Structural Drift: The Population Dynamics of Sequential Learning

We introduce a theory of sequential causal inference in which learners in a chain estimate a structural model from their upstream “teacher” and then pass samples from the model to their downstream “student”. It extends the population dynamics of genetic drift, recasting Kimura's selectively neutral theory as a special case of a generalized drift process using structured populations with memory. We examine the diffusion and fixation properties of several drift processes and propose applications to learning, inference, and evolution. We also demonstrate how the organization of drift process space controls fidelity, facilitates innovations, and leads to information loss in sequential learning with and without memory.


I. "SEND THREE-AND FOUR-PENCE, WE'RE GOING TO A DANCE"
This phrase was heard, it is claimed, over the radio during WWI instead of the transmitted tactical phrase "Send reinforcements we're going to advance" [1].As illustrative as it is apocryphal, this garbled yet comprehensible transmission sets the tone for our investigations here.Namely, what happens to knowledge when it is communicated sequentially along a chain, from one individual to the next?What fidelity can one expect?How is information lost?How do innovations occur?
To answer these questions we introduce a theory of sequential causal inference in which learners in a communication chain estimate a structural model from their upstream "teacher" and then, using that model, pass along samples to their downstream "student".This reminds one of the familiar children's game Telephone.By way of quickly motivating our sequential learning problem, let's briefly recall how the game works.
To begin, one player invents a phrase and whispers it to another player.This player, believing they have understood the phrase, then repeats it to a third and so on until the last player is reached.The last player announces the phrase, winning the game if it matches the original.Typically it does not, and that's the fun.Amusement and interest in the game derive directly from how the initial phrase evolves in odd and surprising ways.The further down the chain, the higher the chance that errors will make recovery impossible and the less likely the original phrase will survive.
The game is often used in education to teach the lesson that human communication is fraught with error.The final phrase, though, is not merely accreted error but the product of a series of attempts to parse, make sense, and intelligibly communicate the phrase.The phrase's evolution is a trade off between comprehensibility and accumulated distortion, as well as the source of the game's entertainment.We employ a much more tractable setting to make analytical progress on sequential learning, based on computational mechanics [2][3][4], intentionally selecting a simpler language system and learning paradigm than likely operates with children.
Specifically, we develop our theory of sequential learning as an extension of the evolutionary population dynamics of genetic drift, recasting Kimura's selectively neutral theory [5] as a special case of a generalized drift process of structured populations with memory.This is a substantial departure from the unordered populations used in evolutionary biology.Notably, this requires a new and more general information-theoretic notion of fixation.We examine the diffusion and fixation properties of several drift processes, demonstrating that the space of drift processes is highly organized.This organization controls fidelity, facilitates innovations, and leads to information loss in sequential learning and evolutionary processes with and without memory.We close by describing applications to learning, inference, and evolution, commenting on related efforts.
To get started, we briefly review genetic drift and fixation.This will seem like a distraction, but it is a necessary one since available mathematical results are key.
Then we introduce in detail our structured variants of these concepts-defining the generalized drift process and formulating a generalized definition of fixation appropriate to it.With the background laid out, we begin to examine the complexity of structural drift behavior.We demonstrate that it is a diffusion process within a space that decomposes into a connected network of structured subspaces.Building on this decomposition, we explain how and when processes jump between these subspacesinnovating new structural information or forgetting itthereby controlling the long-time fidelity of the communication chain.We then close by outlining future research and listing several potential applications for structural drift, drawing out consequences for evolutionary processes that learn.
Those familiar with neutral evolution theory are urged to skip to Sec.V, after skimming the next sections to pick up our notation and extensions.

II. FROM GENETIC TO STRUCTURAL DRIFT
Genetic drift refers to the change over time in genotype frequencies in a population due to random sampling.It is a central and well studied phenomenon in population dynamics, genetics, and evolution.A population of genotypes evolves randomly due to drift, but typically changes are neither manifested as new phenotypes nor detected by selection-they are selectively neutral.Drift plays an important role in the spontaneous emergence of mutational robustness [6,7], modern techniques for calibrating molecular evolutionary clocks [8], and nonadaptive (neutral) evolution [9,10], to mention only a few examples.
Selectively neutral drift is typically modeled as a stochastic process: A random walk that tracks finite populations of individuals in terms of their possessing (or not) a variant of a gene.In the simplest models, the random walk occurs in a space that is a function of genotypes in the population.For example, a drift process can be considered to be a random walk of the fraction of individuals with a given variant.In the simplest cases there, the model reduces to the dynamics of repeated binomial sampling of a biased coin, in which the empirical estimate of bias becomes the bias in the next round of sampling.In the sense we will use the term, the sampling process is memoryless.The biased coin, as the population being sampled, has no memory: The past is independent of the future.The current state of the drift process is simply the bias, a number between zero and one that summarizes the state of the population.
The theory of genetic drift predicts a number of measurable properties.For example, one can calculate the expected time until all or no members of a population possess a particular gene variant.These final states are referred to as fixation and deletion, respectively.Variation due to sampling vanishes once these states are reached and, for all practical purposes, drift stops.From then on, the population is homogeneous; further sampling can introduce no genotypic variation.These states are fixed points-in fact, absorbing states-of the drift stochastic process.
The analytical predictions for the time to fixation and time to deletion were developed by Kimura and Ohta [5,11] in the 1960s and are based on the memoryless models and simplifying assumptions introduced by Wright [12] and Fisher [13] in the early 1930s.The theory has advanced substantially since then to handle more realistic models and to predict additional effects due to selection and mutation.These range from multi-allele drift models and F -statistics [14] to pseudohitchhiking models of "genetic draft" [15].
The following explores what happens when we relax the memoryless restriction.The original random walk model of genetic drift forces the statistical structure at each sampling step to be an independent, identically distributed (IID) stochastic process.This precludes any memory in the sampling.Here, we extend the IID theory to use time-varying probabilistic state machines to describe memoryful population sampling.
In the larger setting of sequential learning, we will show that memoryful sequential sampling exhibits structurally complex, drift-like behavior.We call the resulting phenomenon structural drift.Our extension presents a number of new questions regarding the organization of the space of drift processes and how they balance structure and randomness.To examine these questions, we require a more precise description of the original drift theory.

