofriyba,if bias, anticonformity with namely 0, o any For E < ∞ p t < aeywt ofriyba,if bias, conformity with namely 0, − satisfies 0 0 0 | | . t t (D (D 1 4 (D h olwn eutgvsteanswer: the gives result following The ? →∞ < 2D + 1 p p {p < hra if whereas 0, )(2p 1 2 < then 1, 1 2 p 2 1 ihpoaiiy1w have we 1 probability with , bevto 2 Observation t t p hr sawy ovrec of convergence always is there , t . ) ) ) t t t +1 2 1 t then 1, E 1 2 ∗ then 0, = ) hnEq. then 0, = ) ) {p < < } coincide. then , hnby then , p 0 = t t epciey and respectively, A, > (p t ∞ =0 − t p p t − ,1 2, 1, 0, = {p } = t hnif then 0, t t t −2 0 +1 p t ∞ 1 2 1) if t E =0 n if and n,i o htcnb adaottelimit the about said be can what so, if and, ≤ and t p } (1 0 ∞ (p 2 1 = d < | t ∞ < eivsiaewehr if whether, investigate we 1, p =0 p < (t rmEq. From o all For < 0 0 < − t D p t ec,if Hence, 0. ihprobability with +1 {p p 1 2 ≤ ) ≤ p bevto 1 Observation p t ,1 2, 1, 0, = E p t 0. sbuddb n ,teLebesgue the 1, and 0 by bounded is p n h atthat fact the and ∗ . . . t > A t < = p d t t < t p 1, (p | − < 1 = } ∗ ) < 0 0 p p < and p 1 2 t ∞ and 1, ≥ 2 t n any and , 0 = < t t ≤ =0 0 1 2 +1 1 2 1 lim hrfr h in of signs the Therefore . t ) n if and , then 0, gives ≤ 1 p ,1 2 1, 0, = p > ihpoaiiy1. probability with ihpoaiiy1. probability with 1, orsodt xto ntype in fixation to correspond 1+2D + [1 ∞ , B | 0 eemndb Eq. by determined − 2 E t then 1, p p h process the . . . →∞ p < esethat see we (p t t 4p r qal represented. equally are uhthat such ∗ = ) If . E 0 1 2 d = t pcfidi Eq. in specified ) p t +1 2 1 , (p ≤ p E E (1 ∗ > ehave we 1. E p ∗ t 0 1 2 < . . . d = (p (p t x p t p | and , − +1 0 = Eq. 0, ≤ (p . t p = p t 0 ≤ t t p ` (1 2 1 t 0 +1 p −−−→ p t = ) t ≤ t ) | E +1 1 t , →∞ < {p 0 p 2 1 oa equilibrium. an to − sapolymorphic a is > ) p p (D p 0 | ≤ t t ≥ nal that entails < 0 then 1, E t p = ) | p ∗ t p p ≤ −−−→ 0 } p t 2 h process the 1, < t ≤ ∗ t 1 = t t 0, (p p t ∞ →∞ ) t )]. < ) 1. p =0 1 = t ) if > mle that implies p p 1 < ∞ 0 D ` < < p 0 t and , sthere Is 2. ihprob- with < d p . converges. t t ) hnif then 0, < n a and 1, p t E < p (p 2: r also are > changes If . t 1 2 for ∞ then 1, (p then , t n if and 0 2 1 − p t with and and +1 1 2 0 ∗ ` [3] [4] 1 2 ≤ = < = ) | iia if similar variable dom p p 2) 1) 4. Result values? these of each takes p Var(D Now if applies argument similar A 1. probability with p Proof: h rbblte f0and 0 of probabilities the p that implies which 3. Result Proof: 1)] By probability hrfr,since Therefore, If he aus0 ,and 1, 0, values three 12 But rbblt 1. probability 0 E ∞ 0 ∞ t ∗ < so , d Applying If eut3 Result o all for 0 = If If )(2p ≤ spsil fadol if only and if possible is −−−→ p 0 6= p p t d eut2, Result Var(p 0 →∞ t 0 ∞ 2 2 2 1 P E 1 2 , 0 = < (1 rmEq. From If < hnEq. then , < ∞ (1 p (p (p n yEq. by and Var(p P If ∗ 2 1 E − t p d If 1 2 p E eueEq. use we , − ∞ t − E hnalso then , E 0 nal htterno variable random the that entails 1 = 1; ,1 2, 1, 0, = (p +1 t 0 (p = < d = p +1 (p t Var(p 2, Result < [p 1) (p p < = p = ) t ∞ namely, = t P = ) E p ) ∞ σ p t t +1 ∞ and , = ) 2 t t 2 1 +1 1, ≤ 0 ∞ 2 1 2 −−−→ +1 +1 )=1 = 0) = E (D E (2p ) t , (1 dE E < σ 2(1 = ) →∞ 2 p σ E | E hsi hscase, this in Thus 0. since 2, = ) (D 8 (2p 2 ae w aus0and 0 values two takes E p ∞ t | | 2 1. − p (p and , (p t implies hnEq. then 0, = ) t p p 5, p [ > . . . ∞ p 2 [Var(p = ) t − p 0 = t t t E 2 p t 0, = ) ∞ t 0 ∗ = ) = ) ∞ p ) P (1 (1 ) 2 1 Eq. 0, ∞ 2 1) (p ≤ ∞ = and , −−−→ − o all for (1 − p (p t orsodn otecntn equilibria constant the to corresponding − osethat see to )(2p − or − →∞ 2 t t , D p Var(D p 2 1 1) + ) ∞ p − + t 2p ∞ E p t then −−−→ p htaetepoaiiisthat probabilities the are What . t t then 2 1 , 0 p t +1 ) 2 ∞ t →∞ ), sidpnetof independent is or 0 = E p ∞ 11 ∞ E [p 0 ≤ ) −−−→ r si Eq. in as are ≥ dE ∞ E t , 2 ) t (p (D →∞ | ∞ = − (2p 2 2 1 mle that implies ,1 2, 1, 0, = p (p 0 t )(2p p (2p https://doi.org/10.1073/pnas.2107204118 p ∞ (1 )p t [ t ∞ hrfr rmEq. from Therefore 1)]. ihpoaiiy1 and 1, probability with E ∞ 2 1 ihpoaiiy1 ec,Eq. Hence, 1. probability with p )]+ ∞ )p t = ) t t 2 2 P − (1 p or ∞ (1 − P Var(p (1 ) (1 ∞ t ∞ p (p mle that implies (1 p (p Var p − − ∞ 1) 2 − p p − − − ∞ ∞ = and ∞ ∞ 0 ∞ − (1 2 p 1) p If . p . . . )(2p )=0. 1)]= p (1 ∞ )=2p = 1) = 2 1 ∞ 1 2 ∞ 1 = t [E p t = + 2 − )(2p 1 = ) n as and , ) p t )(2p h ruetis argument The 13. − )(2p ) 2 0 Therefore . )(2p ∞ 0 = Var(p (p 2 1 p E (2p ∞ and < p 2 = ) ∞ p 1 = t [p ∞ a aeol the only take can t +1 − t ec h ran- the Hence . ∞ p . ∞ ) t − , t 2 1 t )(2p 2 0 (1 − o if Now 0. = 1)] − E . (2p − < | − t < 1)]. PNAS p 0 ). E p (p 1) 1), 0 − )=0 = 1) p − t )=0 = 1) 2 1 . ∞ (p ∞ )] 0 t 2 then , p +1 1. . < ∞ t − E − p | )(2p 1. 7 = ) ∞ | (p )=0 = 1) 1) f12 of 3 p (1 [10] [13] [14] [11] [12] with t t with 2 0 p = ) = ) [9] [5] [6] [7] [8] t p ∞ − ≤ − 0 . ,
ANTHROPOLOGY EVOLUTION If d = E(Dt ) 6= 0, the analysis is more complicated because in 3A), it is necessary that E[Dt (2)] < 0, in which case the dynamics this case it is not clear what E(p∞) is. We address this case using are similar to Eq. 1 with anticonformity (D < 0). With n = 3 and the notion developed by Karlin and Liberman (53) of stochastic identically distributed conformity coefficients, there is always ∞ local stability (SLS) of a constant equilibrium. Let {xt }t=0 be convergence of the population to an equilibrium. a sequence of random variables satisfying xt+1 = f (xt ) for t = If Dt (2) are not identically distributed, nonconvergence can ∗ 0, 1, 2, ... . Then x is a constant equilibrium of this sequence if occur with n = 3 role models. In Fig. 4A, Dt (2) alternates x ∗ satisfies x ∗ = f (x ∗). between two fixed values, producing a stable, two-generation Definition: SLS. A constant equilibrium x ∗ is said to be SLS if for cycle in the frequency of A. This cycle is confined to a region 1 1 any ε > 0 there is a δ > 0 such that in (0, 2 ) or ( 2 , 1), depending on whether the initial frequency p0 1 is below or above p = 2 , respectively. [Recall that Observation 1 ∗ ∗ |x0 − x | < δ =⇒ P lim xt = x ≥ 1 − ε. [15] does not rely on the assumption of identically distributed Dt (2).] t→∞
