Diffusing diffusivity selects Pareto tail exponent in random growth with redistribution

This paper demonstrates that introducing diffusing diffusivity into the Bouchaud-Mézard model of random growth with redistribution leads to a stationary Pareto wealth tail whose exponent is determined by the interplay between high-diffusivity states and redistribution rates, rather than simply by the mean diffusivity.

The Gist

Imagine a bustling marketplace where people are constantly trying to grow their wealth. In this market, two main forces are at play:

1. The Rollercoaster of Chance: Your wealth grows or shrinks randomly, like a rollercoaster. Sometimes you hit a huge jackpot; sometimes you take a steep dive.
2. The Redistribution Machine: To keep things from getting out of control, there's a mechanism that constantly takes a little bit from the very rich and gives it to the very poor, trying to keep everyone in the middle.

For decades, scientists used a model (called the Bouchaud–Mézard model) to predict how wealth is distributed in this market. They assumed that the "roughness" of the rollercoaster—the volatility—was fixed. They thought the track was always equally bumpy, no matter what.

The New Discovery: The Rollercoaster Track Itself Changes

This paper argues that in the real world, the "roughness" of the track isn't fixed. It changes over time. Sometimes the track is smooth and predictable (low volatility); other times, it's a chaotic, bumpy mess (high volatility). The authors call this "diffusing diffusivity."

Think of it like this:

• The Old View: Imagine driving a car on a road where the potholes are always the same size.
• The New View: Imagine driving on a road where the potholes themselves are moving. Sometimes you are on a smooth highway; a moment later, you are on a dirt track full of rocks. The nature of the road is fluctuating.

What Happens When You Ride Alone?

If you are just one person riding this changing rollercoaster (without the redistribution machine), the paper finds something interesting about time:

• Short Term (The Immediate Ride): If you look at your journey over a short period, your path looks wild and unpredictable. It doesn't follow a standard bell curve; it's "fat-tailed," meaning extreme ups and downs are more common than usual. This is because you might get stuck on a "bumpy" section of the track for a while.
• Long Term (The Whole Trip): If you ride for a very long time, you eventually experience all types of road conditions—smooth, bumpy, and everything in between. Because you've seen it all, your average journey smooths out and starts to look like a normal, predictable bell curve again. The chaos of the changing road "averages itself out."

What Happens When the Whole Market is Connected?

The real magic happens when we bring back the Redistribution Machine (the system that moves money between people).

In the old model, scientists thought that to predict the wealth of the super-rich, you just needed to know the average roughness of the road. They thought, "If the road is bumpy 50% of the time and smooth 50% of the time, just use the average bumpiness to calculate the results."

The paper proves this is wrong.

When the road conditions change slowly (meaning you stay on a bumpy track for a long time before switching to a smooth one), the "average" doesn't matter anymore. Instead, the most extreme conditions take over.

• The Analogy: Imagine a race where runners switch between a smooth track and a muddy, bumpy track.
  • If they switch tracks instantly, the race result depends on the average speed of both tracks.
  • If they stay on the muddy track for a long time, the runners on the muddy track will go wild and far ahead of everyone else. The final result is dictated entirely by the muddy track, not the average of the two.

The Main Conclusion: Who Wins the Lottery?

The paper shows that the "Pareto tail" (the mathematical rule that describes how many super-rich people there are) is selected by the periods of highest volatility.

• Fast Switching: If the road conditions change very quickly, the system acts like the old model. The wealth distribution follows the "average" road.
• Slow Switching: If the road conditions stay the same for a long time, the people who happen to get stuck in the most volatile (bumpy) state for a long time become the super-rich outliers. Their wealth explodes because they rode the wildest rollercoaster for the longest time.

In simple terms: The paper reveals that in a world where volatility fluctuates, the richest people aren't just the ones who got "lucky" on average. They are the ones who happened to stay in the "high-volatility zone" long enough to ride the biggest waves. The system doesn't care about the average road; it cares about the worst (or best) road you happened to be on for the longest time.

This changes how we calculate the "exponent" (the number that tells us how steep the wealth gap is). It's not a simple average anymore; it's a complex balance between how fast the road changes and how rough the roughest part of the road is.

Technical Summary

Technical Summary: Diffusing Diffusivity Selects Pareto Tail Exponent in Random Growth with Redistribution

Problem Statement
The Bouchaud–Mézard (BM) model describes the emergence of stationary Pareto wealth distributions in exchange economies through a balance between random multiplicative growth and additive redistribution. A critical, implicit assumption in the standard BM framework is that the multiplicative noise intensity (diffusivity, $D$) is constant. However, empirical observations in both physical systems (e.g., single-molecule tracers) and financial markets indicate that volatility is often a stochastic, fluctuating variable with persistence times comparable to the observable process. This paper investigates how replacing the constant diffusivity with a "diffusing diffusivity" (a latent stochastic process) alters the stationary Pareto tail exponent in a system of interacting agents. Specifically, it asks whether the self-averaging properties of diffusing diffusivity observed in additive Brownian motion survive the multiplicative nonlinearity and redistributive interactions of the BM model.