III. GENETIC DRIFT
We begin with the definition of an allele, which is one of several alternate forms of a gene.The textbook example is given by Mendel's early experiments on heredity [16], in which he observed that the flowers of a pea plant were colored either white or violet, this being determined by the combination of alleles inherited from its parents.A new, mutant allele is introduced into a population by the mutation of a wild-type allele.A mutant allele can be passed on to an individual's offspring who, in turn, may pass it on to their offspring.Each inheritance occurs with some probability.
Genetic drift, then, is the change of allele frequencies in a population over time.It is the process by which the number of individuals with an allele varies generation after generation.The Fisher-Wright theory [12,13] models drift as a stochastic evolutionary process with neither selection nor mutation.It assumes random mating between individuals and that the population is held at a finite, constant size.Moreover, successive populations do not overlap in time.
Under these assumptions the Fisher-Wright theory reduces drift to a binomial or multinomial sampling process-a more complicated version of familiar random walks such as Gambler's Ruin or Prisoner's Escape [17].
Offspring receive either the wild-type allele A 1 or the mutant allele A 2 of a particular gene A from a random parent in the previous generation with replacement.A population of N diploid1 individuals will have 2N total copies of these alleles.Given i initial copies of A 2 in the population, an individual has either A 2 with probability i /2N or A 1 with probability 1− i /2N.The probability that j copies of A 2 exist in the offspring's generation given i copies in the parent's generation is: This specifies the transition dynamic of the drift stochastic process over the discrete state space {0, 1 /N, . . ., N −1 /N, 1}.This model of genetic drift is a discrete-time random walk, driven by samples of a biased coin, over the space of biases.The population is a set of coin flips, where the probability of Heads or Tails is determined by the coin's current bias.After each generation of flips, the coin's bias is updated to reflect the number of Heads or Tails realized in the new generation.The walk's absorbing states-all Heads or all Tails-capture the notion of fixation and deletion.

IV. GENETIC FIXATION
Fixation occurs with respect to an allele when all individuals in the population carry that specific allele and none of its variants.Restated, a mutant allele A 2 reaches fixation when all 2N alleles in the population are copies of A 2 and, consequently, A 1 has been deleted from the population.This halts the random fluctuations in the frequency of A 2 , assuming A 1 is not reintroduced.
Let X be a binomially distributed random variable with bias probability p that represents the fraction of copies of A 2 in the population.The expected number of copies of A 2 is E[X] = 2N p.That is, the expected number of copies of A 2 remains constant over time and depends only on its initial probability p and the total number (2N ) of alleles in the population.However, A 2 eventually reaches fixation or deletion due to the change in allele frequency introduced by random sampling and the presence of absorbing states.Prior to fixation, the mean and variance of the change in allele frequency ∆p are: (2) respectively.
On average there is no change in frequency.However, sampling variance causes the process to drift towards the absorbing states at p = 0 and p = 1.The drift rate is determined by the current generation's allele frequency and the total number of alleles.For the neutrally selective case, the average number of generations until fixation (t 1 ) or deletion (t 0 ) is given by Kimura and Ohta [5]: where N e denotes effective population size.For simplicity we take N e = N , meaning all individuals in the population are candidates for reproduction.As p → 0, the boundary condition is given by: That is, excluding cases of deletion, an initially rare mutant allele spreads to the entire population in 4N e generations.
One important consequence of the theory is that when fixation (p = 1) or deletion (p = 0) are reached, variation in the population vanishes: Var[∆p] = 0.With no variation there is a homogeneous population, and sampling from this population produces the same homogeneous population.In other words, this establishes fixation and deletion as absorbing states of the stochastic sampling process.Once there, drift stops.
Figure 1 illustrates this, showing both the simulated and theoretically predicted number of generations until fixation occurs for N = 10, as well as the predicted time to deletion for reference.Each simulation was performed for a different initial value of p and averaged over 400 realizations.Using the same methodology as Kimura and Ohta [5], we include only those realizations whose mutant allele reaches fixation.
Populations are produced by repeated binomial sampling of 2N uniform random numbers between 0 and 1.An initial probability 1 − p is assigned to allele A 1 and probability p to allele A 2 .The count i of A 2 in the initial population is incremented for each random number less than p.This represents an individual acquiring the allele A 2 instead of A 1 .The maximum likelihood estimate of allele frequency in the initial sample is simply the number of A 2 alleles over the sample length: p = i /2N.This estimate of p is then used to generate a new population of offspring, after which we re-estimate the value of p.These steps are repeated each generation until fixation at p = 1 or deletion at p = 0 occurs.This is the Monte Carlo (MC) sampling method.
Kimura's theory and simulations predict the time to fixation or deletion of a mutant allele in a finite population by the process of genetic drift.The Fisher-Wright model and Kimura's theory assume a memoryless population in which each offspring inherits allele A an IID binomial sampling process.We now generalize this to memoryful stochastic processes, giving a new definition of fixation and exploring examples of structural drift behavior.

