Entropic Dynamics: Mechanics without Mechanism

TL;DR

Entropic Dynamics derives physical laws from entropic inference without relying on traditional action principles, offering a universal inference-based framework potentially applicable beyond physics.

Contribution

This work reviews the derivation of quantum theory within Entropic Dynamics, highlighting its novel inference-based approach that does not depend on an underlying mechanism.

Findings

Quantum theory derived from entropic inference

Framework does not require an action principle

Potential applicability to fields beyond physics

Abstract

Entropic Dynamics is a framework in which dynamical laws such as those that arise in physics are derived as an application of entropic methods of inference. No underlying action principle is postulated. Instead, the dynamics is driven by entropy subject to constraints reflecting the information that is relevant to the problem at hand. In this work I review the derivation of quantum theory but the fact that Entropic Dynamics is based on inference methods that are of universal applicability suggests that it may be possible to adapt these methods to fields other than physics.

Equations

$$S(x)=-\int dyp(y|x)\log \frac{p(y|x)}{q(y)}.$$

$$\mathcal{S}[P,Q]=-\int d{x}^{\prime }d{y}^{\prime }P(x^{\prime },y^{\prime }|x,y)\log \frac{P(x^{\prime },y^{\prime }|x,y)}{Q(x^{\prime },y^{\prime }|x,y)},$$

$$Q(x^{\prime },y^{\prime }|x,y)=Q(x^{\prime }|x)Q(y^{\prime }|y).$$

$$S[Q]=-\int d{x}^{\prime }Q(\Delta x)\log \frac{Q(\Delta x)}{\mu },$$

$$⟨{\delta }_{ab}\Delta x_n^a\Delta x_n^b⟩={\kappa }_n,(n=1\dots N),$$

$$Q(x^{\prime }|x)\propto \exp -\frac{1}{2}{\sum }_n{\alpha }_n{\delta }_{ab}\Delta x_n^a\Delta x_n^b.$$

$$Q(x^{\prime },y^{\prime }|x,y)=q(y^{\prime })\exp -\frac{1}{2}{\sum }_n{\alpha }_n{\delta }_{ab}\Delta x_n^a\Delta x_n^b.$$

$$P(x^{\prime },y^{\prime }|x,y)=P(x^{\prime }|x,y)P(y^{\prime }|x^{\prime },x,y),$$

$$P(x^{\prime },y^{\prime }|x)=P(x^{\prime }|x)p(y^{\prime }|x^{\prime }).$$

$$\mathcal{S}[P,Q]=-\int d{x}^{\prime }P(x^{\prime }|x)\log \frac{P(x^{\prime }|x)}{Q(x^{\prime }|x)}+\int d{x}^{\prime }P(x^{\prime }|x)S(x^{\prime }),$$

$$P(x^{\prime }|x)=\frac{1}{\zeta }\exp \left[S(x^{\prime })-\frac{1}{2}{\sum }_n{\alpha }_n{\delta }_{ab}\Delta x_n^a\Delta x_n^b\right],$$

$$S(x^{\prime })=S(x)+{\sum }_n\Delta x_n^a\frac{\partial S}{\partial x_n^a}+\dots$$

$$P(x^{\prime }|x)=\frac{1}{Z}\exp \left[-\frac{1}{2}{\sum }_n{\alpha }_n{\delta }_{ab}\left(\Delta x_n^a-⟨\Delta x_n^a⟩\right)\left(\Delta x_n^b-⟨\Delta x_n^b⟩\right)\right],$$

$$\Delta x_n^a=⟨\Delta x_n^a⟩+\Delta w_n^a,\text{where}⟨\Delta x_n^a⟩=\frac{1}{{\alpha }_n}{\delta }^{ab}\frac{\partial S}{\partial x_n^b},$$

$$⟨\Delta w_n^a⟩=0\text{and}⟨\Delta w_n^a\Delta w_n^b⟩=\frac{1}{{\alpha }_n}{\delta }^{ab}.$$

$$P(x_i)=\int d{x}_{i-1}P(x_i,x_{i-1})=\int d{x}_{i-1}P(x_i|x_{i-1})P(x_{i-1}).$$

$$\rho (x^{\prime },t^{\prime })=\int dxP(x^{\prime }|x)\rho (x,t).$$

$$\rho (x,t)=\int d{x}^{\prime }P(x|x^{\prime })\rho (x^{\prime },t^{\prime }),$$

$$P(x|x^{\prime })=\frac{\rho (x,t)}{\rho (x^{\prime },t^{\prime })}P(x^{\prime }|x).$$

$${\alpha }_n=\frac{m_n}{\eta }\frac{1}{\Delta t}.$$

$${\gamma }_{AB}=C\int d{x}^{\prime }P(x^{\prime }|x)\frac{\partial \log P(x^{\prime }|x)}{\partial x^A}\frac{\partial \log P(x^{\prime }|x)}{\partial x^B}.$$

$${\gamma }_{AB}=\frac{C{m}_n}{\eta \Delta t}{\delta }_{n{n}^{\prime }}{\delta }_{ab}=\frac{C{m}_n}{\eta \Delta t}{\delta }_{AB}.$$

$$m_{AB}=\frac{\eta \Delta t}{C}{\gamma }_{AB}=m_n{\delta }_{AB}.$$

$$m^{AB}=\frac{C}{\eta \Delta t}{\gamma }^{AB}=\frac{1}{m_n}{\delta }^{AB}.$$

$$\Delta x^A=b^A\Delta t+\Delta w^A,$$

$$⟨\Delta x^A⟩=b^A\Delta t\text{with}{b}^A=\eta m^{AB}{\partial }_{B}S,$$

$$⟨\Delta w^A⟩=0\text{and}⟨\Delta w^A\Delta w^B⟩=\eta m^{AB}\Delta t.$$

$${\partial }_t\rho =-{\partial }_A\left(b^A\rho \right)+\frac{1}{2}\eta m^{AB}{\partial }_A{\partial }_B\rho .$$

$${\partial }_t\rho =-{\partial }_A\left(\rho v^A\right).$$

$$v^A=b^A+u^A.$$

Full text

Law without Law:

*“The only thing harder to understand than a law of statistical origin would be a law that is not of statistical origin, for then there would be no way for it — or its progenitor principles — to come into being.” *

Two tests:

*“No test of these views looks like being someday doable, nor more interesting and more instructive, than a derivation of the structure of quantum theory… No prediction lends itself to a more critical test than this, that every law of physics, pushed to the extreme, will be found statistical and approximate, not mathematically perfect and precise.” *