∞ More Role Models. The above results apply when the number of In our case, the sequence {pt }t=0 determined by Eq. 2 has three ∗ ∗ ∗ 1 role models is n = 3. Do such results for global convergence hold constant equilibria: p = 0, p = 1, and p = 2 . Following ref. ∗ n 53, to determine whether a constant equilibrium p is SLS we for any number of role models? To address this, we follow 2 examine the linear approximation of the transformation in Eq. 2 Boyd and Richerson (4) and Eq. above is replaced by “near” p∗. The linear approximations of Eq. 2 near p∗ = 0 and ∗ p = 1 for small νt are both n ! X Dt (j ) n j n−j pt+1 = pt + p (1 − pt ) . [19] ν = ν (1 − D ), [16] n j t t+1 t t j =0 ∗ 1 while that near p = 2 for small ηt is Eq. 19 with the subscript t removed from Dt (j ) is the gen- Dt ηt+1 = ηt (1 + 2 ). [17] eral form of Boyd and Richerson’s (4) deterministic model. The sum contains the probability of each possible sample j In Eq. 16, the frequency of a rare variant increases under anti- of n role models with variant A (a binomial, due to random conformity (Dt < 0), whereas in Eq. 17, the frequency of a 1 sampling) multiplied by the scaled conformity coefficient for common variant (ηt > ) increases under conformity (Dt > 0). D (j ) 2 that sample, t . For any generation t = 0, 1, 2, ... the vector Thus, we have the following: n Dt = [Dt (0), Dt (1), ... , Dt (n)] of conformity coefficients Dt (j ) j = 0, 1, 2, ... , n Result 5. has the following properties for : ∗ 1) If E [log(1 − Dt )]< 0, then both p = 0 and ∗ p = 1 are SLS. If E [log(1 − Dt )] > 0, then Dt (0) = Dt (n) = 0, Dt (n − j ) = −Dt (j ), [20a] P (limt→∞ pt = 0)= P (limt→∞ pt = 1)= 0. Dt ∗ 1 2) If E log(1 + 2 ) < 0, then p = 2 is SLS, whereas if Dt 1 E log(1 + 2 ) > 0, then P limt→∞ pt = 2 = 0. and for j = 1, 2, ... , n − 1, −j < Dt (j ) < n − j . [20b] x As both functions y = log(1 − x) and y = log(1 + 2 ) are con- cave (for x < 1 and x > −2, respectively), we apply Jensen’s Eq. 20b ensures that probabilities of acquiring A and B are in inequality to obtain (0, 1). The left-hand side of Eq. 20a entails that if a sample contains A(B) at frequency 1, an individual acquires A(B) with E [log(1 − Dt )] < log [E(1 − Dt )], [18] probability 1, and the right-hand side entails “equal treatment” Dt Dt E log(1 + 2 ) < log E(1 + 2 ) , of A and B. If an individual samples j of type A and n − j of type B, the amount by which it (anti)conforms to A, namely 2 since −2 < Dt < 1 with probability 1 and σ = Var(Dt ) > 0. Dt (j ), must be equal and opposite to the amount by which it Applying inequalities in Eq. 18 and Result 5, we have the (anti)conforms to B, −Dt (j ), for all probabilities to sum to 1. following: Therefore, when A rather than B is present in n − j role models, its conformity coefficient Dt (n − j ) must equal that for B in the Result 6. same scenario: −Dt (j ). Therefore, Eq. 19 is equivalent to ∗ ∗ 1) If d = E(Dt ) ≥ 0, then p = 0 and p = 1 are SLS. ∗ 1 2) If d = E(Dt ) ≤ 0, then p = is SLS. n ! 