V. SEQUENTIAL LEARNING
How can genetic drift be a memoryful stochastic process?Consider a population of N haploid organisms.Each generation consists of N alleles and so is represented by a string of N symbols, e.g.A 1 A 2 . . .A 1 A 1 , where each symbol corresponds to an individual with a particular allele.In the original drift models, a generation of offspring is produced by a memoryless binomial sampling process, selecting an offspring's allele from a parent with replacement.In contrast, the structural drift model produces a generation of individuals in which the sample order is tracked.The population is now a string of alleles, giving the potential for memory and structure in sampling-spatial, temporal, or other interdependencies between individuals within a sample.
At first, this appears as a major difference from the usual setting employed in evolutionary biology, where populations are treated as unordered collections of individuals and sampling is modeled as an independent, identically distributed stochastic process.That said, the structure we have in mind has several biological interpretations, such as inbreeding and subdivision [18] or the life histories of heterogeneous populations [19].We later return to these alternative interpretations when considering applications.
The model class we select to describe memoryful sampling is the -machine: the unique, minimal, and optimal representation of a stochastic process [4].As we will show, these properties give an important advantage when analyzing structural drift, since they allow one to monitor the amount of structure innovated or lost during drift.We next give a brief overview of -machines and refer the reader to the previous reference for details.
-Machine representations of the finite-memory discretevalued stochastic processes we consider here form a class of (deterministic) probabilistic finite-state machine or unifilar hidden Markov model.An -machine consists of a set of causal states S = {0, 1, . . ., k − 1} and a set of per-symbol transition matrices: where A = {A 1 , . . ., A m } is the set of alleles and where the transition probability gives the probability of transitioning from causal state S i to causal state S j and emitting allele a.The causal state probability Pr(σ), σ ∈ S, is determined as the left eigenvector of the stateto-state transition matrix T = a∈A T (a) .
Maintaining our connection to (haploid) population dynamics, we think of an -machine as a generator of populations or length-N strings: As a model of a sampling process, an -machine gives the most compact representation of the distribution of strings produced by sampling.Consider a simple binary process that alternately generates 0s and 1s called the Alternating Process shown in Fig. 3. Its -machine generates either the string 0101 . . .or 1010 . . .depending on the start state.The per-symbol transition matrices are: Enforcing the alternating period-2 pattern requires two states, A and B, as well as two positive probability transitions T (0) AB = 1.0 and T BA = 1.0.We are now ready to describe sequential learning, depicted in Fig. 5.We begin by selecting an initial population generator M 0 -an -machine.Following a path through M 0 , guided by its transition probabilities, produces a length-N string α N 0 = a 1 . . .a N that represents the first population of N individuals possessing alleles a i ∈ A. We then infer an -machine M 1 from the population α N 0 .M 1 is then used to produce a new population α N  1 , from which a new -machine M 2 is estimated.This new population has the same allele distribution as the previous, plus some amount of variance.The cycle of inference and re-inference is repeated while allele frequencies drift each generation until fixation or deletion is reached.At that point, the populations (and so -machines) cannot vary further.The net result is a stochastically varying time series of -machines (M 0 , M 1 , M 2 , . ..) that terminates when the populations α N t stop changing.Thus, at each step a new representation or model is estimated from the previous step's sample.The inference step highlights that this is learning: a model of the generator is estimated from the given finite data.The repetition of this step creates a sequential communication chain.Sequential learning is thus closely related to genetic drift except that sample order is tracked, and this order is used in estimating the next model.
The procedure is analogous to flipping a biased coin a number of times, estimating the bias from the results, and re-flipping the newly biased coin.Eventually, the coin will be completely biased towards Heads or Tails.In our drift model the coin is replaced by an -machine, which removes the IID model constraint and allows for the sampling process to take on structure and memory.Not only do the transition probabilities T (a) ij change, but the structure of the model itself -the number of states and the presence or absence of transitions-drifts over time to capture the statistics of the sample using as little information as possible.This is an essential and distinctive aspect of structural drift.
Before we can explore this dynamic, we first need to examine how an -machine reaches fixation or deletion.