*J. A. Wheeler *[1]

1 Introduction

The drive to explain nature has always led us to seek the mechanisms hidden behind the phenomena. Descartes, for example, claimed to explain the motion of planets as being swept along in the flow of some vortices. The model did not work very well but at least it gave the illusion of a mechanical explanation and thereby satisfied a deep psychological need. Newton’s theory fared much better. He took the important step of postulating that gravity was a universal force acting at a distance but he abstained from offering any mechanical explanation — a stroke of genius immortalized in his famous “hypotheses non fingo.” At first there were objections. Huygens, for instance, recognized the undeniable value of Newton’s achievement but was nevertheless deeply disappointed: the theory works but it does not explain. And Newton agreed: any action at a distance would represent “so great an absurdity… that no man who has in philosophical matters a competent faculty of thinking can ever fall into it.” [2]

Over the following 18th century, however, impressed by the many successes of Newtonian mechanics, people started to downplay and then even forget their qualms about the absurdity of action at a distance. Mechanical explanations were, of course, still required but the very meaning of what counted as “mechanical” suffered a gradual but irreversible shift. It no longer meant “caused by contact forces” but rather “described according to Newton’s laws.” Over time Newtonian forces, including those mysterious actions at a distance, became “real” which qualified them to count as the causes behind phenomena.

But this did not last too long. With Lagrange, Hamilton, and Maxwell, the notion of force started to lose some of its fundamental status. Indeed, after Maxwell succeeded in extending the principles of dynamics to include the electromagnetic field, the meaning of ‘mechanical explanation’ changed once again. It no longer meant identifying the Newtonian forces, but rather identifying the right Lagrangian, the right action principle. Thus, today gravity is no longer explained through a force but through the curvature of space-time. And the concept of force finds no place in quantum mechanics where interactions are described as the evolution of abstract vectors in Hilbert space.

Our goal in this chapter is to derive useful dynamical models without invoking underlying mechanisms. This does not mean that such mechanisms do not exist, for all we know they might. It is just that useful models can be constructed without having to go through the trouble of keeping track of a myriad of microscopic details that often turn out to be ultimately irrelevant. The idea can be illustrated by contrasting the two very different ways in which the theory of Brownian motion was originally derived by Smoluchowski and by Einstein. In Smoluchowski’s approach one keeps track of the microscopic details of molecular collisions through a stochastic Langevin equation and a macroscopic effective theory is then derived by taking suitable averages. In Einstein’s approach, on the other hand, one focuses directly on those pieces of information that turn out to be relevant for the prediction of macroscopic effects. The advantage of Einstein’s approach is the simplicity that arises from not having to keep track of irrelevant details that are eventually washed out when taking the averages. The challenge, of course, is to identify those pieces of information that happen to be relevant.

Our argument proceeds by stages. We will exhibit three explicit examples. Starting with the simplest, we first tackle a standard diffusion, which allows us to introduce and discuss the non-trivial concept of time. In the next stage we discuss the derivation of Hamiltonian dynamics as a non-dissipative type of diffusion. Finally, in a further elaboration invoking notions of information geometry, we identify the particular Hamiltonian that leads to quantum mechanics.

Ever since its origin in 1925 the interpretation of quantum mechanics has been a notoriously controversial subject. The central question revolves around the interpretation of the quantum state, the wave function: does it represent the actual real state of the system (its ontic state) or does it represent a state of knowledge about the system (an epistemic state)? Excellent reviews with extended references to the literature are given in [3]-[6]. However, the urge to seek underlying mechanisms has motivated a huge effort towards identifying hidden variables and sub-quantum levels of reality from which an effective quantum dynamics might emerge. (See e.g., [7]-[13].) In this work we adopt the point of view that quantum theory itself provides us with the most important clue: the goal of quantum mechanics is to calculate probabilities. Therefore, by its very nature, quantum mechanics must be an instance of Bayesian and entropic inference.

Entropic Dynamics (ED) provides a framework for deriving dynamical laws as an application of entropic methods [14]-[17]. The use of entropy as a tool for inference can be traced to E. T. Jaynes [18]-[20]. For a pedagogical overview and further references see [21].

The literature on the attempts to reconstruct quantum mechanics is vast and there are several approaches that like ED are also based on information theory (see e.g., [22]-[31]). Since all these approaches must sooner or later converge to the same Schrödinger equation it is inevitable that they will show similarities but there are, however, important differences. What distinguishes ED is a strict adherence to Bayesian and entropic methods without invoking mechanisms operating at some deeper sub-quantum level.

This chapter is basically a self-contained review of material developed in [14]-[17] but with one significant change. The gist of Entropic Dynamics is that the system of interest (e.g., particles or fields) undergoes a diffusion with a special systematic drift which is ultimately responsible for such quintessential quantum phenomena as interference and entanglement. A central idea introduced in [14], but not fully developed in later publications, is that this peculiar drift is itself of entropic origin. More precisely, the drift follows the gradient of the entropy of some extra variables which remain hidden or at least inaccessible. Here we intend to pursue this fully entropic interpretation of the drift while taking advantage of later insights that exploit ideas from information geometry.

An important feature of ED is a central concern with the nature of time. In ED “entropic” time is designed to keep track of change. The construction of entropic time involves several ingredients. One must define the notion of ‘instants’; show that these instants happen to be ordered; and finally one must define a convenient measure of the duration or interval between the successive instants. As one might have expected, in an entropic approach to time it turns out that an arrow of time is generated automatically. Entropic time is intrinsically directional.

Here we will focus on the derivation of the Schrödinger equation but the ED approach has been applied to a variety of other topics in quantum mechanics that will not be reviewed here. These include the quantum measurement problem [32][33]; momentum and uncertainty relations [34]; the Bohmian limit [35][36] and the classical limit [37]; the extensions to curved spaces [38] and to relativistic fields [39][40]; and the derivation of symplectic and complex structures from information geometry [41].

2 The statistical model

As in any inference problem we must first decide on the subject matter — what are we talking about? We consider $N$ particles living in a flat Euclidean space $\mathbf{X}$ with metric $\delta_{ab}$. Here is our first assumption: The particles have definite positions $x_{n}^{a}$, collectively denoted by $x$, and it is their unknown values that we wish to infer. (The index $n$ $=1\ldots N$ labels the particles and $a=1,2,3$ its spatial coordinates.) For $N$ particles the configuration space is $\mathbf{X}_{N}=\mathbf{X}\times\ldots\times\mathbf{X}$. The fact that positions are unknown implies that the probability distribution $\rho(x)$ will be the focus of our attention.