2 X Dt (j ) n h j n−j n−j j i ∗ pt+1 = pt + pt (1 − pt ) − pt (1 − pt ) , Therefore, if d = E(Dt ) = 0, all three constant equilibria, p = n j ∗ ∗ 1 j =k 0, p = 1, and p = are SLS. Hence p∞ can take the three 2 [21] values as stated in Result 4. The dynamics with n = 3 role models are illustrated in Figs. 1 n n+1 where k = 2 + 1 if n is even and k = 2 if n is odd. A, D, and G; 2A; and 3A. Dt is denoted by Dt (2) because it cor- Assume that the conformity coefficients change randomly at responds to the case where j = 2 of 3 role models have the same each generation t, and for any t = 0, 1, 2, ... , Dt is a vector of variant [the reason for this Dt (j ) notation becomes clear with random variables such that the Dt s are identically distributed and n > 4 and is described under Eq. 19]. Consistent with Observa- independent of the population state pt . The entries of Dt , Dt (j ), 1 1 1 1 tion 1, runs that start in (0, 2 ) or ( 2 , 1) remain in (0, 2 ) or ( 2 , 1), need not be identically distributed. Since probabilities of differ- ∗ 1 respectively. Equilibria are p = 0, 2 , and 1. It is possible that ent samples j of n, with corresponding Dt (j ), depend on pt , it is all three equilibria are SLS with E[Dt (2)] = 0, E[Dt (2)] < 0, or possible that higher or lower pt corresponds to greater or smaller ∗ 1 E[Dt (2)] > 0 (Fig. 1 A, D, and G, respectively). For p = 2 to differences between pt+1 and pt . be unstable (Fig. 2A), it is necessary that E[Dt (2)] > 0. In this Assume first that E(Dt ) = 0; i.e., E(Dt (j )) = 0 for all t = case, the dynamics are similar to those of deterministic Eq. 1 0, 1, 2, ... and j = 0, 1, 2, ... , n. In this case we have the with conformity (D > 0). For p∗ = 0 and 1 to be unstable (Fig. following result:
4 of 12 | PNAS Denton et al. https://doi.org/10.1073/pnas.2107204118 On randomly changing conformity bias in cultural transmission Downloaded by guest on September 30, 2021 Downloaded by guest on September 30, 2021 hn if Then, for sabuddmrigl n h atnaeconvergence martingale the and martingale that bounded implies theorem a is ihprobability with 0.5. SD and −0.05 truncation) (before mean with ( on distribution normal 2, 1, 0, Proof: nrnol hnigcnomt isi utrltransmission cultural in bias conformity changing randomly On al. et Denton probability with D probability with with equilibria although shown, are 1. Fig. 2) 1) 7. Result t (4) Let probability If, probability process the If edn admvariables, random pendent E ∼ j [D naddition, in E U ,2, 1, = . . . If tcatclclsaiiyof stability local Stochastic t σ ,0] (−4, (D (j j 2 )] Eq. , D = h eno log[1 of mean the , t = ) D t Var . . . t 1. 1. ihprobability with (k sidpnetof independent is 5 4 2 3 3 1 {p 0, 19 (B) . and and D , ), (D t t D n hnsatn rmayiiilstate initial any from starting then G) ( 0.5. SD and 0.05 truncation) (before mean with 2) (−3, on distribution normal truncated a from sampled is (3) } mle that implies D t t ∞ − t D D n =0 (k ,1 ihma bfr truncation) (before mean with 1) −2, t (j t t lim (k = (4) (2) 1 ), )) 4, converges; 1), + = = D t [σ →∞ D etepstv aineof variance positive the be ihprobability with −3 ihprobability with −1.5 t 5adS ..