VI. STRUCTURAL STASIS
Recall the Alternating Process from Fig. 3, producing the strings 0101 . . .and 1010 . . .depending on the start state.Regardless of the initial state, the original -machine is re-inferred from any sufficiently long string it produces.In the context of sequential learning, this means the population at each generation is the same.
However, if we consider allele A 1 to be represented by symbol 0 and A 2 by symbol 1, neither allele reaches fixation or deletion according to current definitions.Nonetheless, the Alternating Process prevents any variance between generations and so, despite the population not being all 0s or all 1s, the population does reach an equilibrium: half 0s and half 1s.For these reasons, one cannot use the original population-dynamics definitions of fixation and deletion.This leads us to introduce structural stasis to combine the notions of fixation, deletion, and the inability to vary caused by periodicity.Said more directly, structural stasis corresponds to a process becoming nonstochastic, since it ceases to introduce variance between generations and so prevents further drift.However, we need a method to detect the occurrence of structural stasis in a drift process.
A state machine representing a periodic sampling process enforces the constraint of periodicity via its internal memory.One measure of this memory is the population diversity H(N ) [20]: where the units are [bits]. 2The population diversity of the Alternating Process is H(N ) = 1 bit at any size N 1.This single bit of information corresponds to the machine's current phase or state.Generally, though, the value diverges-H(N ) ∝ N -for arbitrary sampling processes, and so population diversity is not suitable as a general test for stasis.
Instead, the condition for stasis can be given as the vanishing of the growth rate of population diversity: Equivalently, we can test the per-allele entropy of the sampling process.We call this allelic entropy: where the units are [bits per allele].Allelic entropy gives the average information per allele in bits, and structural stasis occurs when h µ = 0.While closer to a general test for stasis, this quantity is difficult to estimate from population samples since it relies on an asymptotic estimate of the population diversity.However, the allelic entropy can be calculated in closed-form from the -machine representation of the sampling process: When h µ = 0, the sampling process has become periodic and lost all randomness generated via its branching transitions.This new criterion subsumes the notions of fixation and deletion as well as periodicity.An -machine has zero allelic entropy if any of these conditions occur.
More formally, we have the following statement.
Proposition 1. Structural stasis is a fixed point of finite-memory structural drift.
Proof.Finite-memory means that the -machine representing the population sampling process has a finite number of states.Given this, if h µ = 0, then the -machine has no branching in its recurrent states: T (a) ij = 0 or 1, where S i and S j are asymptotically recurrent states.This results in no variation in the inferred -machine when sampling sufficiently large populations.Lack of variation, in turn, means that ∆p = 0 and so the drift process stops.If allelic entropy vanishes at time t and no mutations are allowed, then it is zero for all t > t.Thus, structural stasis is an absorbing state of the drift stochastic process.

VII. EXAMPLES
While more can be said analytically about structural drift, our present purpose is to introduce the main concepts.We will show that structural drift leads to interesting and nontrivial behavior.First, we calibrate the new class of drift processes against the original genetic drift theory.

A. Memoryless Drift
The Biased Coin Process is represented by a singlestate -machine with a self loop for both Heads and Tails symbols.It is an IID sampling process that generates populations with a binomial distribution.Unlike the Alternating Process, the coin's bias p is free to drift during sequential inference.These properties make the Biased Coin Process an ideal candidate for exploring memoryless drift.
Fig. 6 shows structural drift, using two different measures, for a single realization of the Biased Coin Process with initial p = Pr[Heads] = Pr[Tails] = 0.5.Structural stasis (h µ = 0) is reached after 115 generations.The initial Fair Coin -machine occurs at the left of Fig. 6 and the final, completely biased -machine occurs at the right.
Note that the drift of allelic entropy h µ and p = Pr[Tails] are inversely related, with allelic entropy converging quickly to zero as stasis is approached.This reflects the rapid drop in population diversity.After stasis occurs, all randomness has been eliminated from the transitions at state A, resulting in a single transition that always produces Tails.Anticipating later discussion, we note that during this run only Biased Coin Processes were observed.
The time to stasis of the Biased Coin Process as a function of initial p = Pr[Heads] was shown in Fig. 2. Also shown there was the previous Monte Carlo Kimura drift simulation modified to terminate when either fixation or deletion occurs.This experiment illustrates the definition of structural stasis and allows direct comparison of structural drift with genetic drift in the memoryless case.Not surprisingly, we can interpret genetic drift as a special case of the structural drift process for the Biased Coin.Both simulations follow Kimura's theoretically predicted curves, combining the lower half of the deletion curve with the upper half of the fixation curve to reflect the initial probability's proximity to the absorbing states.A high or low initial bias leads to a shorter time to stasis as the absorbing states are closer to the initial state.Similarly, a Fair Coin is the furthest from absorption and thus takes the longest average time to reach stasis.

B. Structural Drift
The Biased Coin Process represents an IID sampling process with no memory of previous flips, reaching stasis when Pr[Heads] = 1.0 or 0.0 and, correspondingly, when h µ (M t ) = 0.0.We now introduce memory by starting drift with M 0 as the Golden Mean Process, which produces binary populations with no consecutive 0s.Its -machine was shown in Fig. 4. Note that one can initialize drift using any stochastic process; for example, see the -machine library of Ref. [22].
Like the Alternating Process, the Golden Mean Process has two causal states.However, the transitions from state A have nonzero entropy, allowing their probabilities to drift as new -machines are inferred from generation to generation.If the A → B transition probability p (Fig. 4) becomes zero the transition is removed, and the Golden Mean Process reaches stasis by transforming into the Fixed Coin Process (top right, Fig. 6).Instead, if the same transition drifts towards probability p = 1, the A → A transition is removed.In this case, the Golden Mean Process reaches stasis by transforming into the Alternating Process (Fig. 3).
To compare structural drift behaviors, consider also the Even Process.Similar in form to the Golden Mean Process, the Even Process produces populations in which blocks of consecutive 1s must be even in length when bounded by 0s [21].Figure 7 compares the drift of Pr[Heads] for a single realization of the Biased Coin, Golden Mean, and Even Processes.One observes that the Even and Biased Coin Processes reach stasis as the Fixed Coin Process, while the Golden Mean Process reaches stasis as the Alternating Process.For different realizations, the Even and Golden Mean Processes might instead reach different stasis points.
It should be noted that the memoryful Golden Mean and Even Processes reach stasis markedly faster than the memoryless Biased Coin.While Fig. 7 shows only a sin- gle realization of each sampling process type, the top panel of Fig. 9 shows the large disparity in stasis times holds across all settings of each process's initial bias.This is one of our first general observations about memoryful processes: The structure of memoryful processes substantially impacts the average time to stasis by increasing variance between generations.In the cases shown, time to stasis is greatly shortened.