Incidentally, this already addresses that old question of determinism vs. indeterminism that has troubled quantum mechanics from the outset. In the standard view quantum theory is considered to be an extension of classical mechanics and therefore deviations from causality demand an explanation. In contrast, according to the entropic view, quantum mechanics is an example of entropic inference, a framework designed to handle insufficient information. From this entropic perspective indeterminism requires no explanation; uncertainty and probabilities are the expected norm. It is certainty and determinism that demand explanations [37].

The assumption that the particles have definite positions that happen to be unknown is a not an innocent statement. It represents a major shift away from the standard Copenhagen interpretation. Such a departure is not in itself a problem: our goal here is not to justify any particular interpretation of quantum mechanics but to reproduce its empirical success. According to the Copenhagen interpretation observable quantities such as positions, do not in general have definite values; they can become definite but only as the immediate result of a measurement. ED differs in that positions play a very special role: at all times particles have definite positions which define the ontic state of the system. The wave function, on the other hand, is a purely epistemic notion; it defines our epistemic state about the system. In ED — just as in the Copenhagen interpretation — other observables such as energy or momentum do not in general have definite values; their values are created by the act of measurement. These other quantities are epistemic in that they do not reflect properties of the particles but of the wave function. Thus positions are ontic while energies and momenta are epistemic. This point deserves to be emphasized: in the ED description of the double slit experiment, we might not know which slit the particle goes through, but it definitely goes through one or the other.

The second assumption is that *in addition to the particles there also exist some other variables denoted *$y$. The assumption does not appear extreme.111As we shall later discuss in more detail in the concluding remarks, the $y$ variables, although hidden from our view, do not at all play the technical role usually ascribed to “hidden variables”. First, the world definitely contains more stuff than the $N$ particles we happen to be currently interested in and, second, it turns out that our main conclusions are very robust in that they turn out to be largely independent of most particular details about these $y$ variables. We only need to assume that the $y$ variables are themselves uncertain and that their uncertainty, described by some probability distribution $p(y|x)$, depends on the location $x$ of the particles. As we shall soon see it is the entropy of the distribution $p(y|x)$ that plays a significant role in defining the dynamics of $x$. Other details of the distribution $p(y|x)$ turn out to be irrelevant. For later reference, the entropy $S(x)$ of $p(y|x)$ relative to an underlying measure $q(y)$ on the space of $y$ variables is

[TABLE]

For notational simplicity in multidimensional integrals such as (1) we will write $dy$ instead of $d^{n}y$. Note also that $S(x)$ is a scalar function on the configuration space $\mathbf{X}_{N}$.

Once the microstates $(x,y)\in\mathbf{X}_{N}\times\mathbf{Y}$ are identified we proceed to tackle the dynamics. Our goal is to find the probability density $P(x^{\prime}|x)$ for a change from an initial position $x\in\mathbf{X}_{N}$ to a new neighboring $x^{\prime}\in\mathbf{X}_{N}$. Since both $x^{\prime}$ and the associated $y^{\prime}$ are unknown the relevant space is not just $\mathbf{X}_{N}$ but $\mathbf{X}_{N}\times\mathbf{Y}$ and, therefore, the distribution we seek is the joint distribution $P(x^{\prime},y^{\prime}|x,y)$. This is found by maximizing the appropriate entropy,

[TABLE]

relative to the joint prior $Q(x^{\prime},y^{\prime}|x,y)$ and subject to the appropriate constraints.

Choosing the prior

In eq.(2) the prior $Q(x^{\prime},y^{\prime}|x,y)$ expresses our beliefs — or more precisely, the beliefs of an ideally rational agent — before any information about the specific motion is taken into account. We adopt a prior that represents a state of considerable ignorance: knowledge of $x$s tells us nothing about $y$s and vice versa. This is represented by a product,

[TABLE]

The prior $Q(y^{\prime}|y)$ for the $y$ variables is chosen to be a uniform distribution. If the volume element in the space $\mathbf{Y}$ is proportional to $q(y)dy$ then $Q(y^{\prime}|y)\propto q(y^{\prime})$. The measure $q(y)$ need not be specified further. The prior $Q(x^{\prime}|x)$ for the $x$ variables is considerably more informative.

In a “mechanics without a mechanism” one does not explain why motion happens. The goal instead is to produce an estimate of what kind of motion one might reasonably expect. The central piece of dynamical information is that the particles follow trajectories that are continuous. This represents yet another significant deviation from the early historical development of quantum mechanics which stressed discreteness and discontinuity. Fortunately, in modern versions of quantum mechanics these two aspects are de-emphasized. Ever since Schrödinger’s pioneering work the discreteness of quantum numbers is not mysterious — certainly not more mysterious than the discrete resonances of classical vibrations. And the discontinuities implicit in abrupt quantum jumps, a relic from Bohr’s old quantum theory, just do not exist; they were eventually recognized as unnecessary and effectively discarded as soon as Schrödinger formulated his equation. Incidentally, there is another source of discontinuity — the so-called wave function collapse. Its mystery disappears the moment one recognizes the epistemic nature of the wave function: the collapse is not physical but merely an instance of probability updating [32][33].

The assumption of continuity represents a tremendous simplification because it implies that motion can be analyzed as the accumulation of many infinitesimally short steps. We adopt a prior $Q(x^{\prime}|x)$ that incorporates the information that the particles take steps that are infinitesimally short, that reflects translational and rotational invariance, but is otherwise maximally ignorant about any correlations. Such a prior can itself be derived from the principle of maximum entropy. Indeed, maximize

[TABLE]

where $\Delta x_{n}^{a}=x_{n}^{\prime a}-x_{n}^{a}$, relative to the uniform measure $\mu=\,$const., subject to normalization, and $N$ independent constraints that enforce short steps and rotational invariance,

[TABLE]

where $\kappa_{n}$ are small constants. The result is a product of Gaussians,

[TABLE]

The Lagrange multipliers $\alpha_{n}$ are constants that may depend on the index $n$ in order to describe non-identical particles. They will eventually be taken to infinity in order to enforce infinitesimally short steps.

The choice of a Gaussian prior turns out to be natural, not just because it arises in an informational context, but also because, as described by the central limit theorem as in the theory of errors, it arises whenever a “macroscopic” effect is the result of the accumulation of a large number of “microscopic” contributions.

Since proportionality constants have no effect on the entropy maximization, our choice for the joint prior is

[TABLE]

Now we are ready to specify the constraints.