(F 0.5. SD and −0.05 t (k 0 p* (3) − 5 1 = then . . . . p D = 1), + ∼ p* D E t t σ (n 0, t U = n , = (3) (p ,1 ihprobability with 1) [0, D p − 2 1 0 = p htis, that 0, ∞ a aelne ob ece;hwotnec qiiru a ece sin is reached was equilibrium each often how reached; be to longer take may 1 , p ∼ t . . . t ∞ n h eno log of mean the and 1)], +1 (n t 1 2 U ae h values the takes .Ec ltsos2 uswt initial with runs 25 shows plot Each 1. , ihpoaiiy1. probability with , ,2 ihprobability with 2) [0, and | − D as p t t 1) )= lim (n 5 1 ) D E . ) n p 3 1 t D t r independent are (D = →∞ o any for ( . t 0 = (0) t Hence . (3) H 5, t ) = ) D = n p t t = ihprobability with 1.5 ,1 ihma bfr truncation) (before mean with 1) (−4, on distribution normal truncated a from sampled is (4) 5adS ..(E 0.5. SD and −0.05 D 0, 4 3 = 0 t 0 4, t and (n r inde- are ≤ p 2 1 {p for D ∞ 1 , h t p 0]. = ) 5 3 (3) D 1 t D 0 } and + with with t t ≤ t ∞ t = (3) (j =0 = D 9wt probability with 0.99 1 ) t D 2 ∼ (2) t (3) U i ,0] (−3, for admvrals Eq. that variables, random ic Var(p have Since ∼ Var(p + U 3 5 ) n ,0] (−3, n and = n X p j Var(p = = =k ihprobability with −1 0 .(A) 3. = 4, n X j D t =k +1 −1 n σ 5(e)ad2 uswith runs 25 and (red) 0.25 t D (3) ihprobability with j 2 2 t t ,1) (−3, on distribution normal truncated a from sampled is (3) = ) n +1
n σ = = j 2 2 n 3 4 t j 5wt probability with −2.85 | +1
Var(p 3, ! and p n t j D 2 = ) ) ! E t (2) D 2 t 1 4 t Var (3) h (C . ∼ ) p h p t U j = (E 5 2 ) (1 t j ,1 ihprobability with 1) [0, (D) . ihprobability with −2.5 (1 n 21 (p − = − 5, t p n +1 p t mle o any for implies = D ) t https://doi.org/10.1073/pnas.2107204118 n ) t | 3, p (4) n −j p 0 5 2 −j D t = . ∼ ))+ t − 2 ssmldfo truncated a from sampled is (2) 5(le.Egt generations Eighty (blue). 0.75 − U p eto A section Appendix, SI [ ,1 ihprobability with 1) 0, p t n E t n −j −j (Varp (1 3 2 (1 1 4 ( . − and − I ) n p n t = p t t D +1 PNAS ) = t ,1 2, 1, 0, = t j 3, ) (2) i 5, j 2 | i D . p 2 D ∼ t t (2) t | )) (4) U . we , ,0] (−2, = f12 of 5 5 4 along = [23] [22] 0.99 and . . . 0.9
ANTHROPOLOGY EVOLUTION 1 Fig. 2. Stochastic local stability of p* = 0, 1 and instability of p* = 2 . Each plot shows 25 runs with initial p0 = 0.499 (red) and 25 runs with p0 = 0.501 (blue). The first 80 generations are shown, although equilibria may take longer to be reached; how often each equilibrium was reached is in SI Appendix, h Dt (2) i section A, along with E[Dt (j)], the mean of log[1 − Dt (n − 1)], and the mean of log 1 + 2 for n = 3. (A) n = 3, Dt (2) ∼ U[0, 1) with probability 0.9 and Dt (2) ∼ U(−2, 0] with probability 0.1. (B) n = 4, Dt (3) ∼ U[0, 1) with probability 0.925 and Dt (3) ∼ U(−3, 0] with probability 0.075. (C) n = 5, Dt (4) ∼ U[0, 1) with probability 0.98 and Dt (4) ∼ U(−4, 0] with probability 0.02. Dt (3) ∼ U[0, 2) with probability 0.55 and Dt (3) ∼ U(−3, 0] with probability 0.45. (D) n = 4, 3 1 3 1 Dt (3) = 0.99 with probability 4 and Dt (3) = −2.99 with probability 4 .(E) n = 5, Dt (4) = 0.99 with probability 4 and Dt (4) = −3.99 with probability 4 . 3 1 Dt (3) = 0.1 with probability 4 and Dt (3) = −0.1 with probability 4 .