VIII. ISOSTRUCTURAL SUBSPACES A. Subspace Diffusion
To illustrate the richness of structural drift and to understand how it affects average time to stasis, we examine the complexity-entropy (CE) diagram [23] of the -machines produced over several realizations of an arbitrary sampling process.The CE diagram displays how the allelic entropy h µ of an -machine varies with the allelic complexity C µ of its causal states: where the units are [bits].The allelic complexity is the Shannon entropy over an -machine's stationary state distribution Pr(S).It measures the memory needed to maintain the internal state while producing stochastic outputs.-Machine minimality guarantees that C µ is the smallest amount of memory required to do so.Since there is a one-to-one correspondence between processes and their -machines, a CE diagram is a projection of process space onto the two coordinates (h µ , C µ ).Used in tandem, these two properties differentiate many types of sampling process, capturing both their intrinsic memory (C µ ) and the diversity (h µ ) of populations they generate.Two such CE diagrams are shown in Fig. 8, illustrating different subspaces and stasis points reachable by the Golden Mean Process during structural drift.Consider the left panel first.An -machine reaches stasis by transforming into either the Fixed Coin or the Alternating Process.To reach the former, the -machine begins on the upper curve in the left panel and drifts until the A → B transition probability nears zero and the inference algorithm decides to merge states in the next generation.This forces the -machine to jump to the Biased Coin subspace on the line C µ = 0 where it will most likely diffuse until the Fixed Coin stasis point at (h µ , C µ ) = (0, 0) is reached.If instead the A → B transition probability drifts towards zero, the Golden Mean stays on the upper curve until reaching the Alternating Process stasis point at (h µ , C µ ) = (0, 1).Thus, the two stasis points are differentiated not by h µ but by C µ , with the Alternating Process requiring 1 bit of memory to track its internal state and the Biased Coin Process requiring none.
What emerges from these diagrams is a broader view of how population structure drifts in process space.Roughly, the M t diffuse locally in the parameter space specified by the current, fixed architecture of states and transitions.During this, transition probability estimates vary stochastically due to sampling variance.Since C µ and h µ are continuous functions of the transition probabilities, this variance causes the M t to fall on well defined curves or regions corresponding to a particular process subspace.(See Figs. 4 and 5 in Ref. [23] and the theory for these curves and regions there.) We refer to these curves as isostructural curves and the associated sets of -machines as isostructural sub-spaces.They are metastable subspaces of sampling processes that are quasi-invariant under the structural drift dynamic.When one or more -machine parameters diffuse sufficiently so that inference is forced to change topology by adding or removing states or transitions to reflect the statistics of the sample, this quasiinvariance is broken.We call such topological shifts subspace jumps to reflect the new region occupied by the resulting -machine in process space, as visualized by the CE diagram.Movement between subspaces is often not bidirectional-innovations from a previous topology may be lost either temporarily (when the innovation can be restored by returning to the subspace) or permanently.For example, the Golden Mean subspace commonly jumps to the Biased Coin subspace but the opposite is highly improbable without mutation.(We consider the latter type of structured drift elsewhere.) Before describing the diversity seen in the CE diagram of Fig. 8's right panel, we first turn to analyze in some detail the time-to-stasis underlying the behavior illustrated in the left panel.

B. Subspace Decomposition
A pathway is a set of subspaces passed through by any drift realization starting from some initial process and reaching a specific stasis point.The time to stasis of a drift process P is the sum of time spent in the subspaces γ visited by its pathways to stasis ρ, weighted by the probabilities that these pathways and subspaces will be reached.The time spent in a subspace γ i+1 only depends on the transition parameter(s) of the -machine at the time of entry and is otherwise independent of the prior subspace γ i .Thus, calculating the stasis time of a structured population can be broken down into independent subspace times when we know the values of the transition parameters at subspace jumps.These values can be derived both empirically and analytically, and we aim to develop the latter for general drift processes in future work.
More formally, the time-to-stasis t s of a drift process P is simply the weighted sum of the stasis times for its connected pathways ρ: Similarly, the stasis time of a particular pathway decomposes into the time spent diffusing in its connected subspaces γ: To demonstrate, Fig. 9 shows the stasis time of the Golden Mean Process (GMP) with initial bias p 0 in more detail.Regression lines along with their 95% confidence  intervals are displayed for simulations with initial biases p 0 ∈ {0.1, 0.2, . . ., 0.9}.The middle panel shows the total time-to-stasis as the weighted sum of its Fixed Coin (FC) and Alternating Process (AP) pathways: t s (GMP(p 0 )) = Pr(FC|GMP(p 0 ))t s (FC|GMP(p 0 )) + Pr(AP|GMP(p 0 ))t s (AP|GMP(p 0 )) .
For low p 0 , the transition from state A to state B is unlikely, so 0s are rare and the AP pathway is reached infrequently.Thus, the total stasis time is initially dominated by the FC pathway (Pr(FC|GMP(p 0 )) is high).
As p 0 → 0.3 and above, the AP pathway is reached more frequently (Pr(AP|GMP(p 0 )) grows) and its stasis time begins to influence the total.The FC pathway is less likely as p 0 → 0.6 and the total time becomes dominated by the AP pathway (Pr(AP|GMP(p 0 )) is high).
Since the AP pathway visits only one subspace, the bottom panel shows the stasis time of the FC pathway as the weighted sum of the Golden Mean (GM) and Biased Coin (BC) subspace times: This corresponds to time spent diffusing in the GM subspace before the subspace jump and time spent diffusing in the BC subspace after the subspace jump.Note that the times quoted are simply diffusion times within a subspace, since not every subspace in a pathway contains a stasis point.These expressions emphasize the dependence of stasis time on the transition parameters at jump points as well as on the architecture of isostructural subspaces in drift process space.For example, if the GM jumps to the BC subspace at p = 0.5, the stasis time will be large since the -machine is maximally far from either stasis point.However, the inference algorithm will typically jump at very low values of p resulting in a small average stasis time for the BC subspace in the FC pathway.Due to this, calculating the stasis time for the GMP requires knowing the AP and FC pathways as well as the value of p where the GM → BC jump occurs.