The constraints

We first write the posterior as a product,

[TABLE]

and we specify the two factors separately. We require that the new $x^{\prime}$ depends only on $x$, so we set $P(x^{\prime}|x,y)=P(x^{\prime}|x)$ which is the transition probability we wish to find. The new $x^{\prime}$ is independent of the actual values of $y$ or $y^{\prime}$ but, as we shall soon see, it does depend on their entropies.

As for the second factor in (8), we require that $P(y^{\prime}|x^{\prime},x,y)=p(y^{\prime}|x^{\prime})$ — the uncertainty in $y^{\prime}$ depends only on $x^{\prime}$. Therefore, the first constraint is that the joint posterior be of the form

[TABLE]

To implement this constraint substitute (9) into the joint entropy (2) to get

[TABLE]

where $S(x)$ is given in eq.(1).

Having specified the prior and the constraints the ME method takes over. We vary $P(x^{\prime}|x)$ to maximize (10) subject to normalization. The result is

[TABLE]

where $\zeta$ is a normalization constant. In eq.(11) it is clear that infinitesimally short steps are obtained in the limit $\alpha_{n}\rightarrow\infty$. It is therefore useful to Taylor expand,

[TABLE]

and rewrite $P(x^{\prime}|x)$ as

[TABLE]

where $Z$ is the new normalization constant. A generic displacement $\Delta x_{n}^{a}=x_{n}^{\prime a}-x_{n}^{a}$ can be expressed as an expected drift plus a fluctuation,

[TABLE]

[TABLE]

These equations show that as $\alpha_{n}\rightarrow\infty$, for very short steps, the dynamics is dominated by fluctuations $\Delta w_{n}^{a}$ which are of order $O(\alpha_{n}^{-1/2})$, while the drift $\langle\Delta x_{n}^{a}\rangle$ is much smaller, only of order $O(\alpha_{n}^{-1})$. Thus, just as in Brownian motion, the trajectory is continuous but not differentiable. In ED particles have definite positions but their velocities are completely undefined. Notice also that the particle fluctuations are isotropic and independent of each other. The directionality of the motion and correlations among the particles are introduced by a systematic drift along the gradient of the entropy of the $y$ variables.

The introduction of the auxiliary $y$ variables, which has played a central role of conveying relevant information about the drift, deserves a comment. It is important to realize that the same information can be conveyed through other means, which demonstrates that quantum mechanics, as an effective theory, turns out to be fairly robust222In the case of physics, the $y$ variables might for example describe the microscopic structure of spacetime. In the case of an economy, they may describe the hidden neural processes of various agents. It is a strength of the ED formulation that details about these auxilliary variables are not needed. [35][36]. It is possible, for example, to avoid the $y$ variables altogether and invoke a drift potential directly (see e.g., [15]-[17]). The advantage of an explicit reference to some vaguely defined $y$ variables is their mere existence may be sufficient to account for drift effects.

3 Entropic time

Having obtained a prediction, given by eq.(13), for what motion to expect in one infinitesimally short step we now consider motion over finite distances. This requires the introduction of a parameter $t$, to be called time, as a book-keeping tool to keep track of the accumulation of successive short steps. Since the rules of inference are silent on the subject of time we need to justify why the parameter $t$ deserves to be called time.

The construction of time involves three ingredients. First, we must identify something that one might call an “instant”; second, it must be shown that these instants are ordered; and finally, one must introduce a measure of separation between these successive instants — one must define “duration.” Since the foundation for any theory of time is the theory of change — that is, dynamics — the notion of time constructed below will reflect the inferential nature of entropic dynamics. Such a construction we will call entropic time.

3.1 Time as an ordered sequence of instants

Entropic dynamics is given by a succession of short steps described by $P(x^{\prime}|x)$, eq.(13). Consider, for example, the $i$th step which takes the system from $x=x_{i-1}$ to $x^{\prime}=x_{i}$. Integrating the joint probability, $P(x_{i},x_{i-1})$, over $x_{i-1}$ gives

[TABLE]

This equation follows directly from the laws of probability, it involves no assumptions and, therefore, it is sort of empty. To make it useful something else must be added. Suppose we interpret $P(x_{i-1})$ as the probability of different values of $x_{i-1}$ *at one “instant” labelled *$t$, then we can interpret $P(x_{i})$ as the probability of values of $x_{i}$ at the next “instant” which we will label $t^{\prime}$. Writing $P(x_{i-1})=\rho(x,t)$ and $P(x_{i})=\rho(x^{\prime},t^{\prime})$ we have

[TABLE]

Nothing in the laws of probability leading to eq.(16) forces the interpretation (17) on us — this is the additional ingredient that allows us to construct time and dynamics in our model. Equation (17) defines the notion of “instant”: If the distribution $\rho(x,t)$ refers to one instant $t$, then the distribution $\rho(x^{\prime},t^{\prime})$ generated by $P(x^{\prime}|x)$ defines what we mean by the “next” instant $t^{\prime}$. Iterating this process defines the dynamics. Entropic time is constructed instant by instant: $\rho(t^{\prime})$ is constructed from $\rho(t)$, $\rho(t^{\prime\prime})$ is constructed from $\rho(t^{\prime})$, and so on.

The construction is intimately related to information and inference. An ‘instant’ is an informational state that is complete in the sense that it is specified by the information — codified into the distributions $\rho(x,t)$ and $P(x^{\prime}|x)$ — that is sufficient for predicting the next instant. To put it briefly: the instant we call the present is such that the future, given the present, is independent of the past.

In the ED framework the notions of instant and of simultaneity are intimately related to the distribution $\rho(x)$. It is instructive to discuss this further. When we consider a single particle at a position $\vec{x}=(x^{1},x^{2},x^{3})$ in $3$-d space it is implicit in the notation that the three coordinates $x^{1}$, $x^{2}$, and $x^{3}$ occur simultaneously — no surprises here. Things get a bit more interesting when we describe a system of $N$ particles by a single point $x=(\vec{x}_{1},\vec{x}_{2},\ldots\vec{x}_{N})$ in $3N$-dimensional configuration space. Here it is also implicitly assumed that all the $3N$ coordinate values refer to the same instant; we take them to be simultaneous. But this is no longer so trivial because the particles are located at different places — we do not mean particle 1 is at $\vec{x}_{1}$ now, and particle 2 is at $\vec{x}_{2}$ later. There is an implicit assumption linking the very idea of a configuration space with that of simultaneity and something similar occurs when we introduce probabilities. Whether we talk about one particle or about $N$ particles, a distribution such as $\rho(x)$ describes our uncertainty about the possible configurations $x$ of the system at a given instant. The different values of $x$ refer to the same instant; they are meant to be simultaneous. And therefore, in ED, a probability distribution $\rho(x)$ provides a criterion of simultaneity.

