Comeback kids: an evolutionary approach of the long-run innovation process
CComeback kids: an evolutionary approach of thelong-run innovation process ∗ Shidong Wang † Mathematical Institute, University of OxfordAndrew Wiles Building, Radcliffe Observatory Quarter, Woodstock Road, OX2 6CG, Oxford, United Kingdom
Renaud Foucart ‡ Department of Economics, Humboldt UniversityUnter den Linden 6, 10099 Berlin , Germany
Cheng Wan § Department of Economics; Nuffield College, University of OxfordNew Road, OX1 1NF, Oxford, United Kingdom
Abstract
We provide a theoretical framework to understand when firms may benefit from ex-ploiting previously abandoned technologies and brands. We model for the long run pro-cess of innovation, allowing for sustainable diversity and comebacks of old brands andtechnologies. We present two extensions to the logistic and Lotka-Volterra equations,which describe the diffusion of an innovation. First, we extend the short-term competi-tion to a long-term process characterized by a sequence of innovations and substitutions.Second, by allowing the substitutions to be incomplete, we extend the one-dimensionalprocess to a tree-form multidimensional one featuring diversification throughout the long-term development.
Keywords. competition, migration, brand rejuvenation, innovation diffusion, diversification,long-term development tree, birth-death process ∗ July 24, 2016 † Email: [email protected] , S. Wang is supported by a EPSRC Grant EP/I01361X/1 atthe University of Oxford ‡ Email: [email protected] § Corresponding author.
Email: [email protected] a r X i v : . [ q -f i n . E C ] J u l here is an Indian proverb that goes, “Sit on the bank of a river and wait: Your enemy’scorpse will soon float by” The process of innovation is traditionally seen as a sequence of one-shot games, in which anew (and better) product or technology replaces an outdated standard. As discussed at lengthbelow, examples abound however, of declining brands and technological standards who cameback from the dead following the arrival of a disruptive innovation. A very timely one is the“comeback of vinyl.” While this technology had been almost abandoned in the late nineties,vinyls are now the fastest growing area of music sales, with very positive future prospects,to the point that engineers predict that the vinyl record has come back to stay as the onlyanalogue medium (Bartmanski and Woodward, 2013).While the appeal to nostalgia might be a driver of successful and temporary brand rejuve-nation, little is known about the underlying process of innovation that make long lastingcomebacks possible. We aim at providing a theoretical framework to understand a feature ofinnovation that is largely overlooked: when a new technology arrives and makes a previouslydominant one obsolete, it may leave room for an increase in the demand for an even oldertechnology.We use stochastic dynamics in order to provide the micro-foundations for a deterministicmodeling of the long-term process of innovation, competition and market diversification, basedon the idea of “close competitors.” To keep the example of the vinyl, we claim to be able torecover the market evolution by the fact that, while compact discs were a direct competitorto vinyl, digital music is much less so. Consumers indeed claim to retrieve satisfaction fromthe consumption of a vinyl that cannot be obtained using a dematerialized device. Hence, asthe market for CD gradually shrinks to be replaced by digitalized music (a direct competitorof CD), a market space opens for a limited but sustainable comeback of the vinyl. Studies inpsychology, sociology and anthropology have shown that the choice to buy a new album invinyl goes beyond simply being matters of nostalgia or fetish. It is precisely because musichas become mostly dematerialized that more and more consumers buy new albums in whatthey perceive to be the most tangible format, the vinyl (Magaudda, 2011, Bartmanski andWoodward, 2013, Negus, 2015).When a brand or technology becomes dominant and starts gaining market shares at theexpense of a previously dominant one, our results suggest that one could benefit from goinga step backward, and wonder what the previously dominant one replaced. If this abandonedbrand or technology vanished due to better substitute features that are also present in thenew dominant one, there is no hope for a long lasting return. If the new dominant brand ortechnology does have such features, a successful comeback is possible.Extensive research in business and economics has been carried out on the diffusion of innova-tion, with objective to understand the diffusion of new technologies over the course of industrialhistory (Griliches, 1957, 1960, Mansfield, 1961, Geroski, 2000, Young, 2009, Peres et al. , 2010).More recent models in business and economics have studied simultaneous launches and co- Umberto Eco, “the Name of the Rose,” p.550 et al. , 2009, Yan and Ma, 2011 and Guseoand Mortarino, 2014), including applications to the case of the music industry (Guidolin andGuseo, 2015). None of them however studies the conditions for a fading technology to comeback.In other contexts, the subject of innovation can be a product, an idea, a practice, a phe-nomenon, a social norm or convention, or even a religion. In this paper, we call the subjectin question an “alternative.” For example, market analysts wish to predict and influence howa new product gradually occupies the market or how old ones vanish from it (Mahajan andMuller, 1979, Mahajan et al. , 1990, Parker, 1994, Chandrasekaran and Tellis, 2007); commu-nication specialists try to understand the mechanism of diffusion in order to intervene in thepropagation of an idea or information; anthropologists and sociologists investigate on how aparticular practice spreads from one tribe, culture or region to another (Katz et al. , 1963,Brown, 1981, Clark, 1984, Rogers, 2003); and political scientists are interested in the adoptionof new policies among different states.In most of the cases, the innovation does not arise out of nothing. One or several olderalternatives with similar functions may well have existed before the arrival of the innovation,though the innovation can be more advanced or sophisticated and ready to be accepted by alarger population. In this sense, we talk rather of “substitution” than “diffusion.” For example,synthetic fibers replaced natural fibers in cloth making, and the diesel locomotive replaced thesteam locomotive in railroad transportation.The seminal work of Mansfield (1961) initiates the rigorous mathematical modeling of thissubstitution process, in particular in economics research. Fisher and Pry (1971) provides anempirical logistic S-type equation to describe the evolution of the market occupation rate ofan innovation. A lot of studies follow this approach by extending and improving the model(Blackman, 1972, 1974, Bretschneider and Mahajan, 1980; Easingwood et al. , 1981, Sharifand Kabir, 1976a,b; Sharif and Ramanathan, 1981, Skiadas, 1985, etc.) A typical logisticdifferential equation is(1.1) ˙ n = bn ( N − n ) , where n is the market share of the alternative in question, N is the upper bound of the marketshare, b is the coefficient of imitation or other mechanism leading to the adoption of thealternative by the population. In particular, b can be a constant in a most elementary modelor a function of n in more sophisticated models.Another approach consists in modeling the competition between the old and the new alter-natives in an explicit way, by constructing a Lotka-Volterra dynamical system (Batten, 1982,Karmeshu et al. , 1985, Bhargava, 1989). For example, let m (resp. n ) represent the numberof individuals adopting the old alternative x (resp. the new alternative y ). Then, a system ofLotka-Volterra equations governing their competitive relation is:(1.2) (cid:40) ˙ m = α m ( M − m − βn ) − γ m, ˙ n = α n ( N − n − βm ) − γ n. Here, α is the rate of adoption, γ the rate of abandonment and M the upper bound of themarket capacity for x ; the notations are similar for y . Compared with the logistic equation(1.1), the parameter β introduces, in addition, the competition between x and y .3oth logistic and Lotka-Volterra equations were originally introduced in the study of animalpopulation or epidemics, where they remain an important tool in the study of populations(see for instance Lafferty et al. , 2015). They are applied by analogy in the model of innovationdiffusion and substitution in economics (Pistorius and Utterback, 1997), and are supportedby empirical studies on the macroscopic level (the population level). In this framework, aninnovation eventually becomes old, then a newer innovation appears and another round ofcompetition takes place.In actual markets, the process is however more subtle. The substitution of a new alternative forthe old one is not always complete, and the coexistence of several generations of technologiesis not rare. Moreover, when the environment of the competition changes and, in particular,with the periodical arrival of innovations, the characteristic of the current competition maybe influenced and changed as well.We provide real world examples of such processes of coexistence and comebacks of alternativesin Section 2. We present in Section 3 an individual-based foundation of the logistic substi-tution and Lotka-Volterra competition processes, and then extend it to a long-term processcharacterized by a sequence of innovations, competitions and substitutions. This recovers thewell-known representation of innovation as a succession of one-shot competitions between twoalternatives (Champagnat, 2006). We then extend in Section 4 the model to a tree-form one,in which several alternatives survive after each short-term competition and, in the long run,innovations drive the growth of a “development tree.”We use a stochastic birth-death model as the foundation of the deterministic logistic andLotka-Volterra equations. As remarked by Fisher and Pry (1971) on their logistic diffusionmodel, deterministic models are “not to be applied to substitutions prior to their achieving amagnitude of a few percent, at which time a definite growth pattern is established and thevery early history has little effect upon the trend extrapolation.” The deterministic, continuouslogistic equation and Lotka-Volterra dynamical system on the aggregate level is a macroscopicapproximation of the stochastic, discrete process on the individual level. The approximationis therefore valid only if the size of the population is large enough so that the fluctuation oruncertainty on the individual level can be averaged. According to Nielsen Soundscan, more than 9.2 million vinyl records were sold in the U.S. in2014, marking a 52% increase over the year before, the highest numbers recorded by SoundScansince the music industry monitor started tracking them back in 1991. These numbers are likelyto be under-evaluated, as a large share of vinyl albums are purchased at independent recordstores who do not necessarily report their figures. Depending on the sources, this represents The facts in this subsection are borrowed from: Eric Felten, “It’s Alive! Vinyl Makes a Comeback,” theWall Street Journal, January 27, 2012, Megan Gibson, “Here’s Why Music Lovers Are Turning to Vinyl andDropping Digital,” The Time, January 13, 2015, Allan Kozinn, “Weaned on CDs, They’re Reaching for Vinyl,”The New York Times, June 9, 2013, and Glenn Peoples and Russ Crupnick, “The True Story of How VinylSpun Its Way Back From Near-Extinction,” Billboard, December 17, 2014 thecloud ,” and that high involvement in music is connected to a perception of tangible records asmore valuable (Styvén, 2010). And, indeed, when talking about the advantages of vinyl, thebenchmark of today’s mass consumption of music is not the CD anymore, but digital music. In
Billboard magazine, Gleen Peoples and Russ Crupnick relate a research study carried on behalfof the Music Business Association in 2006, when the industry was trying to save the marketfor physical records from the competition of illegal downloads. Consumers were surveyed overa number of innovating products, such as interactive and connected CDs, branded cards witha download code, or music placed on branded storage devices. Among these alternatives wasa “premium” vinyl including a digital download card. Of all these products, vinyl tested theworse, and didn’t particularly appeal to any important fan segments. Almost ten years later,all these relatively successfully tested products became, in the best case, very small niches,while premium vinyl records are now widely sold.Our claim is that in 2006 the music industry was looking for an alternative able to improveover the CD standard. Today’s vinyl success comes from the fact that it is not so muchcompared to CD anymore, but mostly to digital music. In the words used in the paper, CDwas a close competitor of vinyl, and a close competitor of digital music. However, vinyl is nota close competitor of digital music, and therefore as CD is supplanted by digital music, thereis room for a comeback of vinyl.