∗ 1 Now as pt −−−→ p∞ with probability 1, by the Lebesgue con- near the constant equilibrium p = given in ref. 49, t→∞ 2 ` ` equation B8: vergence theorem E(pt ) −−−→ E(p∞) for any ` = 1, 2, ... , and t→∞ Eq. 23 implies that n−1 ! 0 n−2 X D(j ) n ν = ν 1 + 1 ((2j − n)) = νφ . [25] 2 n j n j =k n−1 !2 2 2 X σj n h j n−j n−j j i E p∞(1 − p∞) − p∞ (1 − p∞) = 0. Ref. 49’s table S1 shows that depending on n, φn can be negative, n2 j ∗ j =k in which case log φn is not defined and conditions for SLS of p = 1 [24] 2 cannot be determined. Also, Observation 1 does not hold. With n > 3 role models and E(Dt ) =6 0, there need not be j n−j n−j j convergence to an equilibrium (Fig. 5). Below we determine con- Hence p∞(1 − p∞) = p∞ (1 − p∞) for j = k, k + ditions for convergence to an equilibrium with n = 4 role models 1, ... , n − 1 and p∞ = 0, p∞ = 1, or p∞ = 1 − p∞ in which case 1 1 and, subsequently, a general number n of role models. p∞ = 2 . Thus, with probability 1, p∞ takes values 0, 2 , 1. ∗ ∗ The following result concerning the SLS of p = 0 and p = 1 Example: The Case of n = 4 Role Models. Recall that the recursion holds for E(Dt ) = 0 as well as E(Dt ) 6= 0. with fixed conformity coefficients is the same for n = 3 and 4 (4). However, with random conformity coefficients the case n = 4 Result 8. exhibits different evolutionary dynamics from those with n = 3 role models. When n = 4, we have five conformity coefficients ∗ ∗ Dt (j ) for j = 0, 1, ... , 4, Eq. 20a and inequality in Eq. 20b, 1) If E [log (1 − Dt (n − 1))] < 0, then p = 0 and p = 1 are SLS. 2) If E [log (1 − Dt (n − 1))] > 0, then P (limt→∞ pt = 0) = 0 Dt (0) = Dt (2) = Dt (4) = 0, Dt (1) = −Dt (3), and P (limt→∞ pt = 1)= 0. [26] − 3 < Dt (3) < 1.
The proof is similar to that of Result 5 based on the linear Hence, from Eq. 21 with n = 4 and k = 3, we have approximation of Eq. 19 near p∗ = 0 and p∗ = 1 in equation B2 of Denton et al. (49) and the assumption that Dt (n − 1) < 1 pt+1 = pt + Dt pt (1 − pt )(2pt − 1), [27] by Eq. 20b. If E (Dt (n − 1))> 0, then using Jensen’s inequality, ∗ ∗ E [log (1 − Dt (n − 1))]< 0, so p = 0 and p = 1 are SLS. where Dt are identically distributed and correspond to Dt (3) We are unable to obtain corresponding results for the in Eq. 26. Therefore in this case [1 + 2Dt pt (1 − pt )] with ∗ 1 1 SLS of p = 2 . Consider the linear approximation of Eq. 19 −3 < Dt < 1 can be negative, and by Eq. 3, pt+1 − 2 and
6 of 12 | PNAS Denton et al. https://doi.org/10.1073/pnas.2107204118 On randomly changing conformity bias in cultural transmission Downloaded by guest on September 30, 2021 Downloaded by guest on September 30, 2021 nrnol hnigcnomt isi utrltransmission cultural in bias conformity changing randomly On al. et Denton (C generations. Now Eq. of when convergence convergence global be global d still of there proof Will hold. the not in does factor key a was A 10 for ran 4. Fig. and 0.05 probability reached. equilibrium The (blue). 3. Fig. ..