In a relativistic theory there is a greater freedom in the choice of instants and this translates into a greater flexibility with the notion of simultaneity. Conversely, as we have shown elsewhere, the requirement that these different notions of simultaneity be consistent with each other places strict constraints on the allowed forms of ED [40].

It is common to use equations such as (17) to define a special kind of dynamics, called Markovian, that unfolds in a time defined by some external clocks. In such a Markovian dynamics the specification of the state at one instant is sufficient to determine its evolution into the future.333The application of entropic dynamics to continuous motion is, of course, heavily motivated by the application to physics; for all we know space and time form a continuum. The derivation of discrete Markov models as a form of entropic dynamics is a subject for future research. But even in the case of an essentially discrete dynamics (such as e.g. the daily updates of the stock market) the idealization of continuous evolution is still useful as an approximation over somewhat longer time scales. It is important to recognize that in ED we are not making a Markovian assumption. Although eq.(17) is formally identical to the Chapman-Kolmogorov equation it is used for a very different purpose. We do not use (17) to define a (Markovian) dynamics in a pre-existing background time because in ED there are no external clocks. The system is its own clock and (17) is used both to define the dynamics and to construct time itself.

3.2 The arrow of entropic time

The notion of time constructed according to eq.(17) is intrinsically directional. There is an absolute sense in which $\rho(x,t)$ is prior and $\rho(x^{\prime},t^{\prime})$ is posterior. If we wanted to construct a time-reversed evolution we would write

[TABLE]

where according to the rules of probability theory $P(x|x^{\prime})$ is related to $P(x^{\prime}|x)$ in eq.(13) by Bayes’ theorem,

[TABLE]

This is not, however, a mere exchange of primed and unprimed quantities. The distribution $P(x^{\prime}|x)$, eq.(13), is a Gaussian derived from the maximum entropy method. In contrast, the time-reversed $P(x|x^{\prime})$ is given by Bayes’ theorem, eq.(19), and is not in general Gaussian. The asymmetry between the inferential past and the inferential future is traced to the asymmetry between priors and posteriors.

The puzzle of the arrow of time (see e.g. [42][43]) has been how to explain the asymmetric arrow from underlying symmetric laws. The solution offered by ED is that there are no underlying laws whether symmetric or not. The time asymmetry is the inevitable consequence of entropic inference. From the point of view of ED the challenge is not to explain the arrow of time but the reverse: how to explain the emergence of symmetric laws within an entropic framework that is intrinsically asymmetric. As we shall see below some laws of physics derived from ED, such as the Schrödinger equation, are indeed time-reversible even though entropic time itself only flows forward.

3.3 Duration: a convenient time scale

To complete the construction of entropic time we need to specify the interval $\Delta t$ between successive instants. The basic criterion is convenience: duration is defined so that motion looks simple. We saw in eqs.(14) and (15) that for short steps (large $\alpha_{n}$) the motion is largely dominated by fluctuations. Therefore specifying $\Delta t$ amounts to specifying the multipliers $\alpha_{n}$ in terms of $\Delta t$.

The description of motion is simplest when it reflects the symmetry of translations in space and time. In a flat spacetime this leads us to an entropic time that resembles Newtonian time in that it flows “equably everywhere and everywhen.” Thus, we choose $\alpha_{n}$ to be independent of $x$ and $t$, and we choose $\Delta t$ so that $\alpha_{n}\propto 1/\Delta t$. Furthermore, it is convenient to express the proportionality constants in terms of some particle-specific constants $m_{n}$ and an overall constant $\eta$ that fixes the units of the $m_{n}$s relative to the units of time. The result is

[TABLE]

The constants $m_{n}$ will eventually be identified with the particle masses and the constant $\eta$ will be regraduated into $\hbar$.

4 The information metric of configuration space

Before we proceed to study the dynamics defined eq.(13) with its corresponding notion of entropic time it is useful to consider the geometry of the $N$-particle configuration space, $\mathbf{X}_{N}$. We have assumed that the geometry of the single particle spaces $\mathbf{X}$ is described by the Euclidean metric $\delta_{ab}$. We can expect that the $N$-particle configuration space, $\mathbf{X}_{N}=\mathbf{X}\times\ldots\times\mathbf{X}$, will also be flat, but a question remains about the relative scales associated to each $\mathbf{X}$ factor. Information geometry provides the answer.

To each point $x\in\mathbf{X}_{N}$ there corresponds a probability distribution $P(x^{\prime}|x)$. This means that $\mathbf{X}_{N}$ is a statistical manifold the geometry of which is uniquely determined (up to an overall scale factor) by the information metric,

[TABLE]

Here the upper case indices label both the particle and its coordinate, $x^{A}=x_{n}^{a}$, and $C$ is an arbitrary positive constant (see e.g., [21][44]). Substituting eqs.(13) and (20) into (21) in the limit of short steps ($\alpha_{n}\rightarrow\infty$) yields

[TABLE]

Note the divergence as $\Delta t\rightarrow 0$. Indeed, as $\Delta t\rightarrow 0$ the distributions $P(x^{\prime}|x)$ and $P(x^{\prime}|x+\Delta x)$ become more sharply peaked and they become increasingly easier to distinguish. This leads to an increasing information distance, $\gamma_{AB}\rightarrow\infty$. To define a distance that remains useful for arbitrarily small $\Delta t$ we choose $C\propto\Delta t$. In fact, since $\gamma_{AB}$ will always appear in the combination $\gamma_{AB}\Delta t/C$, it is best to define the “mass” tensor,

[TABLE]

and its inverse,

[TABLE]

Thus, up to overall constants the metric of configuration space is the mass tensor. This result may be surprising. Ever since the work of Heinrich Hertz in 1894 [45] it has been standard practice to describe the motion of systems with many particles as the motion of a single point in an abstract space — the configuration space. Choosing the geometry of this configuration space had so far been regarded as a matter of convenience and was suggested through an examination of the kinetic energy of the system. In contrast, in ED there is no choice. Up to a global scale the metric follows uniquely determined by information geometry.