A private company can benefit from understanding the impact of new technologies on a possiblecomeback of previous standards. In 2003, the once flourishing Danish company Lego wasvirtually out of cash, on its way to bankruptcy. The decay of the company began in the mid-nineties, when children started to abandon the traditional building blocks in favor of moresophisticated toys. It was a time of video games, electronic and flashy innovations, or cheaperand more disposable toys. The strategy of Lego was therefore to mimic this trend, producinga large variety of innovating products, without a lot of commercial success.In 2005 however, the company decided to return to its basics, simplifying its product line byreducing the number of components produced in its factories from 12,400 to 7,000, and goingback to its core principle of building bricks. This strategy seems to have largely paid. Lego isagain a profitable company and, from 2012 to 2015, Lego’s sales have gone up an average of24% annually and profits have grown by 41%.Our claim is that part of this success can be explained by innovations in the entertainmentindustry, leaving more and more children playing on dematerialized games, such as computersor tablets. Consumers in search of easy and innovative toys started to prefer those to a number The facts in this subsection are borrowed from: Craig McLean, “Lego, play it again,” The Telegraph,December 17, 2009, “Innovation Almost Bankrupted Lego - Until It Rebuilt with a Better Blueprint,” Time,July 23, 2012, Jay Greene, “How LEGO revived its brand,” Business Week, July 23, 2010, and Nick Watt andHana Karar, “The land where Lego comes to life,” Abc News, November 16, 2009.
5f relatively short-living physical toys. However, this substitution opened a new avenue forlong-lasting, high quality, physical toys such as Lego’s, which are not “just on a screen” andno close competitors to tablets or software.
In the early seventies, following the peak in the prices of fossil fuels, most developed countriesstarted to massively invest in nuclear power plants. The technology was rather recent and, atthe time, relatively cheaper and socially acceptable, so that in the early 2000s, countries likeFrance, Sweden or Belgium had more than of their electricity coming from nuclear sources(with almost for France). One of the drawbacks of nuclear power as a “baseload power”however, is that it takes time to turn on and off, and therefore needs to be complemented bymore flexible sources of energy.As environmental concerns become more and more important and the social acceptabilityof nuclear power decreases, renewable technologies are pushed forward in order to becomecompetitive on the electricity market, with the ambition of making it the baseload power ina large number of developed countries. An important characteristic of renewable electricitysources is their intermittency, which makes them less and less compatible with nuclear powerplants as soon as renewable electricity represents a higher share of the production. Hence,with the technologies available today, a direct consequence of higher shares of renewable isthat it increases the cost of nuclear power and makes the more flexible sources such as gaspower plants much more attractive, even more so than as a complement of nuclear power.
A last example can be found in the release of the Nintendo 64 in the late nineties. Nintendowas a direct competitor of Sega and both were the remaining historical producers of video gameconsoles, sharing the common characteristics of having very identifiable characters followingfrom one platform to the other (for instance, Mario for Nintendo and Sonic for Sega). In1994, both Sega and Nintendo had planned to release a new platform. The Sega Saturn wasinnovating by providing a new technology: video games were produced on CD instead of romcartridges, allowing for cheaper and much more flexible production of games, at the cost ofbeing slower to load. However, two unexpected elements arrived. First, a new competitor(Sony), launched a high quality console based on the CD technology (the Playstation), whichwas quickly adopted as the new standard instead of the Sega Saturn. Second, for technicalreasons, Nintendo had to delay the release of its new console. When the Nintendo 64 wasreleased two years behind schedule, the Sega Saturn had failed to become a standard, andNintendo was competing mostly against the Playstation. At this time, the two products were The facts in this subsection are borrowed from Neuhoff (2005), Verbruggen (2008), Sovacool (2009),Eurostat (2013) and data available on the website of the world nuclear association (world-nuclear.org) The facts in this subsection are borrowed from: Lawrence Fisher, “Nintendo Delays Introduction Of Ultra64 Video-Game Player,” The New York Times, May 6, 1995, Mark Langshaw, “Sony PlayStation vs Nintendo64: Gaming’s Greatest Rivalries,” Digital Spy, December 9, 2012, Michael Krantz, “Super Mario’s dazzlingcomeback,” Time, June 24, 2001 and Cesar Bacani and Murakami Mutsuko, “Nintendo’s new 64-bit platformsets off a scramble for market share,” Asia Week, April 18, 1997.