C. Structural Innovation and Loss
Inference of -machines from finite populations is computationally expensive, particularly in a sequential setting with many realizations.
The topology of the -machine is inferred directly from the statistics of finite samples; both states and transitions are added and removed over time to capture innovation and loss of population structure.In the spirit of Kimura's pseudo-sampling variable method [24], we introduce a pseudo-drift algorithm for efficient drift simulation and increased control of the trade-off between structural innovation and loss.
Instead of inferring and re-inferring an -machine each generation, we explicitly define the conditions for topological changes to the -machine of the previous generation.To test for structural innovation, a random causal state from the current M t is cloned and random incoming transitions are routed instead to the cloned state.This creates a new model M t that describes the same process.Gaussian noise is then added to the cloned state's outgoing transitions to represent change in population structure.The likelihood of the population α N t is calculated for both M t and M t and the model with the maximum a posteriori (MAP) likelihood is retained: If the original M t was retained, its transition parameters are updated by feeding the sample through the model to obtain edge counts which are normalized to obtain probabilities.This produces a generator for the next generation's population in a way that allows for innovation.As well, it side-steps the computational cost of the inference algorithm.
To capture structural loss, we monitor near-zero transition probabilities where the -machine inference algorithm would merge states.When such a transition exists we test for structural simplification by considering all pairwise mergings of causal states and select the topology via the MAP likelihood.However, unlike above, we penalize likelihood using the Akaike Information Criterion (AIC) [25]: and, in particular, the AIC corrected for finite sample sizes [26]: where k is the number of model parameters, L is the sample likelihood, and n is the sample size.A penalized likelihood is necessary because a smaller -machine is more general and cannot fit the data as well.When penalized by model size, however, a smaller model with sufficient fit to the data may be selected over a larger, better fitting model.This method allows loss to occur while again avoiding the expense of the full -machine inference algorithm.Extensive comparisons with several versions of the latter show that the new pseudo-drift algorithm produces qualitatively the same behavior.
Having explained how the pseudo-drift algorithm introduces structural innovation and loss we can now describe the drift runs of Fig. 8's right panel.In contrast to the left panel, structural innovation was enabled.The immediate result is that the drift process visits a much wider diversity of isostructural subspaces-sampling processes that are markedly more complex.-Machines with 8 or more states are created, some of which are quite entropic and so produce high sampling variance.Stasis -machines with periods 3, 4, 5, and 6 are seen, while only those with periods 1 and 2 are seen in runs without innovation (left panel).
By way of closing this first discussion of structural drift, it should be emphasized that none of the preceding phenomena occur in the limit of infinite populations or infinite sample size.The variance due to finite sampling drives sequential learning, the diffusion through process space, and the jumps between isostructural subspaces.

IX. APPLICATIONS AND EXTENSIONS
Much of the previous discussion focused on structural drift as a kind of stochastic process, with examples and behaviors selected to emphasize the role of structure.Although there was a certain terminological bias toward neutral evolution theory since the latter provides an entree to analyzing how structural drift works, our presentation was intentionally general.Motivated by a variety of potential applications and extensions, we describe these now and close with several summary remarks on structural drift itself.

A. Emergent Semantics and Learning in Communication Chains
Let's return to draw parallels with the opening example of the game of Telephone or, more directly, to the sequential inference of temporal structure in an utterance passed along a serially coupled communication chain.There appears to be no shortage of related theories of language evolution.These range from the population dynamics of Ref. [27] and the ecological dynamics of Ref. [28] to the cataloging of error sources in human communication [29] and recent efforts to understand cultural evolution as reflecting learning biases [30,31].
By way of contrast, structural drift captures the language-centric notion of dynamically changing semantics and demonstrates how behavior is driven by finitesample fluctuations within a semantically organized subspace.The symbols and words in the generated strings have a semantics given by the structure of a subspace's -machine; see Ref. [3].A particularly simple example was identified quite early in the informationtheoretic analysis of natural language: The Golden Mean -machine (Fig. 4) describes the role of isolated space symbols in written English [32, Fig. 1].Notably, this structure is responsible for the Mandelbrot-Zipf powerlaw scaling of word frequencies [33,34].More generally, though, the semantic theory of -machines shows that causal states provide dynamic contexts for interpretation as individual symbols and words are recognized.Quantitatively, the allelic complexity C µ (M t ) is the total amount of semantic content that can be generated by an M t [3].In this way, shifts in the architecture of the M t during drift correspond to semantic changes.That is, diffusion within an isostructural subspace corresponds to constant semantics, while jumps between isostructural subspaces correspond to semantic innovations (or losses).
In the drift behaviors explored above, the M t went to stasis (h µ = 0) corresponding to periodic formal languages.Clearly, such a long-term condition falls far short as a model of human communication chains.The resulting communications, though distant from those at the beginning of the chain, are not periodic.To more closely capture emergent semantics in the context of sequential language learning, we have extended structural drift to include mutation and selection.In future work we will use these extensions to investigate how the former prevents permanent stasis and the latter enables a preference for intelligible phrases.