To recap our results so far: with the multipliers $\alpha_{n}$ chosen according to (20), the dynamics given by $P(x^{\prime}|x)$ in (13) is a standard Wiener process. A generic displacement, eq.(14), is

[TABLE]

where $b^{A}(x)$ is the drift velocity,

[TABLE]

and the uncertainty $\Delta w^{A}$ is given by

[TABLE]

I finish this section with two remarks. The first is on the nature of clocks: In Newtonian mechanics time is defined to simplify the motion of free particles. The prototype of a clock is a free particle: it moves equal distances in equal times. In ED time is also defined to simplify the dynamics of free particles. The prototype of a clock is a free particle too: as we see in (27) free particles (because for sufficiently short times all particles are free) undergo equal fluctuations in equal times.

The second remark is on the nature of mass. As we shall soon see the constants $m_{n}$ will be identified with the particles’ masses. Then, eq.(27) provides an interpretation of what ‘mass’ is: mass is an inverse measure of fluctuations. It is not unusual to treat the concept of mass as an unexplained primitive concept that measures the amount of stuff (or perhaps the amount of energy) and then state that quantum fluctuations are inversely proportional to mass. In an inference scheme such as ED the presence of fluctuations do not require an explanation. They reflect uncertainty, the natural consequence of incomplete information. This opens the door to an entropic explanation of mass: mass is just a measure of uncertainty about the expected motion.

5 Diffusive dynamics

The dynamics of $\rho(x,t)$, given by the integral equation (17), is more conveniently re-written in a differential form known as the Fokker-Planck (FP) equation [46],

[TABLE]

(For the algebraic details see e.g., [21].) The FP equation can also be written as a continuity equation,

[TABLE]

The product $\rho v^{A}$ in (29) represents the probability current, and $v^{A}$ is interpreted as the velocity of the probability flow — it is called the current velocity. From (28) the current velocity in (29) is the sum of two separate contributions,

[TABLE]

The first term is the drift velocity $b^{A}\propto\partial^{A}S$ in (26) and describes the tendency of the distribution $\rho(x)$ to evolve so as to increase the entropy of the $y$ variables. One could adopt a language that resembles a causal mechanism: *it is as if *the system were pushed by an entropic force. But, of course, such a mechanistic language should not be taken literally. Strictly speaking, entropic forces cannot be causal agents because they are purely epistemic. They might influence our beliefs and expectations but they are not physically capable of pushing the system — the “real causes,” if any, lie elsewhere. Or perhaps, there are no real causes. In classical mechanics, for example, there is nothing out there that causes a free particle to persist in its uniform motion in a straight line. Similarly, in ED we are not required to identify what makes changes happen the way they do.

The second term in (30) is the osmotic velocity,

[TABLE]

It represents diffusion, the tendency for probability to flow down the density gradient. Indeed, one might note that the osmotic or diffusive component of the current, $\rho u^{A}$, obeys a version of Fick’s law in configuration space,

[TABLE]

where $D^{AB}=\eta m^{AB}/2$ is the diffusion tensor [46]. Here too one might be tempted to adopt a causal mechanistic language and say that the diffusion is driven by the fluctuations expressed in eq.(27). *It is as if *the system were bombarded by some underlying random field. But eq.(27) need not represent actual fluctuations; it represents our uncertainty about where the particle will be found after $\Delta t$. The mere fact that we happen to be uncertain about the position of the particles does not imply that something must be shaking them.444E. T. Jaynes warned us that “the fact that our information is able to determine [a quantity] $F$ to 5 percent accuracy, is not enough to make it fluctuate by 5 percent!” This mistake he called the Mind Projection Fallacy [47].

Next we note that since both $b^{A}$ and $u^{A}$ are gradients, their sum — the current velocity — is a gradient too,

[TABLE]

The FP equation,

[TABLE]

can be conveniently rewritten in yet another equivalent but very suggestive form involving functional derivatives. For some suitably chosen functional $\tilde{H}[\rho,\Phi]$ we have

[TABLE]

It is easy to check that the appropriate functional $\tilde{H}$ is

[TABLE]

where the integration constant $F[\rho]$ is some unspecified functional of $\rho$.

We have just exhibited our first example of a mechanics without mechanism: a standard diffusion derived from principles of entropic inference. Next we turn our attention to the derivation of quantum mechanics.

6 Hamiltonian dynamics

The previous discussion has led us to a standard diffusion. It involves a single dynamical field, the probability density $\rho(x)$, that evolves in response to a fixed non-dynamical “potential” given by the entropy of the $y$ variables, $S(x)$. In contrast, a quantum dynamics consists in the coupled evolution of two dynamical fields: the density $\rho(x)$ and the phase of the wave function. This second field can be naturally introduced into ED by allowing the entropy $S(x)$ to become dynamical: the entropy $S$ guides the evolution of $\rho$, and in return, the evolving $\rho$ reacts back and induces a change in $S$. This amounts to an entropic dynamics in which each short step is constrained by a slightly different drift potential; the constraint (9) is continuously updated at each instant in time.

Clearly, different updating rules lead to different forms of ED. The rule that turns out be particularly useful in physics is inspired by an idea of Nelson’s [48]. We require that $S$ be updated in such a way that a certain functional, later to be called “energy,” remains constant. Such a rule may appear to be natural — how could it be otherwise? Could we possibly imagine physics without a conserved energy? But this naturalness is deceptive because ED is not at all like a classical mechanics. Indeed, the classical interpretation of a Langevin equation such as (25) is that of a Brownian motion in the limit of infinite friction. This means that in order to provide a classical explanation of quantum behavior we would need to assume that the particles were subjected to infinite friction while undergoing zero dissipation. Such a strange dynamics could hardly be called ‘classical’; the relevant information that is captured by our choice of constraints cannot be modelled by invoking some underlying classical mechanism — this is mechanics without a mechanism. Furthermore, while it is true that an updating rule based on the notion of a conserved total energy happens to capture the relevant constraints for a wide variety of physics problems, its applicability is limited even within physics. For example, in the curved spacetimes that are used to model gravity it is not possible to even define a global energy, much less require its global conservation. (For the updating rules that apply to curved spaces see [40].)

The ensemble Hamiltonian –

In the standard approach to mechanics one starts with an action and from its invariance under time translations one derives the conservation of energy. Our derivation proceeds in the opposite direction: we first identify energy conservation as the relevant piece of information and from it we derive Hamilton’s equations and their associated action principle.

It turns out that the empirically successful energy functionals are of the form (36). We impose that, irrespective of the initial conditions, the entropy $S$ or, equivalently, the potential $\Phi$ in (33), will be updated in such a way that the functional $\tilde{H}[\rho,\Phi]$ in (36) is always conserved,

[TABLE]

or,

[TABLE]

Using eq.(35) we get

[TABLE]

We want $d\tilde{H}/dt=0$ to hold for arbitrary choices of the initial values of $\rho$ and $\Phi$. Using eq.(34) this translates into requiring $d\tilde{H}/dt=0$ for arbitrary choices of $\partial_{t}\rho$. Therefore, $\Phi$ must be updated according to

[TABLE]

Equations (35) and (40) are recognized as a conjugate pair of Hamilton’s equations and the conserved functional $\tilde{H}[\rho,\Phi]$ in (36) is then called the ensemble Hamiltonian.