This subsection constructs a stochastic birth-death model describing the short-term compe-tition between alternatives on the individual level, which leads to the substitution of a newalternative (i.e. innovation) for the old one. This is the elementary model and the base of thepaper. We discuss two particular examples, which correspond to the logistic diffusion modeland the Lotka-Volterra competition model. The next subsection introduces a long-term pro-cess made up of successive innovations sparsely distributed along the time line, each followedby an elementary short-term competition/substitution process.The mathematical tools used in this section are borrowed from population dynamics studiedby probability theorists. We first describe the original model and adapt it to our context.Then we introduce and interpret the analytical properties of the dynamics.Alternatives achieving the same goal or having similar functions compete with each otherwhen they are both present in the market. An alternative is characterized by its trait , whosevalue belongs to a given trait space X . For example, coal, oil, gas, hydroelectric power andnuclear have distinct traits as energy resources. An individual, who can be a person, a firm,etc., is a (potential) adopter of an alternative. We consider, for each trait, the number ofindividuals having adopted an alternative of that trait. To fit into the language of birth-deathprocess, we call the event whereby an individual who has not yet adopted any alternativeadopts an alternative of trait x “the birth of an individual of trait x .” For example, it canbe the entry of a firm into the market. Conversely, the “death of an individual of trait x ”represents the event that an individual actually holding an alternative of trait x abandonsthe trait while not adopting another one. For example, it can be the exit of a firm from themarket, either for a reason linked only to its own performance, or because of its loss of marketshare due to competition against better-equipped firms. Finally, “mutation” means the birthof an individual with a new trait that has never existed before, which is interpreted as thearrival of an innovation.Recall that in continuous-time birth-death processes, an individual gives birth or dies at inde-pendent, random exponential times. Besides, the reproduction rate, death rate and mutationrate vary among different traits.Now let us consider a multi-trait birth-death process with mutations. At any time t , thepopulation is composed of a finite number I t of individuals, respectively characterized by theirtrait x ( t ) , . . . , x I t ( t ) . The traits belong to a given closed subset of R d , called the trait space X . Let δ x stand for the Dirac distribution concentrated on x . The population state at time t
7s specified by the counting measure on X ν t = I t (cid:88) i =1 δ X i ( t ) . To write the dynamics of the process ( ν t ) t> , let us introduce the following parameters: • b ( x ) is the clonal birth rate from an individual with trait x , who reproduces an offspringwith the same trait as its parent. • d ( x ) is the natural death rate of an individual with trait x . • α ( x, y ) is the competition kernel felt by some individual with trait x from another indi-vidual with trait y , which results in the death of an individual with trait x . • µ ( x ) is the mutant birth rate of an parental individual with trait x . Mutation is indeedanother type of birth event, which results in a newborn with a different trait from itsparent. • p ( x, dh ) is the law of mutation variation h = y − x between a mutant y and its parentaltrait x . Since the mutant trait y = x + h is in X , the law has support in X − x := { y − x : y ∈ X } ⊂ R d .The natural death is due to “aging.” For example, it can be the natural death of a person, orthe exit of a firm from the market for reasons independent from the competition with otheralternatives. The competition captures the influence of the presence of an alternative on themarket on the survival rate of a given trait. For instance, if vinyl is directly threatened byCD, the massive presence of CDs in the market will increase the death rate of vinyl retailers.Competition is also present within an alternative: all other things being equal entry on themarket for vinyl increases competition in this market. The total death rate of an individual oftrait x is hence D ( x ) = d ( x ) + (cid:82) α ( x, y ) (cid:104) ν , { y } (cid:105) dy , where (cid:104) ν t , { y } (cid:105) dy is the size of individualswith trait y .Let us give an algorithmic, pathwise construction of the dynamics of process ( ν t ) t ≥ . Theconstruction can be summarized as follows. (1) At any given time, say initial time 0, choose an individual from the population ν atrandom. Assume that its trait is x . (2) Simulate an exponential random variable T of parameter b ( x ) + D ( x ) + µ ( x ) . (3) Simulate a uniform random variable θ on interval [0 , . According to its value, decidewhat happens to this individual of trait x at time T : its death, a clonal birth from itor a birth with mutation from it. • If θ ∈ [0 , D ( x ) b ( x )+ D ( x )+ µ ( x ) ) , the individual is killed and the sub-population size of trait x decreases by one. • If θ ∈ [ D ( x ) b ( x )+ D ( x )+ µ ( x ) , D ( x )+ b ( x ) b ( x )+ D ( x )+ µ ( x ) ) , a clone birth of trait x is reproduced and thesub-population size of trait x increases by one.8 If θ ∈ [ D ( x )+ b ( x ) b ( x )+ D ( x )+ µ ( x ) , , a birth with mutation occurs from parental trait x , andan innovative sub-population is created. The innovative trait x + h is determinedby the mutation transition law p ( x, dh ) . (4) Return to step (1) and continue the iteration.This construction serves as the definition of the individual-based model as well as a basicsimulation algorithm for this model. The following assumption guarantees that the process iswell defined.
Assumption 1.
There exist constants ¯ b , ¯ d , α , ¯ α , such that < b ( x ) ≤ ¯ b , < d ( x ) ≤ ¯ d , < α ≤ α ( x, y ) ≤ ¯ α , and b ( x ) − d ( x ) > for all x in X . Indeed, owing to the upper-bounds of the transition rates given in Assumption 1, the dynamicsare stochastically dominated by a Poisson process with birth rate ¯ b and death rate ¯ d + ¯ αν t .This fact justifies the existence of the process ( ν t ) t ≥ .For each K ∈ N ∗ , consider a population with initial size K . The proportion of individualswith each trait in the population is given by X Kt = ν t K . Our objective is to see whether thedynamics of X K can be approximated by deterministic equations when K is large enough.To give some flavor of our approach, let us cite two specific examples of process ( ν t ) t ≥ ,respectively with only one trait and two traits in the trait space, and without further mutations,i.e. mutation rate µ ( x ) ≡ for any x ∈ X . Hence the process depicts the short-term diffusionor competition behavior of the trait(s) on the individual level. It leads to an approximationof the usual logistic and Lotka-Volterra equations (cf. equations (1.1) and (1.2)) on thepopulation level. Proposition 3.1.
Assume X = { x } , µ ( x ) ≡ , and K × α ( x, x ) → α ( x, x ) as K → + ∞ . Sup-pose that the initial population size (cid:104) X K , { x } (cid:105) converges to n as K → + ∞ , then (cid:104) X Kt , { x } (cid:105) is converges in probability to the solution of the following (logistic) equation: (3.1) ˙ n t ( x ) = ( b ( x ) − d ( x ) − α ( x, x ) n t ( x )) n t ( x ) , which has a unique stable point ¯ n ( x ) = b ( x ) − d ( x ) α ( x,x ) . The result has been proved in (Fournier and Méléard, 2004, Theorem 5.3) and is its one-dimensional realization. Similarly, there is a two-dimensional limiting process.
Proposition 3.2.
Assume X = { x, y } , µ ( z ) ≡ for z ∈ X , and K × α ( w, z ) → α ( w, z ) as K → + ∞ for w, z ∈ X . Suppose that the initial sub-population sizes (cid:104) X K , { x } (cid:105) and (cid:104) X K , { y } (cid:105) converges to n ( x ) and n ( y ) as K → + ∞ , then ( (cid:104) X Kt , { x } (cid:105) , (cid:104) X Kt , { y } (cid:105) ) convergesin probability to the solution of the following (Lotka-Volterra) equations: (3.2) (cid:40) ˙ n t ( x ) = ( b ( x ) − d ( x ) − α ( x, x ) n t ( x ) − α ( x, y ) n t ( y )) n t ( x ) , ˙ n t ( y ) = ( b ( y ) − d ( y ) − α ( y, x ) n t ( x ) − α ( y, y ) n t ( y )) n t ( y ) . Proposition 3.3.
Consider the Lotka-Volterra system ( n ( x ) , n ( y )) satisfying equations (3.2) .Suppose that n ( x ) , n ( y ) > and ¯ f ( y, x ) := b ( y ) − d ( y ) − α ( y, x )¯ n ( x ) > and its symmetricform ¯ f ( x, y ) < . Then, we conclude that (0 , ¯ n ( y )) is the only stable point. The proof is given in Appendix. 9 .2 Development with sequential substitutions
This subsection takes mutation into consideration. This allows us to recover the well-knownprocess consisting in a sequence of innovations, each followed by a short-term competition/substitutionprocess introduced in the previous subsection (Champagnat, 2006). In this model, innovationsoccur only rarely and do not take place on the same timescale as competition and substitu-tion. In order to identify the proper range of mutation timescale in this sequential innovation-substitution model, let us replace the mutation rate µ ( x ) by σµ ( x ) . Besides, define the fitnessfunction of trait x with respect to y by ¯ f ( x, y ) = b ( x ) − d ( x ) − α ( x, y )¯ n ( y ) . The following assumption gives a non-coexistence condition for any pair of distinct traits.
Assumption 2.
For all x, y in X , ¯ f ( x, y ) · ¯ f ( y, x ) < . Champagnat (Champagnat, 2006, Theorem 1) proved the following result.
Proposition 3.4 (Champagnat, 2006) . Consider a sequence of processes ( X Kt ) t ≥ , K ∈ N ∗ ,with rescaled mutation law σµ ( x ) . Suppose that the initial points X K = N K K δ x satisfy that N K K law → n as K → + ∞ , where n is a strictly positive constant. Also suppose that, for all C > , (3.3) exp {− CK } (cid:28) Kσ (cid:28) K .