B. Cultural Evolution and Iterated Learning
Extending these observations, the Iterated Learning Model (ILM) of language evolution [35,36] is of particular interest.In this model, a language evolves by repeated production and acquisition by agents under cultural pressures and the "poverty of the stimulus" [35].Via this process, language is effectively forced through a transmission bottleneck that requires the learning agent to generalize from finite data.This, in turn, exerts pressure on the language to adapt to the bias of the learner.Thus, in contrast to traditional views that the human brain evolved to learn language, ILM suggests that language also adapts to be learnable by the human brain.
ILM incorporates the sequential learning and propagation of error we discuss here and provides valuable insight into the effects of error and cultural mutations on the evolution of language for the "human niche".There are various simulation approaches to ILM with both single and multiple agents based on, for example, neural networks and Bayesian inference, as well as experiments with human subjects.We suggest that structural drift could also serve as the basis for single-agent ILM experiments, as found in Swarup et al. [37], where populations of alleles in the former are replaced by linguistic features of the latter.The benefits are compelling: an informationtheoretic framework for quantifying the trade-off between learner bias and transmission bottleneck pressures, visualization of cultural evolution via the CE diagram, and decomposition of the time-to-stasis of linguistic features in terms of isostructural subspaces as presented above.

C. Epochal Evolution
Beyond applications to knowledge transmission via serial communication channels, structural drift gives an alternative view of drift processes in population genetics.In light of new kinds of evolutionary behavior, it reframes the original questions about underlying mechanisms and extends their scope to phenomena that exhibit memory in the sampling process or that derive from structure in populations.Examples of the latter include niche construction [38], the effects of environmental toxins [39], changes in predation [40], and socio-political factors [41] where memory lies in the spatial distribution of populations.In addition to these, several applications to areas beyond population genetics proper suggest themselves.
An intriguing parallel exists between structural drift and the longstanding question about the origins of punctuated equilibrium [42] when modeled as the dynamics of epochal evolution [43,44].The possibility of evolution's intermittent progress-long periods of stasis punctuated by rapid change-dates back to Fisher's demonstration of metastability in drift processes with multiple alleles [13].
Epochal evolution, though, presented an alternative to the views of metastability posed by Fisher's model and Wright's adaptive landscapes [45].Within epochal evolutionary theory, equivalence classes of genotype fitness, called subbasins, are connected by fitness-changing portals to other subbasins.A genotype is free to diffuse within its subbasin via selectively neutral mutations, until an advantageous mutation drives genotypes through a portal to a higher-fitness subbasin.An increasing number of genotypes derive from this founder and diffuse in the new subbasin until another portal to higher fitness is discovered.Thus, the structure of the subbasin-portal architecture dictates the punctuated dynamics of evolution.
Given an adaptive system which learns structure by sampling its past organization, structural drift theory implies that its evolutionary dynamics are inevitably described by punctuated equilibria.Diffusion in an isostructural subspace corresponds to a period of structured equilibrium in a subbasin and subspace jumps correspond to rapid innovation or loss of organization during the transit of a portal.In this way, structural drift establishes a connection between evolutionary innovation and structural change, identifying the conditions for creation or loss of organization.Extending structural drift to include mutation and selection will provide a theoretical framework for epochal evolution using any number of structural constraints in a population.

D. Evolution of Graph-Structured Populations
We focused primarily on the drift of sequentially ordered populations in which the generator (an -machine) captured the structure and randomness in that ordering.We aimed to show that a population's organization plays a crucial role in its dynamics.This was, however, only one example of the general class of drift process we have in mind.For example, computational mechanics also describes structure in spatially extended systems [46,47].Given this, it is straightforward to build a model of drift in geographically distributed populations that exhibit spatiotemporal structure.
Though they have not tracked the structural complex-ity embedded in populations as we have done here, a number of investigations consider various classes of structured populations.For example, the evolutionary dynamics of structured populations have been studied using undirected graphs to represent correlations between individuals.Edge weights w ij between individuals i and j give the probability that i will replace j with its offspring when selected to reproduce.By studying fixation and selection behavior on different types of graphs, Lieberman et al found that graph structures can sometimes amplify or suppress the effects of selection, even guaranteeing the fixation of advantageous mutations [48].Jain and Krishna [49] investigated the evolution of directed graphs and the emergence of self-reinforcing autocatalytic networks of interaction.They identified the attractors in these networks and demonstrated a diverse range of behaviors from the creation of structural complexity to its collapse and permanent loss.
Graph evolution is a model of population structure complementary to that presented by structural drift.In the latter, -machine structure evolves over time with nodes representing equivalence classes of the distribution of selectively neutral alleles.Unlike -machines, the multinomial sampling of individuals in graph evolution is a memoryless process.A combined approach will allow one to examine how amplification and suppression of selection and the emergence of autocatalysis are affected by external influences on the population structure.For example, this could include how a population uses temporal memory to maintain desirable properties in anticipation of structural shifts in the environment.The result would provide a theory for niche construction in which a nonlinear dynamics of pattern formation spontaneously changes population structure.