Thus, the form of the ensemble Hamiltonian $\tilde{H}$ is chosen so that the first Hamilton equation (35) is the FP eq.(29), and then the second Hamilton equation (40) becomes a generalized Hamilton-Jacobi equation,

[TABLE]

This is our second example of a mechanics without mechanism: a non-dissipative ED leads to Hamiltonian dynamics.

The action –

We have just seen that the field $\rho(x)$ is a generalized coordinate and $\Phi(x)$ is its canonical momentum. Now that we have Hamilton’s equations, (35) and (40), it is straightforward to invert the usual procedure and construct an action principle from which they can be derived. Just define the differential

[TABLE]

and then integrate to get an “action”,

[TABLE]

Thus, by construction, imposing $\delta A=0$ leads to (35) and (40). Thus, in the ED approach, actions are not particularly fundamental; they are just clever ways to summarize the dynamics in a very condensed form.

7 Information geometry and the Quantum Potential

Different choices of the functional $F[\rho]$ in (36) lead to different dynamics. Earlier we used information geometry, eq.(21), to define the metric $m_{AB}$ of configuration space. Here we use information geometry once again to motivate the particular choice of the functional $F[\rho]$ that leads to quantum theory.

The special role played by the particle positions leads us to consider the family of distributions $\rho(x|\theta)$ that are generated from a distribution $\rho(x)$ by translations in configuration space by a vector $\theta^{A}$, $\rho(x|\theta)=\rho(x-\theta)$. The extent to which $\rho(x|\theta)$ can be distinguished from the slightly displaced $\rho(x|\theta+d\theta)$ or, equivalently, the information distance between $\theta^{A}$ and $\theta^{A}+d\theta^{A}$, is given by

[TABLE]

where

[TABLE]

Changing variables $x-\theta\rightarrow x$ yields

[TABLE]

Note that these are translations in configuration space. They are not translations in which the system is displaced as a whole in $3$-d space by the same constant amount, $(\vec{x}_{1},\vec{x}_{2},\ldots\vec{x}_{N})\rightarrow(\vec{x}_{1}+\vec{\varepsilon},\vec{x}_{2}+\vec{\varepsilon},\ldots\vec{x}_{N}+\vec{\varepsilon})$. The metric (46) measures the extent to which a distribution $\rho(x)$ can be distinguished from another distribution $\rho^{\prime}(x)$ in which just one particle has been slightly shifted while all others remain untouched, e.g., $(\vec{x}_{1},\vec{x}_{2},\ldots\vec{x}_{N})\rightarrow(\vec{x}_{1}+\vec{\varepsilon},\vec{x}_{2},\ldots\vec{x}_{N})$.

The simplest choice of functional $F[\rho]$ is linear in $\rho$, $F[\rho]=\int dx\,\rho(x)V(x)$, and the function $V(x)$ will play the role of the familiar scalar potential. In an entropic dynamics one might also expect contributions that are of a purely informational nature. Information geometry provides us with two tensors: one is the metric of configuration space $\gamma_{AB}\propto m_{AB}$, and the other is $I_{AB}[\rho]$. The simplest nontrivial scalar that can be constructed from them is the trace $m^{AB}I_{AB}$. This suggests

[TABLE]

where $\xi>0$ is a constant that controls the relative strength of the two contributions. The case $\xi<0$ leads to instabilities and is therefore excluded. (From eq.(46) we see that $m^{AB}I_{AB}$ is a contribution to the energy such that those states that are more smoothly spread out tend to have lower energy.) The case $\xi=0$ leads to a qualitatively different theory — a hybrid dynamics that is both indeterministic and yet classical [35]. The term $m^{AB}I_{AB}$ is usually called the “quantum” potential or the “osmotic” potential.555To my knowledge the relation between the quantum potential and the Fisher information was first pointed out in [49]. It is the crucial term that accounts for all quintessentially ‘quantum’ effects — superposition, entanglement, wave packet expansion, tunnelling, and so on.

With the choice (47) for $F[\rho]$ the generalized Hamilton-Jacobi equation (41) becomes

[TABLE]

8 The Schrödinger equation

Once we have the coupled equations (34) and (48) we are done — it may not yet be obvious yet but this is quantum mechanics. Purely for the sake of convenience it is useful to combine $\rho$ and $\Phi$ into a single complex function

[TABLE]

where $k$ is an arbitrary positive constant which amounts to rescaling the constant $\eta$ in eq.(33). Then the two equations (34) and (48) can be written into a single complex equation,

[TABLE]

which is quite simple and elegant except for the last non-linear term. It is at this point that we can take advantage of our freedom in the choice of $k$. Since the dynamics is fully specified through $\rho$ and $\Phi$ the different choices of $k$ in $\Psi_{k}$ all lead to different versions of the same theory. Among all these equivalent descriptions it is clearly to our advantage to pick the $k$ that is most convenient — a process sometimes known as ‘regraduation’.666Other notable examples of regraduation include the Kelvin choice of absolute temperature and the Cox derivation of the sum and product rule for probabilities [21]. The optimal choice, $k_{\text{opt}}=(8\xi)^{1/2}$, is such that the non-linear term drops out and is identified with Planck’s constant,

[TABLE]

Then eq.(50) becomes the Schrödinger equation,

[TABLE]

where the wave function is

[TABLE]

The constant \xi=\hbar^{2}/8\in eq.(47) turned out to play a crucial role: it defines the numerical value of what we call Planck’s constant and sets the scale that separates quantum from classical regimes.

The conclusion is that for any positive value of the constant $\xi$ it is always possible to combine $\rho$ and $\Phi$ to a physically equivalent but more convenient description where the Schrödinger equation is linear. From the ED perspective the linear superposition principle together with its attendant complex linear Hilbert spaces are definitely important. But they are important because they are convenient calculational tools and not because they are fundamental.

9 Some final comments

A theory such as ED can lead to many questions. Here are a few.

Is ED equivalent to quantum mechanics?