Then, the trajectory { X Kt/Kσ } t ≥ converges in probability to { Y t } t ≥ such that Y t = (cid:40) n δ x , t = 0 , ¯ n ( η t ) δ η t , t > , where { η t } t ≥ with η = x is a Markov jump process in trait space X with jump rate from any z ∈ X to y ∈ X ¯ n ( z ) [ ¯ f ( y, z )] + b ( y ) m ( z, dy ) . Here [ ¯ f ( y, z )] + denotes the positive part of ¯ f ( y, z ) ∈ R . His arguments of proof gave rigorous mathematical justification of the following ideas. Thetimescale for the occurrence of mutations is of order Kσ , whereas the fixation timescale (forthe competition between a pair of traits to attain an equilibrium) is of order ln K . Accordingto (3.3), the interval between two occurrences of mutation is much longer than the fixationtimescale. Proposition 3.4 implies that, in this case, the population trait appears monomor-phic (which means that all the individuals bear the same trait) at any moment on the mutationtimescale. In other words, when a new trait appears by mutation, its competition with theexisting one takes place very quickly and one of them is beaten and vanishes. Then all the However, e − CK (cid:28) Kσ guarantees that the mutation is not so rare that the fixed monomorphic populationdoes not drift out of its unique stable equilibrium (cf. Freidlin and Wentzell Freidlin and Wentzell (1984)). Simulations of the monomorphic substitution sequence model arising in Proposition 3.4. individuals bear the same trait, namely the one that survives from the competition. This lastsfor a long time until another mutation takes place.Figure 1 shows some simulations of a sequential substitution model, with time on the x-axis andpopulation on the y-axis . The trait space contains respectively three (resp. four) types for theleft (resp. right) panel. The population densities of trait x , x , x , x are marked respectivelyby red, blue, green and black colored curves. Take b ( x ) = 3 , b ( x ) = 6 , b ( x ) = 8 , b ( x ) = 10 and death rates d ( x i ) ≡ for i = 0 , , , . Other parameters are: competition kernel α ≡ ,migration kernel m ≡ . , and mutation scale σ = K − , with initial population size K = 100 .One can see that the competition process of two successive alternatives, which leads to thefixation of the fitter and the vanishing of the less fit, takes very little time compared withthe time interval between two innovations. Therefore, the typical double S -curves of theLotka-Volterra equation (3.2) are almost invisible. One has to “stretch” the time line at each“competition moment” to recover them. In the previous section, an individual always bears the same trait from its birth to its death.For example, when a firm enters the market and adopts a certain technology, it never changesthat choice until it quits the market. However, before a firm with poor performance quitsthe market because it is losing market share to its competitors, it may well try to improvethe performance by changing the technology. If it adopts its better performing competitor’stechnology, it is “imitating”. If it tries some other technology randomly, it is doing a “trial”.In the terminology of birth-death process models, this is called migration .In this section, by allowing migration of individuals, we extend the previous sequential sub-stitution model into a tree-form development model. On the one hand, migration from an oldtrait to a new trait accelerates the diffusion of the new one in the population; on the otherhand, migration from a new trait to an old trait prevents the old from completely vanishing.11t any period, somebody will trial to revive an alternative, and depending on how competitiveit is in the current market structure, it can make a comeback. Therefore, instead of having a“monomorphic” market all the time as in the sequential substitution model, several traits cancoexist in the population under the refined model. In the long term, we obtain a “developmenttree” of the alternatives, with a diversity of sustainable alternatives.Studying more than two alternatives at the same time brings another issue. When severalalternatives coexist, does each of them compete with all of the others? Is there migrationbetween each pair of alternatives? In this paper, we assume that the competition and themigration are rather local than global. For example, as means of transportation, bicyclescompete with cars, cars compete with trains, trains compete with airplanes, but bicycles donot compete with airplanes. It is precisely because of this assumption that we can obtaincomebacks in equilibrium: if the direct competitor of an alternative disappears, it is possiblefor this alternative to prosper again.For these reasons, this paper considers only the competition and migration between traitswith fitnesses that are “close” to each other. While in each specific context, an alternativemay compete with several alternatives that are more or less close to it, we consider a particularcase where each trait only compete with the two “nearest neighbors” in terms of their fitness.This makes the explicit form of the “development tree” tractable, and illustrates the main ideaof the model.In Section 4.2, we consider the case without mutation, in order to see how migration changesthe process of short-term competition and substitution. Section 4.3 extends the model toa long-term tree process by taking mutation, i.e. innovation, into consideration. Unlike inthe previous section, where we apply established mathematical results to our context, in thissection we develop new mathematical tools in probability theory in order to construct the treemodel.
Let us introduce migration into the basic mutation-free model specified in Section 3.1. Denoteby m ( x, dy ) the transition law for an individual migrating from trait x to trait y . In thissubsection, we assume that there is no mutation, µ ( x ) ≡ . Suppose that competition andmigration take place only between two alternatives that are neighbors on the fitness landscape .Explicitly, for a given trait space X = { x , x , x , · · · , x L } containing L + 1 distinct traits( L ∈ N ), admit the following assumption: Assumption 3.
The L + 1 traits are ordered on an increasing fitness landscape, i.e. x ≺ x ≺ . . . ≺ x L , where the order ≺ is defined as follows: (4.1) x ≺ y if ¯ f ( x, y ) < and ¯ f ( y, x ) > . In addition, m ( x i , x j ) = α ( x i , x j ) = 0 , ∀ | i − j | > . Note that the order is not a total order because it doesn’t have the transitivity and comparability property.For instance, ≺ doesn’t hold between x and x even if x ≺ x and x ≺ x . x competes with trait x , but not with trait x , and each trait is “fitter” thanthe previous one. We would like to study the birth-death process with such traits and localcompetition and migration behavior. Before stating Proposition 4.1, let us introduce somenotation and a technical assumption.Recall that ¯ n ( x ) is the rest point of the logistic equation (3.1) for trait x (cf. Proposition 3.1).Following Bovier and Wang (2013), define a specific configuration Γ ( L ) by(4.2) Γ ( L ) := l (cid:88) i =0 ¯ n ( x i ) δ x i , if L = 2 l. l +1 (cid:88) i =1 ¯ n ( x i − ) δ x i − , if L = 2 l + 1 . In other words, when the traits are ordered by their fitness, at Γ ( L ) , only every other trait ispresent, and its size is its equilibrium size with only intra-trait competition but not inter-traitcompetition. Proposition 4.1.
Consider the process ( X Kt ) t ≥ with rescaled migration law (cid:15)m ( x, dy ) on thetrait space X = { x , x , . . . , x L } . Suppose that for all K , X K = N K K δ x , and N K K → n in lawas K → + ∞ , where n > . If (4.3) (cid:28) K(cid:15) (cid:28) K, then, there exists a constant ¯ t L > such that, lim K →∞ X Kt ln (cid:15) (d) = Γ ( L ) , ∀ t > ¯ t L , under the total variation norm. The proof is in the Appendix. If, in a large population, the frequency of migration events ismuch higher than that of birth and death events, but not high enough for the whole populationto be migrating (cf. 4.3), then the outcome of the nearest-neighbor competition and migrationis the elimination of every other trait. The remaining traits do not compete with each otherand there is no migration between them. Therefore, the size of the subpopulation bearing acertain trait is just the equilibrium size of the trait if it is the only trait available, i.e. the onedetermined by (3.1).Two numerical simulations are illustrated in Figure 2. The parameters are the same as forFigure 1 except that the initial population size is now K = 1000 . In addition, (cid:15) = K − inthe three-type case (LHS) and (cid:15) = K − in the four-type case (RHS). One can verify thatcondition (4.3) holds in both cases. According to Proposition 4.1, the fixation timescale (i.e.the time after which the population can be approximated by Γ ( L ) ) is of order ln (cid:15) . The limit,stable configuration is Γ (2) = 3 δ x + 8 δ x for the three-type case and Γ (3) = 6 δ x + 10 δ x forthe four-type case. In this subsection, we incorporate mutation into the model and, in particular, focus on thejoint effect of migration and mutation on different timescales. It turns out that a trait that13 time popu l a t i on den s i t y n ( x i ) trait x trait x trait x time popu l a t i on den s i t y n ( x i ) trait x trait x trait x trait x Figure 2:
Simulations of a development tree model arising in Proposition 4.1 on a three- and four-type traitspace. almost dies out on the migration timescale has a chance to recover itself and further to bestabilized on the mutation timescale. This phenomenon results from a change in the fitnesslandscape, which is due to the arrival of a mutant.As in Section 3.2, mutation takes place according to the mutation kernel p ( x, dh ) . Due to thislaw, a previously non-presenting trait appears from time to time, and enlarges the space ofcurrently presenting traits. On the contrary, the migration kernel m ( x, dy ) only acts on thespace of currently presenting traits. The asymptotic behavior of the birth-death process withmigration and mutation depends on the timescales of both events. Let us rescale the mutationlaw by σ and the migration law by (cid:15) .First we need some assumptions and the definition of the trait substitution tree model, fol-lowing Bovier and Wang (2013). Assumption 4.