(D) 0.9. n p (A) . = = t Since 5, − 3 E p D n efcson focus we Q (D 2 1 t n t = xc yls ahpo hw 5rn ihinitial with runs 25 shows plot Each cycles. Exact tcatclclsaiiyof stability local Stochastic (1 (4) t eto A section Appendix, SI = 6 +1 a aedfeetsgs Therefore signs. different have can t 3, p ) eeain;n qiirawr ece.Dsrpin ftednmc,eatiiilfeunis and frequencies, initial exact dynamics, the of Descriptions reached. were equilibria no generations; − ∼ 5, p t D 6 0 = = +1 t U D p t +1 2 lentsbetween alternates (2) ) ,1 ihpoaiiy04and 0.4 probability with 1) [0, t t Q (4) n = ) − with − = t = 2 1 5, 2D + 1 2 1 D Q 1 4 ihprobability with −3.9 t D − n t and (2) 2 p* t (p = (4) = 4 = Q ∼ = t = U t n Eq. and , p ? 2 1 p t ,0] (−2, 9and −3.99 t t − srahdi vr ae lhuhol h rt8 eeain r hw n hseulbimmytk ogrt be to longer take may equilibrium this and shown are generations 80 first the only although case, every in reached is 1 4 − reports − − 1 2 2 1 p* 2 1 ) Q 2 ihpoaiiy09.(B) 0.95. probability with 2 5ad09eeyohrgnrto.(B) generation. other every 0.9 and −1.15 Then . = t a aedfeetsgs from signs, different have can 1+2D + [1 28 E 2 1 D 2 [ D t , n ntblt of instability and seuvln to equivalent is (3) 4 3 t (j D , h eno log[1 of mean the )], D t = t (4) t t ,1 2, 1, 0, = (4) vr hr eeainadi h te generations other the in and generation third every −2.1 p ∼ which 1, Observation t = (1 U ihprobability with 0.9 ,0] (−4, − p p 0 t . . . < )] p* ihpoaiiy0.6. probability with 2 2 1 {p . n = rd n 5rn with runs 25 and (red) = t ,1 ahpo hw 5rn ihinitial with runs 25 shows plot Each 1. 0, } n 4, − when [28] [29] 3 = D D t t (n (3) , 4 1 − ∼ , D n h eno log of mean the and 1)], U t σ Since 9. Result process the to applies result following The Proof: {Q then (3) ,1 ihpoaiiy02 and 0.25 probability with 1) [0, 2 n = D t = = } t Q (3) ihprobability with 0.9 4, t ∞ Var(D Q D =0 ∞ ert Eq. Rewrite D p t ∼ t 0 +1 t sidpnetof independent is (3) a aeol w values, two only take can If ovre ihprobability with converges > U ,2 ihpoaiiy01and 0.1 probability with 2) [0, = = 2 1 t d ehave we ), bu) ihygnrtosaeson lhuhsimulations although shown, are generations Eighty (blue). Q 9eeytidgnrto and generation third every −2.99 = t E 4D + 1 (D 29 h t 1 ) as + 4 3 > and , D t D p t 0, t (4) 2 0 (2) p E 4 1 = t D hnfrany for then [D i D = − t n oas of also so and 0 rd n 5rn with runs 25 and (red) 0.001 for t (3) t (3) 9and 0.99 (j Q )] https://doi.org/10.1073/pnas.2107204118 ∼ n t = r eotdin reported are Q = U ihprobability with −1.9 4 + ∞ 1, ,0] (−3, D .(A) 3. t 0 = (3) D n if and D t t (3) 2 ∼ 0 n or ihpoaiiy07.(C 0.75. probability with {Q ≤ U = D = Q 4 1 ,0] (−3, t Q 0.5. (3) Q − t 3, Q t if , ∞ } 0 section Appendix, SI Q D ∞ = t ∞ ≤ =0 t = d t (2) 9i h other the in 0.99 lim = ihprobability with PNAS 4 1 = : 2 4 1 ∼ . h process the E . U 4 1 p (D . ,1 with 1) [0, t 0 →∞ | = t f12 of 7 ) 0.999 [30] and Q t ) ,
ANTHROPOLOGY EVOLUTION Fig. 5. Stochastic fluctuation. Each plot shows 25 runs with initial p0 = 0.49 (red) and 25 runs with p0 = 0.51 (blue). Eighty generations are shown, although 6 simulations ran for 10 generations; no equilibria were reached. Mean frequencies and E[Dt (j)] are reported in SI Appendix, section A. (A) n = 5, Dt (4) = −3.9 1 1 1 1 with probability 2 , Dt (4) = −3.4 with probability 2 , Dt (3) = −2.9 with probability 2 , and Dt (3) = −2.4 with probability 2 .(B) n = 5, Dt (4) = −0.6 with 1 1 1 1 probability 2 , Dt (4) = −0.8 with probability 2 , Dt (3) = 1.85 with probability 2 , and Dt (3) = 1.95 with probability 2 .