Are the Fokker-Planck eq.(34) and the generalized Hamilton-Jacobi eq.(48) fully equivalent to the Schrödinger equation? This question, first raised by Wallstrom [50] in the context of Nelson’s stochastic mechanics [7], concerns the single- or multi-valuedness of phases and wave functions. Briefly Wallstrom’s objection is that Nelson’s stochastic mechanics led to phases $\Phi$ and wave functions $\Psi$ that are either both multi-valued or both single-valued. Both alternatives are unsatisfactory: quantum mechanics forbids multi-valued wave functions, while single-valued phases can exclude physically relevant states (e.g., states with non-zero angular momentum). We will not discuss this issue except to note that the objection does not apply once particle spin is incorporated into ED [51][52]. A similar argument was developed by Takabayasi in the very different context of his hydrodynamical approach to quantum theory [53].

Is ED a hidden-variable model?

Let us return to these mysterious auxiliary $y$ variables. Should we think of them as hidden variables? There is a trivial sense in which the $y$ variables are “hidden”: they are not directly observable.777The $y$ variables are not observable at the current stage of development of the theory. It may very well happen that once we learn where to look we will find that they have been staring us in the face all along. But being unobserved is not sufficient to qualify as a hidden variable. The original motivation behind attempts to construct hidden variable models was to explain or at least ameliorate certain aspects of quantum mechanics that clash with our classical preconceptions. But the $y$ variables address none of these problems. Let us mention a few of them:

(1) Indeterminism: Is ultimate reality random? Do the gods play dice? In the standard view quantum theory is considered an extension of classical mechanics — indeed, the subject is called quantum mechanics — and therefore deviations from causality demand an explanation. In the entropic view, on the other hand, *quantum theory is not mechanics; it is inference — *entropic inference is a framework designed to handle insufficient information. From the entropic perspective indeterminism requires no explanation. Uncertainty and probabilities are the norm; it is the certainty and determinism of the classical limit that demand an explanation [37].

(2) Non-classical mechanics: A common motivation for hidden variables is that a sub-quantum world will eventually be discovered where nature obeys essentially classical laws. But in ED there is no underlying classical dynamics — both quantum and its classical limit are derived. The peculiar non-classical effects associated with the wave-particle duality arise not so much from the $y$ variables themselves but rather from the specific non-dissipative diffusion which leads to a Schrödinger equation. The important breakthrough here was Nelson’s realization that diffusion phenomena could be much richer than previously expected — non-dissipative diffusions can account for wave and interference effects.

(3) Non-classical probabilities: It is often argued that classical probability fails to describe the double slit experiment; this is not true [21]. It is the whole entropic framework — and not the $y$ variables — that is incompatible with the notion of quantum probabilities. From the entropic perspective it makes just as little sense to distinguish quantum from classical probabilities as it is would be to talk about economic or medical probabilities.

(4) Non-locality: Realistic interpretations of the wave function often lead to such paradoxes as the wave function collapse and the non-local Einstein-Podolski-Rosen (EPR) correlations [4][5]. Since in the ED approach the particles have definite positions and we have introduced auxiliary $y$ variables that might resemble hidden variables it is inevitable that one should ask whether this theory violates Bell inequalities? Or, to phrase the question differently: where precisely is non-locality introduced? The answer is that the theory has been formulated directly in $3N$-dimensional configuration space and the Hamiltonian has been chosen to include the highly non-local quantum potential (47). So, yes, the ED model developed here properly describes the highly non-local effects that lead to EPR correlations and to violations of Bell inequalities.

On interpretation

We have derived quantum theory as an example of entropic inference. The problem of interpretation of quantum mechanics is solved because instead of starting with the mathematical formalism and then seeking an interpretation that can be consistently attached to it, one starts with a unique interpretation and then one builds the formalism. This allows a clear separation between the ontic and the epistemic elements. In ED there is no risk of confusing which is which. “Reality” is represented through the positions of the particles, and our “limited information about reality” is represented by probabilities as they are updated to reflect the physically relevant constraints. In ED all other quantities, including the wave function, are purely epistemic tools. Even energy and momentum and all other so-called “observables” are epistemic; they are not properties of the particles but of the wave functions. [32][33]

To reiterate a point we made above: since “quantum” probabilities were never mentioned one might think that entropic dynamics is a classical theory. But this is misleading: in ED probabilities are neither classical nor quantum; they are tools for inference. All those non-classical phenomena, such as the non-local effects that arise in double-slit interference experiments, or the entanglement that leads to non-local Einstein-Podolski-Rosen correlations, are the natural result of the linearity that follows from including the quantum potential term in the ensemble Hamiltonian.

In ED neither action principles nor Hilbert spaces are fundamental. They are convenient tools designed to summarize the dynamical laws derived from the deeper principles of entropic inference. The requirement that an energy be conserved is an important piece of information (that is, a constraint) which will be fully justified once the extension of entropic dynamics to gravity is developed.

The derivation of laws of physics as examples of inference has led us to discuss the concept of time. The notion of entropic time was introduced to keep track of the accumulation of changes. It includes assumptions about the concept of instant, of simultaneity, of ordering, and of duration. A question that is bound to be raised is whether and how entropic time is related to the actual, real, “physical” time. In a similar vein, to quantify the uncertainties in the motion — the fluctuations — to each particle we associated one constant $m_{n}$. We are naturally led to ask: How are these constants related to the masses of the particles?

The answers are provided by the dynamics itself: by deriving the Schrödinger equation from which we can obtain its classical limit (Newton’s equation, $F=ma$ [37]) we have shown that the $t$ that appears in the laws of physics is entropic time and the constants $m_{n}$ are masses. The argument is very simple: it is the Schrödinger equation and its classical limit that are used to design and calibrate our clocks and our mass-measuring devices. We conclude that by their very design, the time measured by clocks is entropic time, *and what mass measurements yield are the constants *$m_{n}$. No notion of time that is in any way more “real” or more “physical” is needed. Most interestingly, even though the dynamics is time-reversal invariant, entropic time is not. The model automatically includes an arrow of time.

Finally, here we have focused on the derivation of examples of dynamics that are relevant to physics, but the fact that ED is based on inference methods that are of universal applicability and, in particular, the fact that in the entropic dynamics framework one deliberately abstains from framing hypothesis about underlying mechanisms, suggests that it may be possible to adapt these methods to fields other than physics [54].

Acknowledgments

My views on this subject have benefited from discussions with many students and collaborators including M. Abedi, D. Bartolomeo, C. Cafaro, N. Caticha, A. Demme, S. DiFranzo, A. Giffin, S. Ipek, D.T. Johnson, K. Knuth, S. Nawaz, P. Pessoa, M. Reginatto, C. Rodriguez, and K. Vanslette.