1. For any given distinct traits { x , x , · · · , x n } ⊂ X , n ∈ N , there exists aunique total order permutation x n ≺ x n ≺ · · · ≺ x n n − ≺ x n n . .2. When there are n +1 distinct traits currently presenting, they are labeled by x ( n )0 , x ( n )1 , . . . , x ( n ) n such that x ( n )0 ≺ x ( n )1 ≺ · · · ≺ x ( n ) n . When a new trait x appears, the n + 2 presenting traits arerelabeled by x ( n +1)0 , x ( n +1)1 , . . . , x ( n +1) n +1 such that x ( n +1)0 ≺ x ( n +1)1 ≺ · · · ≺ x ( n +1) n ≺ x ( n +1) n +1 .3. Competition and migration only occurs between nearest neighbor and we force the inter-acting kernels to vanish between non-nearest neighbor sites, i.e. for all n , m ( x ( n ) i , x ( n ) j ) = α ( x ( n ) i , x ( n ) j ) ≡ for | i − j | > . Hence, we retain the order ≺ as defined in Assumption 3. At the end of the previous subsection, we point out that, on the migration timescale, thereis a variety of paths to approach the equilibrium configuration. However, the equilibrium14onfiguration of any L -trait system without mutation is always the same, i.e. Γ ( N ) defined by(4.2), and the timescale for convergence is always of order O (ln (cid:15) ) as shown in Proposition4.1. Definition 4.2.
A Markov jump process { Γ t : t ≥ } is a trait substitution tree (or TST forshort) with ancestor Γ = ¯ n ( x ) δ x if the following hold: (i) For all l ∈ N , all k ∈ { , . . . , l } , and all h ∈ X − x (2 l )2 k , it jumps from configuration Γ (2 l ) = (cid:80) li =0 ¯ n ( x (2 l )2 i ) δ x (2 l )2 i , with transition rate ¯ n ( x (2 l )2 k ) µ ( x (2 l )2 k ) p ( x (2 l )2 k , dh ) , to a new configuration Γ (2 l +1) determined as follows: Γ (2 l +1) = j (cid:88) i =1 ¯ n ( x (2 l )2 i − ) δ x (2 l )2 i − + ¯ n ( x (2 l )2 k + h ) δ x (2 l )2 k + h + l (cid:88) i = j +1 ¯ n ( x (2 l )2 i ) δ x (2 l )2 i , if ∃ ≤ j ≤ l s.t. x (2 l )2 j ≺ x (2 l )2 k + h ≺ x (2 l )2 j +1 , j (cid:88) i =1 ¯ n ( x (2 l )2 i − ) δ x (2 l )2 i − + l (cid:88) i = j ¯ n ( x (2 l )2 i ) δ x (2 l )2 i , if ∃ ≤ j ≤ l s.t. x (2 l )2 j − ≺ x (2 l )2 k + h ≺ x (2 l )2 j . The traits presenting in Γ (2 l +1) are relabeled according to the total order relation: x (2 l +1)0 ≺ x (2 l +1)1 ≺ · · · ≺ x (2 l +1)2 l ≺ x (2 l +1)2 l +1 . (ii) For all l ∈ N , all k ∈ { , . . . , l + 1 } , and all h ∈ X − x (2 l +1)2 k − , it jumps from configuration Γ (2 l +1) = (cid:80) l +1 i =1 ¯ n ( x (2 l +1)2 i − ) δ x (2 l +1)2 i − , with transition rate ¯ n ( x (2 l +1)2 k − ) µ ( x (2 l +1)2 k − ) p ( x (2 l +1)2 k − , dh ) , toa new configuration Γ (2 l +2) determined as follows: Γ (2 l +2) = j (cid:88) i =1 ¯ n ( x (2 l +1)2 i − ) δ x (2 l +1)2 i − +¯ n ( x (2 l +1)2 k − + h ) δ x (2 l +1)2 k − + h + l +1 (cid:88) i = j +1 ¯ n ( x (2 l +1)2 i − ) δ x (2 l +1)2 i − , if ∃ ≤ j ≤ l +1 s.t. x (2 l +1)2 j − ≺ x (2 l +1)2 k − + h ≺ x (2 l +1)2 j , j (cid:88) i =1 ¯ n ( x (2 l +1)2 i − ) δ x (2 l +1)2 i − + l +1 (cid:88) i = j ¯ n ( x (2 l +1)2 i − ) δ x (2 l +1)2 i − , if ∃ ≤ j ≤ l +1 s.t. x (2 l +1)2 j − ≺ x (2 l +1)2 k − + h ≺ x (2 l +1)2 j − . The traits presenting in Γ (2 l +2) are relabeled according to the total order relation: x (2 l +2)0 ≺ x (2 l +2)1 ≺ · · · ≺ x (2 l +2)2 l +1 ≺ x (2 l +2)2 l +2 . Roughly speaking, when there are n + 1 traits in the current trait space and ordered on anincreasing fitness landscape, with the nearest-neighbor competition and migration rule, everyother trait is absent from the current configuration. When a new trait x appears by mutation,put it into the queue of the n + 1 old traits according to its fitness. If the nearest trait withhigher fitness than x is absent (resp. present) in the current configuration, then trait x survives(resp. vanishes) in the new configuration. In both cases, the old traits with higher fitness than x remain present or absent in the new configuration as before, while the old traits with lowerfitness than x switch their presence to absence or vice verse in the new configuration. Proposition 4.3.
Consider the process { X K,(cid:15),σt } t ≥ with rescaled migration law (cid:15)m ( x, y ) andrescaled mutation law σµ ( x ) . Suppose that for all K , X K,(cid:15),σ = N K K δ x , and N K K → ¯ n ( x ) inlaw as K → ∞ . If, in addition to condition (4.3) that (cid:28) K(cid:15) (cid:28) K , one has (4.4) ln 1 (cid:15) (cid:28) Kσ (cid:28) e KC , ∀ C > , hen ( X K,(cid:15),σt/Kσ ) t ≥ converges, as K → ∞ , to the trait substitution tree (Γ t ) t ≥ defined in Defi-nition 4.2 in the sense of f.d.d. on M F ( X ) equipped with the topology induced by mappings ν (cid:55)→ (cid:104) ν, f (cid:105) with f a bounded measurable function on X .. The proof is in the Appendix. According to (4.4), the fixation timescale of order ln (cid:15) is muchshorter than the mutation timescale of order Kσ . At the same time, Kσ (cid:28) e KC prevents thesystem from drifting away from the TST equilibrium configuration on the mutation timescale(Freidlin and Wentzell, 1984). time popu l a t i on den s i t y n ( x i ) trait x trait x trait x trait x output.T popu l a t i on den s i t y n ( x i ) trait x trait x trait x trait x trait x Figure 3:
Simulations of a trait substitution tree on the mutation timescale arising in Proposition 4.3 onfour- and five-type trait space.
Return to the examples shown in Figure 2 and add mutations to them. Let us consider theparticular case where a mutant is always fitter than all the existing traits. In both examples,the birth rates of the red, the blue, the green, the black and the purple-colored populationsare respectively 3, 6, 8, 10, and 12, while their death rates are all constantly . Take (cid:15) = K − . and σ = K − . , and the initial scaling parameter K = 400 . The simulation results are shownin Figure 3. Remark how the arrival of each mutant “reshuffles the cards” by initiating a newround of competition which leads to a new equilibrium. Let us compare these examples withthe ones shown in Figure 1, where no migration is concerned. One sees that, with migration,a once-eliminated trait x now has a chance to be revived, because its competitor y whicheliminated x in the previous round may itself now be eliminated by some other trait. Observethat the TST process jumps from Γ ( n ) to Γ ( n +1) on a mutation timescale of order Kσ . Besides,on this timescale, the fixation process (describing short-term pair-wise competition) is notvisible anymore. However, by zooming into the infinitesimal fixation period, S -type curves asin Figure 2 will still emerge.The specific rule of nearest-neighbor migration and competition is adopted in this section as anexample to illustrate the development tree model. When the model is applied to a particularproblem, the local migration and competition rules should be determined according to thecontext. Although the specific form of the process limits in Propositions 4.1 and 4.3 (everyother trait survives on the increasing fitness landscape) may no longer hold, it gives us some16avor of the tree form depicting a long-term dynamic, where short-term competitions and raremutations lead to diversification and development. A notable feature of our tree model is the three distinct timescales related with differentevents. The shortest one is the lifecycle time of a single generation, which is of order .The intermediate timescale is identified as the migration time of order ln (cid:15) . The relationbetween these two timescale (cf. equation (4.3)) justifies the macroscopic approximation ofa population. Models limited to these two timescales have been applied in the framework ofshort-term competition/substitution, which leads to a temporary equilibrium. The noveltyof our model is to add a third, the longest timescale of order Kσ , which corresponds to theinterval between two mutations. Observing from this timescale, the process is composed of aseries of rapid transitions from one temporary equilibrium to another. Indeed, every temporaryequilibrium, though lasting a long time, is eventually broken by the arrival of a mutant oran innovation. These three timescales of distinct orders allow us to draw a tree form graphdepicting the long-term development of the alternatives in question.Besides, though our model aims to explain the long-term development and diversification oftechnologies, ideas, custom etc. in various contexts, it remains an illustrative tool on thetheoretical level. For example, the events such as birth, death and migration have differentsignification in different context. We have adopted an evolutionary setting in this paper sothat the parameters such as birth-death rates or migration laws are exogenously given. Thereis however a large body of literature that focuses on the strategic side of individual decisionmaking, and builds models with different mechanisms generating rational choices or imitativebehavior. The usual approaches consist of non cooperative games (with decision-making onan individual level) and social learning (at the population level). The former often takesthe uncertainty and heterogeneity on the individual level into account. Integrating theseapproaches in our specification constitutes an obvious avenue for future research.