X. FINAL REMARKS
The Fisher-Wright model of genetic drift can be viewed as a random walk of coin biases, a stochastic process that describes generational change in allele frequencies based on a strong statistical assumption: The sampling process is memoryless.Here, we developed a generalized structural drift model that adds memory to the process and examined the consequences of such population sampling memory.
Memoryful sampling is a substantial departure from modeling evolutionary processes with unordered populations.Rather than view structural drift as a replacement for the well understood theory of genetic drift, and given that the latter is a special case of structurally drifting populations, we propose that it be seen as a new avenue for theoretical invention.Given its additional ties to language and cultural evolution, we believe it will provide a novel perspective on evolution in nonbiological domains, as well.
The representation selected for the population sam-pling mechanism was the class of probabilistic finitestate hidden Markov models called -machines.We discussed how a sequential chain of -machines inferred and re-inferred from the finite data they generate parallels the drift of alleles in a finite population, using otherwise the same assumptions made by the Fisher-Wright model.The mathematical foundations developed for the latter and its related models provide a good deal of quantitative, predictive power.Much of this has yet to be exploited.In concert with this, -machine minimality allowed us to monitor information processing, information storage, and causal architecture during the drift process.We introduced the information-theoretic notion of structural stasis to combine the concepts of deletion, fixation, and periodicity for drift processes.Generally, structural stasis occurs when the population's allelic entropy vanishes-a quantity one can calculate in closed form due to the -machine representation of the sampling process.We revisited Kimura and Ohta's early results measuring the time to fixation of drifting alleles and showed that the generalized structural drift process reproduces these well known results when staying within the memoryless sampling process subspace.Starting with structured populations outside of that subspace led the sampling process to exhibit memory effects including structural innovation and loss, complex transients, and greatly reduced stasis times.
Simulations demonstrated how an -machine diffuses through isostructural process subspaces during sequential learning.The result was a very complex time-tostasis dependence on the initial probability parametermuch more complicated than Kimura's theory describes.Nonetheless, we showed that a process' time-to-stasis can be decomposed into sums over these independent subspaces.Moreover, the time spent in an isostructural subspace depends on the value of the -machine probability parameters at the time of entry.This suggests an extension to Kimura's theory for predicting the time to stasis for each isostructural component independently.Much of the phenomenological analysis was facilitated by the global view of drift process space given by the complexity-entropy diagram.
Drift processes with memory generally describe the evolution of structured populations without mutation or selection.Nonetheless, we showed that structure leads to substantially shorter stasis times.This was seen in drifts starting with the Biased Coin and Golden Mean Processes, where the Golden Mean jumps into the Biased Coin subspace close to an absorbing state.This suggests that even without selection, population structure and sampling memory matter in evolutionary dynamics.The temporal or spatial memory captured by the -machine can be interpreted as nonrandom mating, reducing the effective population size N e and, in doing so, increasing sampling variance.It also suggests that memoryless models restrict sequential learning and overestimate stasis times for structured populations.
We demonstrated how structural drift-diffusion, structural innovation and loss-are controlled by the architecture of connected isostructural subspaces.Many questions remain about these subspaces.What is the degree of subspace-jump irreversibility?Can we predict the likelihood of these jumps?What does the phase portrait of a drift process look like?Thus, to better understand structural drift, we need to analyze the high-level organization of generalized drift process space.
Fortunately, -machines are in one-to-one correspondence with structured processes [22].Thus, the preceding question reduces to understanding the space of -machines and how they can be connected by diffusion processes.Is the diffusion within each process subspace predicted by Kimura's theory or some simple variant?We have given preliminary evidence that it does.And so, there are reasons to be optimistic that in face of the open-ended complexity of structural drift, a good deal can be predicted analytically.And this, in turn, will lead to quantitative applications.

FIG. 3 .| 1 FIG. 4 .
FIG. 3. -Machine for the Alternating Process consisting of two causal states S = {A, B} and two transitions.Each transition is labeled p | a to indicate the probability p = T (a) ij of taking that transition and emitting allele a ∈ A. State A emits allele 0 with probability one and transitions to state B, while B emits allele 1 with probability one and transitions to A.

FIG. 5 .
FIG.5.Sequential inference of a chain of -machines.An initial population generator M0 produces a length-N string α N 0 = a1 . . .aN from which a new model M1 is inferred.These steps are repeated using M1 as the population generator and so on, until a terminating condition is met.

FIG. 6 .FIG. 7 .
FIG. 6. Drift of allelic entropy hµ and Pr[Heads] for a single realization of the Biased Coin Process with sample size N = 100.The drift of Pr[Heads] is annotated with its initial machine M0 (left inset) and the machine at stasis M115 (right inset).

FIG. 8 .
FIG. 8. Complexity-entropy diagram for 30 realizations of the Golden Mean Process with N = 1000, both without (left) and with (right) structural innovation.Alternating Process and Fixed Coin pathways are clearly visible in the left panel where the Golden Mean subspace exists on the upper curve and the Biased Coin subspace exists on the line Cµ = 0. -Machines within the same isostructural subspace have identical colors.
FIG. 9. Top: Time to stasis of the Golden Mean, Even, and Biased Coin Processes.Middle: Stasis time of the Golden Mean Process as the weighted sum of stasis times for the Fixed Coin (FC) and Alternating Process (AP) pathways.Bottom: Stasis time of the FC pathway as the weighted sum of Golden Mean (GM) and Biased Coin (BC) subspace diffusion times.
1 or A 2 via Monte Carlo (MC) simulation of Kimura's model.Kimura's predicted times to fixation and deletion are shown for reference.Each estimated time is averaged over 100 realizations with sample size N = 1000.