E (Q | Q ) 1 1 d σ2 t+1 t As 0 ≤ − Q∞ ≤ and d 1 + + < 0, we have [31] 4 4 4 4 1 2 2 1 2 = Qt 1 + 4d 4 − Qt + 4 σ + d 4 − Qt .
If d > 0, then E (Qt+1 | Qt )≥ Qt for t = 0, 1, 2, ... . Thus 1 h d σ2 i ∞ 1 0 ≤ 4E Q∞ 4 − Q∞ d 1 + 4 + 4 ≤ 0. [36] {Qt }t=0 is a submartingale and since 0 ≤ Qt ≤ 4 , we can apply the martingale convergence theorem to deduce that with ∞ probability 1, {Qt }t=0 converges to Q∞. Also applying the ` ` 1 Lebesgue convergence theorem, E Qt −−−→ E Q∞ for any Inequalities in Eq. 36 imply that E Q∞ − Q∞ = 0 and t→∞ 4 1 1 ` = 1, 2, ... . Q∞ 4 − Q∞ = 0 so that Q∞ = 0 or Q∞ = 4 with probabil- Thus Eq. 31 entails that ity 1. Dynamics with n = 4 role models are illustrated in Figs. 1 B, E, and H; 2 B and D; and 3B. The conformity coef- 1 E(Q∞) = E(Q∞) + 4dE Q∞ − Q∞ ficient, Dt , is denoted by Dt (3) (Eqs. 26 and 27). Because 4 [32] 2 2 1 2 Observation 1 no longer holds with n > 3, runs that start in + 4 σ + d E Q∞ 4 − Q∞ . 1 1 1 (0, 2 ) (red) or ( 2 , 1) (blue) can cross the p = 2 line (e.g., 1 Fig. 1B). Fig. 1 B, E, and H shows examples in which all Since d > 0, we deduce from Eq. 32 that E Q∞ − Q∞ = 0 ∗ 1 4 three equilibria p = 0, , and 1 are SLS with E[Dt (3)] = 0, 1 2 1 2 and E Q∞ 4 − Q∞ = 0. But 0 ≤ Q∞ ≤ 4 with probability 1. E[Dt (3)] < 0, and E[Dt (3)] > 0, respectively. Fig. 2 B and D Q 1 − Q = 0 Q ∗ 1 Hence ∞ 4 ∞ with probability 1 and ∞ can take shows that p = 2 can be unstable with either E[Dt (3)] > 0 or 1 E[Dt (3)] < 0 E[Dt (3)] < 0 only the two values Q∞ = 0 or Q∞ = 4 . . However, is necessary for instability 1 2 1 of p∗ = 0 and 1 (Fig. 3B and Result 8). Here, the dynam- If Q∞ = 0, then limt→∞ 2 − pt = 0 and limt→∞ pt = 2 . 1 ics are similar to those of Eq. 1 with anticonformity (D < If Q∞ = 4 , then limt→∞ pt (1 − pt ) = 0. Therefore for t large, ∗ 1 ∗ ∗ 0), apart from fluctuations across p = before reaching this either pt is near p = 0 or pt is near p = 1. But if d > 0, 2 ∗ ∗ 1 equilibrium. both p = 0 and p = 1 are SLS. Therefore, if Q∞ = 4 , either Removing the assumption that Dt (3) are identically dis- limt→∞ pt = 0 or limt→∞ pt = 1. ∞ tributed, Fig. 4B shows an example in which Dt (3) takes one fixed Conclusion. If E (Dt )> 0, then the process {pt } converges t=0 value every third generation and another fixed value in the other with probability 1. If p∞ = limt→∞ pt , then p∞ can assume only 1 generations. This system produces an exact six-generation cycle the three values 0, 1, 2 . in the frequency pt . Unlike in the case of n = 3 (Fig. 4A), this The situation is more complicated when d = E(Dt ) < 0. Eq. 31 cycle occurs around p = 1 because Observation 1 does not hold. can be rewritten as E (Qt+1 | Qt )= Qt (1 + ut ) with 2