Appendix
Proof of Proposition 3.3.
In fact, there are four fixed points of the two-dimensional L-V system(3.2), namely, (0 , , (¯ n ( x ) , , (0 , ¯ n ( y )) , and ( n ∗ ( x ) , n ∗ ( y )) , where ( n ∗ ( x ) , n ∗ ( y )) is such that (cid:26) b ( x ) − d ( x ) − α ( x, x ) n t ( x ) − α ( x, y ) n t ( y ) = 0 b ( y ) − d ( y ) − α ( y, x ) n t ( x ) − α ( y, y ) n t ( y ) = 0 . By simple calculation, we obtain that (cid:40) n ∗ ( x ) = α ( y,y ) f ( x,y ) α ( x,x ) α ( y,y ) − α ( x,y ) α ( y,x ) n ∗ ( x ) = α ( x,x ) f ( y,x ) α ( x,x ) α ( y,y ) − α ( x,y ) α ( y,x ) . An incomplete list of references in economics includes Hiebert (1974), Stoneman (1981), Jensen (1982),Chatterjee and Eliashberg (1990), Sinha and Chandrashekaran (1992), Meade (1989), Bruckner et al. (1996),Chatterjee et al. (2000), Young (2009), Young (2011) and Kreindler and Young (2014).
17o make sense of the solution as a population density (which must be non-negative), one needs f ( x, y ) · f ( y, x ) > . It contradicts the assumption f ( x, y ) < , f ( y, x ) > . We thus excludethe solution ( n ∗ ( x ) , n ∗ ( y )) .The Jacobian matrix for the system (3.2) at point (0 , is (cid:18) b ( x ) − d ( x ) 00 b ( y ) − d ( y ) (cid:19) . Obviously its eigenvalues are both positive. Thus (0 , is unstable.The Jacobian matrix at point (¯ n ( x ) , is (cid:18) − ( b ( x ) − d ( x )) − α ( x, y )¯ n ( x )0 b ( y ) − d ( y ) − α ( y, x )¯ n ( x ) (cid:19) = (cid:18) − ( b ( x ) − d ( x )) − α ( x, y )¯ n ( x )0 f ( y, x ) (cid:19) . Since one of its eigenvalue − ( b ( x ) − d ( x )) is negative whereas the other one is f ( y, x ) > ,the equilibrium (¯ n ( x ) , is unstable.The Jacobian matrix of system (3.2) at point (0 , ¯ n ( y )) is (cid:18) b ( x ) − d ( x ) − α ( x, y )¯ n ( y ) 0 − α ( y, x )¯ n ( y ) − ( b ( y ) − d ( y )) (cid:19) = (cid:18) f ( x, y ) 0 − α ( y, x )¯ n ( y ) − ( b ( y ) − d ( y )) (cid:19) , whose eigenvalues are both negative because of the condition f ( x, y ) < . Thus (0 , ¯ n ( y )) isthe only stable equilibrium of the system (3.2). Proof of Proposition 4.1.
Let us prove the result for the three-trait toy model (see Figure 4)to illustrate the basic idea of proof, though our analysis is not limited to the three-trait caseonly. Indeed, the whole machinery is available for any finite-trait space.Assume X = { x , x , x } . Let ξ Kt ( x ) := (cid:104) X Kt , { x } (cid:105) and ξ Kt ( x i ) := (cid:104) X Kt , { x i } (cid:105) for i = 1 , . Step 1 . Firstly, consider the emergence and growth of population at trait site x . Set S (cid:15) = inf { t > ξ Kt ( x ) ≥ (cid:15) } . Thanks to N K K → n > in law as K → ∞ and by applyingthe Law of Large Numbers of random processes (see Chap.11, Ethier and Kurtz 1986), oneobtains that, for any δ > , T > , lim K →∞ P (cid:18) sup ≤ t ≤ T (cid:12)(cid:12)(cid:12)(cid:12) ξ Kt ( x ) (cid:15) − n t ( x ) (cid:12)(cid:12)(cid:12)(cid:12) < δ (cid:19) = 1 where n t ( x ) is governed by equation ˙ n ( x ) = m ( x , x ) n with initial n ( x ) = 0 . Therefore,(5.1) lim K →∞ P (cid:18) m ( x , x ) n − δ < S (cid:15) < m ( x , x ) n + δ (cid:19) = 1 , that is, S (cid:15) is of order 1. 18 (cid:15) (cid:101) S η S η density time η(cid:15) ¯ n ( x ) S η (cid:101) S η S η recovery of x ¯ n ( x ) ¯ n ( x ) ¯ n ( x ) ¯ n ( x ) growth of x growth of x Figure 4:
Phase evolution of mass bars in early time window on the three-trait site space .For any η > , set S η = inf { t : t > S (cid:15) , ξ Kt ( x ) ≥ η } . Consider a sequence of rescaled processes (cid:16) N K, t K(cid:15) (cid:17) t ≥ S (cid:15) with N K, S(cid:15) K(cid:15) = ξ KS(cid:15) ( x ) (cid:15) → as K → ∞ . As before, by law of large numbers of randomprocesses (see Chap.11 Ethier and Kurtz 1986), one obtains, for any δ > , T > ,(5.2) lim K →∞ P (cid:32) sup ≤ t ≤ T (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) N K, t K(cid:15) − m t (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) < δ (cid:33) = 1 , where m t is governed by equation ˙ m = ¯ f ( x , x ) m = ( b ( x ) − d ( x ) − α ( x , x )¯ n ( x )) m with m = 1 .Set T η/(cid:15) = inf { t − S (cid:15) : t > S (cid:15) , N K, t K(cid:15) ≥ η/(cid:15) } , and t η/(cid:15) = inf { t > m t ≥ η/(cid:15) } . Then, for any δ > , there exists δ (cid:48) > such that(5.3) lim K →∞ P (cid:18)(cid:18) f ( x , x ) − δ (cid:19) ln 1 (cid:15) < S η − S (cid:15) < (cid:18) f ( x , x ) + δ (cid:19) ln 1 (cid:15) (cid:19) = lim K →∞ P (cid:18)(cid:18) f ( x , x ) − δ (cid:19) ln 1 (cid:15) < T η/(cid:15) < (cid:18) f ( x , x ) + δ (cid:19) ln 1 (cid:15) (cid:19) = lim K →∞ P (cid:32) (cid:18) f ( x , x ) − δ (cid:19) ln 1 (cid:15) < t η/(cid:15) < (cid:18) f ( x , x ) + δ (cid:19) ln 1 (cid:15) , sup ≤ t ≤ t η/(cid:15) | N K, t K(cid:15) − m t | < δ (cid:48) (cid:33) =1 where the last equal sign is due to (5.2).After population of trait x reaches some η threshold, the dynamics (cid:0) ξ Kt ( x ) , ξ Kt ( x ) (cid:1) can beapproximated by the solution of a two-dimensional Lotka-Volterra equations. By Proposition3.3, it takes time of order 1 (mark this time coordinator by (cid:101) S η ) for the two subpopulations19witching their mass distribution and gets attracted into η − neighborhood of the stable equi-librium (0 , ¯ n ( x )) . Step 2.
Now consider the emerging and growth of population ξ Kt ( x ) := (cid:104) X Kt , { x } (cid:105) at traitsite x . Set S (cid:15) = inf { t : t > (cid:101) S η , ξ Kt ( x ) ≥ (cid:15) } . Similarly as is done for S (cid:15) in (5.1), one can getthat lim K →∞ P ( S (cid:15) − (cid:101) S η = O (1)) = 1 . On a longer time scale, we will not distinguish S (cid:15) from (cid:101) S η .Set S η = inf { t : t > S (cid:15) , ξ Kt ( x ) ≥ η } . One follows the same procedure to derive (5.3) andasserts that for any δ > ,(5.4) lim K →∞ P (cid:18)(cid:18) f ( x , x ) − δ (cid:19) ln 1 (cid:15) < S η − (cid:101) S η < ( 1¯ f ( x , x ) + δ ) ln 1 (cid:15) (cid:19) = 1 . Note that assumption ( B b ( x ) − d ( x ) ≥ f ( x ,x ) + f ( x ,x ) guarantees that ξ Kt ( x ) can not growso fast in exponential rate b ( x ) − d ( x ) such that it reaches some η -level before S η .During time period ( (cid:101) S η , S η ) , population at site x , on one hand, decreases due to the com-petition from more fitter trait x . On the other hand, it can not go below (cid:15) level due tothe successive migration in a portion of (cid:15) from site x . More precisely, by neglecting migrantcontribution, ξ Kt ( x ) converges n t ( x ) in probability as K tends to ∞ , where(5.5) ˙ n t ( x ) = ( b ( x ) − d ( x ) − α ( x , x )¯ n ( x )) n t ( x ) = ¯ f ( x , x ) n t ( x ) with n ( x ) = η . Let ∆ S η = S η − (cid:101) S η . Then, for any δ > ,(5.6) lim K →∞ P (cid:16) ξ KS η ( x ) ∈ ( n ∆ S η ( x ) − δ, n ∆ S η ( x ) + δ ) (cid:17) = lim K →∞ P (cid:16) ηe ¯ f ( x ,x )∆ S η − δ < ξ KS η ( x ) < ηe ¯ f ( x ,x )∆ S η + δ (cid:17) = lim K →∞ P (cid:16) η(cid:15) | ¯ f ( x ,x ) | / ¯ f ( x ,x ) − δ < ξ KS η ( x ) < η(cid:15) | ¯ f ( x ,x ) | / ¯ f ( x ,x ) + δ (cid:17) = 1 where the second equality is due to (5.4). Taking the migration from site x into account, wethus have(5.7) lim K →∞ P (cid:16) ξ KS η ( x ) = O ( (cid:15) | ¯ f ( x ,x ) | / ¯ f ( x ,x ) ∨ (cid:15) ) (cid:17) = 1 . We proceed as before for (cid:101) S η in step 1. After time S η , the mass bars on dimorphic system ( ξ Kt ( x ) , ξ Kt ( x )) can be approximated by ODEs and will be switched again in time of order 1(marked by (cid:101) S η as in Figure 4), and they are attracted into η − neighborhood of (0 , ¯ n ( x )) . Asfor the population density on site x , one obtains from (5.7)(5.8) lim K →∞ P (cid:16) ξ K (cid:101) S η ( x ) = O ( (cid:15) c ) (cid:17) = 1 where c = | ¯ f ( x ,x ) | ¯ f ( x ,x ) ∧ ≤ . Step 3.
We now consider the recovery of subpopulation at trait site x . Recovery arisesbecause of the lack of effective competitions from its neighbor site x , or under negligible20ompetitions since the local population density on x is very low under the control of its fitterneighbor x . Without loss of generality, we suppose c := | ¯ f ( x ,x ) | ¯ f ( x ,x ) < in (5.7).Set S η = inf { t : t > (cid:101) S η , ξ Kt ( x ) ≥ η } . We proceed as before in step 1. From (5.7), ξ K (cid:101) Sη ( x ) (cid:15) c converges to some positive constant (say m ) in probability as K → ∞ . Thus, by applyinglaw of large numbers to the sequence of processes N Kt K(cid:15) c , for any δ > , T > , (5.9) lim K →∞ P (cid:18) sup ≤ t ≤ T (cid:12)(cid:12)(cid:12)(cid:12) ξ Kt ( x ) (cid:15) c − m t (cid:12)(cid:12)(cid:12)(cid:12) < δ (cid:19) = 1 where m t is governed by logistic equation ˙ m = ( b ( x ) − d ( x )) m starting with a positive initial m .Following the same way to obtain (5.3), time length S η − (cid:101) S η can be approximated by timeneeded for dynamics m to approach η/(cid:15) c level, which is of order c ( b ( x ) − d ( x )) ln (cid:15) , i.e. for any δ > ,(5.10) lim K →∞ P (cid:18)(cid:18) c b ( x ) − d ( x ) − δ (cid:19) ln 1 (cid:15) < S η − (cid:101) S η < (cid:18) c b ( x ) − d ( x ) + δ (cid:19) ln 1 (cid:15) (cid:19) = 1 . At the same time, ξ Kt ( x ) converges in probability to ψ t which satisfies equation ˙ ψ = ¯ f ( x , x ) ψ with ψ (cid:101) S η = η . Then, we can justify the following estimate for population density at site x ,(5.11) lim K →∞ P (cid:16) ξ KS η ( x ) = O ( (cid:15) c ∨ (cid:15) ) (cid:17) = 1 where c = c | ¯ f ( x ,x ) | b ( x ) − d ( x ) .We now combine all these estimates (5.3), (5.4), (5.10) together, and conclude that(5.12) lim K →∞ P (cid:16) (cid:107) X Kt ln (cid:15) − Γ (2) (cid:107) < δ (cid:17) = 1 for t > ¯ t := f ( x ,x ) + f ( x ,x ) + c b ( x ) − d ( x ) under the total variation norm (cid:107) · (cid:107) on M F ( X ) . Proof of Proposition 4.3.
First recall that all Markov jump processes can be specified by twofeatures: the exponentially distributed waiting time, and the one-step transition rule (Ethierand Kurtz (1986)). Our proof, which shows that the limiting process is a Markov jump process Γ t , is hence made up of two parts.The first part of the proof consists in the characterization of exponential waiting time of eachmutation arrival, as can be seen from the construction of the process in Section 3.1. Let L = 2 l , X K,(cid:15),σ = Γ ( L ) , and τ be the first mutation arrival time after time 0. Then, similararguments used in (Champagnat, 2006, Lemma 2 (c)) shows us that Lemma 5.1. (5.13) lim K →∞ P (cid:18) τ > tKσ (cid:19) = exp (cid:32) − t l (cid:88) i =0 ¯ n ( x (2 l )2 i ) µ ( x (2 l )2 i ) (cid:33) , and (5.14) lim K →∞ P (cid:16) at time τ, mutant comes from trait x (2 l )2 k (cid:17) = ¯ n ( x (2 l )2 k ) µ ( x (2 l )2 k ) (cid:80) li =0 ¯ n ( x (2 l )2 i ) µ ( x (2 l )2 i ) .
21e will not repeat the details of the proof.The second part of the proof can been seen as a corollary of Proposition 4.1. It specifies thenew equilibrium configuration and shows that fixation time of the new configuration is of order ln (cid:15) , which is invisible on the mutation timescale. Lemma 5.2.
Assume that X K,(cid:15),σ = Γ (2 l ) + K δ x (2 l )2 k + h for some ≤ k ≤ l . Then there exists aconstant C > , for any δ > , such that (5.15) lim K →∞ P (cid:16) τ > C ln 1 (cid:15) , sup t ∈ ( C ln (cid:15) ,τ ) (cid:107) X K,(cid:15),σt − Γ (2 l +1) (cid:107) < δ (cid:17) = 1 where Γ (2 l +1) associated to Definition 4.2 (i) is defined as the following, in the first case x (2 l +1) i = x (2 l ) i ∀ ≤ i ≤ j, x (2 l +1)2 j +1 = x (2 l )2 k + h, x (2 l +1) i = x (2 l ) i − ∀ j + 2 ≤ i ≤ l + 1; in the second case x (2 l +1) i = x (2 l ) i ∀ ≤ i ≤ j − , x (2 l +1)2 j = x (2 l )2 k + h, x (2 l +1) i = x (2 l ) i − ∀ j + 1 ≤ i ≤ l + 1 . Proof of the lemma.
Indeed, equation (5.13) implies that for all
C > ,(5.16) lim (cid:15) → P ( τ (cid:15) > C ln 1 (cid:15) ) = 1 . According to the fitness landscape, there is one and only one ordered position for the newarising trait x (2 l )2 k + h in Γ (2 l ) . Suppose there exists j ∈ { , , . . . , l } such that x (2 l )2 k + h fitsbetween x (2 l )2 j and x (2 l )2 j +1 , i.e.(5.17) x (2 l )2 j − ≺ x (2 l )2 j ≺ x (2 l )2 k + h ≺ x (2 l )2 j +1 . Since both traits x (2 l )2 j − and x (2 l )2 j +1 are absent in Γ (2 l ) , the pair ( x (2 l )2 j , x (2 l )2 k + h ) is isolated andhence without competition from others. By the same argument as in the proof of Proposition4.1, the two-type system converges to (0 , ¯ n ( x (2 l )2 k + h ) δ x (2 l )2 k + h ) in time of order O (ln (cid:15) ) . Forthose traits of higher fitness than the isolated pair, nothing changes due to their isolation.Whereas for the traits of lower fitness than the pair, trait x (2 l )2 j − increases exponentially due tothe decay of its fitter neighbor x (2 l )2 j . So on and so forth, the mass occupation flips on the lefthand side of x (2 l )2 j . The entire rearrangement process takes time of order O (ln (cid:15) ) .In a similar way, we can prove the case where there is j ∈ { , , . . . , l } such that x (2 l )2 j − ≺ x (2 l )2 k + h ≺ x (2 l )2 j ≺ x (2 l )2 j +1 .The case where L = 2 l + 1 can be proved similarly.22 eferences Bartmanski, D. and
Woodward, I. (2013). The vinyl: The analogue medium in the ageof digital reproduction.
Journal of Consumer Culture , p. 1469540513488403.
Batten, D. (1982). On the dynamics of industrial evolution.
Regional Science and UrbanEconomics , (3), 449–462. Bhargava, S. (1989). Generalized lotka-volterra equations and the mechanism of technolog-ical substitution.
Technological Forecasting and Social Change , (4), 319–326. Blackman, A. W. (1972). A mathematical model for trend forecasts.
Technological Fore-casting and Social Change , , 441–452. — (1974). The market dynamics of technological substitutions. Technological Forecasting andSocial Change , , 41–63. Bovier, A. and
Wang, S.-D. (2013). Trait substitution trees on two time scales analysis.
Markov Processes and Related Fields , , 607–642. Bretschneider, S. I. and
Mahajan, V. (1980). Adaptive technological substitution mod-els.
Technological Forecasting and Social Change , (2), 129–139. Brown, L. (1981).
Innovation Diffusion: A New Perspective . Methuen.
Bruckner, E. , Ebeling, W. , Montaño, M. J. and
Scharnhorst, A. (1996). Nonlin-ear stochastic effects of substitution – an evolutionary approach.
Journal of EvolutionaryEconomics , (1), 1–30. Champagnat, N. (2006). A microscopic interpretation for adaptive dynamics trait substitu-tion sequence models.
Stochastic Processes and their Applications , (8), 1127–1160. Chandrasekaran, D. and
Tellis, G. (2007).
A critical review of marketing research ondiffusion of new products , M. E. Sharpe, pp. 39–80.
Chatterjee, R. and
Eliashberg, J. (1990). The innovation diffusion process in a hetero-geneous population: A micromodeling approach.
Management Science , (9), 1057–1079. — , — and Rao, V. (2000).
Dynamic models incorporating competition , Kluwer AcademicPublisher.
Clark, G. (1984).
Innovation Diffusion: Contemporary Geographical Approaches . Geo Books.
Easingwood, C. , Mahajan, V. and
Muller, E. (1981). A nonsymmetric respondinglogistic model for forecasting technological substitution.
Technological Forecasting and SocialChange , (3), 199–213. Ethier, S. and
Kurtz, T. (1986).
Markov Processes: Characterization and Convergence .Wiley.
Fisher, J. and
Pry, R. (1971). A simple substitution model of technological change.
Tech-nological Forecasting and Social Change , , 75–88.23 ournier, N. and Méléard, S. (2004). A microscopic probabilistic description of a locallyregulated population and macroscopic approximation.
Ann. Appl. Probab. , , 1880–1919. Freidlin, M. I. and
Wentzell, A. D. (1984).
Random perturbations of dynamical systems .New York: Springer.
Geroski, P. (2000). Models of technology diffusion.
Research Policy , (4-5), 603–625. Griliches, Z. (1957). Hybrid corn: an exploration in the economics of technological change.
Econometrica , (4), 501–522. — (1960). Hybrid corn and the economics of innovation. Science , (3422), 275–280. Guidolin, M. and
Guseo, R. (2015). Technological change in the us music industry: Within-product, cross-product and churn effects between competing blockbusters.
TechnologicalForecasting and Social Change , , 35–46. Guseo, R. and
Mortarino, C. (2014). Within-brand and cross-brand word-of-mouth forsequential multi-innovation diffusions.
IMA Journal of Management Mathematics , (3),287–311. Hiebert, L. D. (1974). Risk, learning, and the adoption of fertilizer responsive seed varieties.
American Journal of Agricultural Economics , (4), 764–768. Jensen, R. (1982). Adoption and diffusion of an innovation of uncertain profitability.
Journalof Economic Theory , (1), 182–193. Karmeshu , Bhargava, S. and
Jain, V. (1985). A rationale for law of technological substi-tution.
Regional Science and Urban Economics , (1), 137–141. Katz, E. , Levin, M. L. and
Hamilton, H. (1963). Traditions of research on the diffusionof innovation.
American Sociological Review , (2), 237–252. Kreindler, G. E. and
Young, H. P. (2014). Rapid innovation diffusion in social networks.
Proceedings of the National Academy of Sciences , (Supplement 3), 10881–10888. Lafferty, K. D. , DeLeo, G. , Briggs, C. J. , Dobson, A. P. , Gross, T. and
Kuris,A. M. (2015). A general consumer-resource population model.
Science , (6250), 854–857. Libai, B. , Muller, E. and
Peres, R. (2009). The role of within-brand and cross-brandcommunications in competitive growth.
Journal of Marketing , (3), 19–34. Magaudda, P. (2011). When materiality ‘bites back’: Digital music consumption practicesin the age of dematerialization.
Journal of Consumer Culture , (1), 15–36. Mahajan, V. and
Muller, E. (1979). Innovation diffusion and new product growth modelsin marketing.
Journal of Marketing , (4), 55–68. — , — and Bass, F. M. (1990). New product diffusion models in marketing: A review anddirections for research.
Journal of Marketing , (1), 1–26.24 ansfield, E. (1961). Technical change and the rate of imitation. Econometrica , (4),741–766. Meade, N. (1989). Technological substitution: A framework of stochastic models.
Techno-logical Forecasting and Social Change , (4), 389–400. Negus, K. (2015). Digital divisions and the changing cultures of the music industries (or, theironies of the artefact and invisibility).
Journal of Business Anthropology , (1), 151–157. Neuhoff, K. (2005). Large-scale deployment of renewables for electricity generation.
OxfordReview of Economic Policy , (1), 88–110. Parker, P. M. (1994). Aggregate diffusion forecasting models in marketing: A critical review.
International Journal of Forecasting , (2), 353–380. Peres, R. , Muller, E. and
Mahajan, V. (2010). Innovation diffusion and new productgrowth models: A critical review and research directions.
International Journal of Researchin Marketing , (2), 91–106. Pistorius, C. W. and
Utterback, J. M. (1997). Multi-mode interaction among technolo-gies.
Research Policy , (1), 67–84. Rogers, E. M. (2003).
Diffusion of Innovations . Free Press, 5th edn.
Sharif, M. and
Ramanathan, K. (1981). Binomial innovation diffusion models with dy-namic potential adopter population.
Technological Forecasting and Social Change , (1),63–87. Sharif, M. N. and
Kabir, C. (1976a). A generalized model for forecasting technologicalsubstitution.
Technological Forecasting and Social Change , (4), 353–364. — and — (1976b). System dynamics modeling for forecasting multilevel technological substi-tution. Technological Forecasting and Social Change , (1-2), 89–112. Sinha, R. K. and
Chandrashekaran, M. (1992). A split hazard model for analyzing thediffusion of innovations.
Journal of Marketing Research , (1), 116–127. Skiadas, C. (1985). Two generalized rational models for forecasting innovation diffusion.
Technological Forecasting and Social Change , (1), 39–61. Sovacool, B. K. (2009). The intermittency of wind, solar, and renewable electricity gener-ators: Technical barrier or rhetorical excuse?
Utilities Policy , (3), 288–296. Stoneman, P. (1981). Intra-firm diffusion, Bayesian learning and profitability.
The EconomicJournal , (362), 375–388. Styvén, M. E. (2010). The need to touch: Exploring the link between music involvementand tangibility preference.
Journal of Business Research , (9), 1088–1094. Verbruggen, A. (2008). Renewable and nuclear power: A common future?
Energy Policy , (11), 4036–4047. 25 an, H.-S. and Ma, K.-P. (2011). Competitive diffusion process of repurchased products inknowledgeable manufacturing.
European Journal of Operational Research , (3), 243–252. Young, H. P. (2009). Innovation diffusion in heterogeneous populations: Contagion, socialinfluence, and social learning.
American Economic Review , (5), 1899–1924. — (2011). The dynamics of social innovation. Proceedings of the National Academy of Sciencesof the United States of America ,108