Methodology
The authors employ a two-tiered analytical approach:

1. Single-Agent Dynamics (Non-Interacting):
   The wealth $z_t$ of a single agent is modeled via a stochastic differential equation (SDE) with geometric Brownian motion where the diffusivity $D_t$ is a piecewise-constant stochastic process (refreshed via a Poisson process). By applying Itô's formula to the logarithmic variable $y_t = \ln z_t$, the system is reduced to an additive equation driven by an integrated diffusivity $\Lambda_t = \int_0^t D_s ds$.

   • For an exponential redraw law of diffusivity, the authors derive the exact Fourier–Laplace propagator.
   • They analyze the asymptotic regimes, distinguishing between short-time behavior (where diffusivity is effectively frozen) and long-time behavior (where the process self-averages).

2. Interacting System (Mean-Field Redistribution):
   To address stationary tails, the authors introduce $N$ agents coupled by a mean-field redistribution matrix $J_{ij} = J/N$.

   • They define relative wealth $x_i = z_i / \bar{z}$ and derive the effective single-agent SDE for $x$ under the large-population limit, where the common growth term is removed.
   • To avoid finite-time moment divergences associated with unbounded redraw laws, they restrict the diffusivity to a two-state model ($D_{low}$ and $D_{high}$) with equilibrium weights $\beta$ and $1-\beta$.
   • The joint process $(x_t, D_t)$ is Markovian. The authors formulate coupled stationary Fokker–Planck equations for the probability densities in the low- and high-diffusivity states.
   • An exact tail analysis is performed by assuming algebraic decay modes $P(x) \sim x^{-1-\alpha}$ and solving for the condition where the determinant of the resulting linear system vanishes.

Key Contributions and Results

• Transient vs. Stationary Effects:

  • Single Agent: In the absence of redistribution, diffusing diffusivity induces non-Gaussian statistics at short times (a mixture of Gaussian kernels with varying variances). However, at long times, the logarithmic process self-averages, converging to a Gaussian distribution with an effective variance that includes a contribution from the diffusivity fluctuations ($2\sigma_D^2/\mu_r$).
  • Interacting System: Crucially, this self-averaging does not determine the stationary Pareto tail. The tail exponent is selected by the full statistics of the diffusivity trajectory, not merely its mean.

• Exact Pareto Exponent:
  For the two-state diffusivity model, the stationary Pareto exponent $\alpha_c$ is identified as the smallest root greater than unity of an explicit polynomial determinant $\Delta_\alpha = 0$:
  $$\Delta_\alpha = F_\ell(\alpha)F_h(\alpha) + \mu_r[\alpha J - \alpha(\alpha-1)\bar{D}] = 0$$
  where $F_i(\alpha) = \alpha J - \alpha(\alpha-1)D_i$ and $\bar{D}$ is the mean diffusivity.

• Interpolation Between Limits:
  The derived exponent $\alpha_c$ interpolates between two distinct physical limits:

  1. Fast-Refresh Limit ($\mu_r \to \infty$): The system recovers the standard Bouchaud–Mézard prediction using the mean diffusivity: $\alpha_{fast} = 1 + J/\bar{D}$.
  2. Slow-Refresh Limit ($\mu_r \to 0$): The system is dominated by the most volatile channel. The exponent becomes $\alpha_{slow} = 1 + J/D_{high}$.

  For any finite refresh rate, the authors prove that $\alpha_{slow} < \alpha_c < \alpha_{fast}$. This indicates that persistent volatility fluctuations systematically lower the Pareto exponent (making the tail heavier) compared to the prediction based on average volatility.

Significance and Claims
The paper claims that diffusing diffusivity serves a dual role: it controls transient, non-Gaussian fluctuations in single-agent growth but fundamentally selects the stationary Pareto tail in interacting systems.

The primary theoretical contribution is demonstrating that the stationary tail is not determined by replacing the fluctuating diffusivity with its mean value. Instead, agents that remain in high-diffusivity states for extended periods dominate the rare, large-wealth events. Consequently, the Pareto exponent is strictly lower than the classical Bouchaud–Mézard prediction ($1 + J/\bar{D}$) whenever volatility exhibits persistence.

The authors suggest this result offers a potential empirical signature for real-world socioeconomic systems. Since empirical data often shows persistent and heterogeneous volatility (e.g., in firm growth or financial returns), the displacement of the observed Pareto exponent below the value predicted by mean diffusivity could be attributed to the "diffusing diffusivity" mechanism described here. The paper concludes by noting that this framework provides a testable hypothesis: datasets with identical mean diffusivity but different persistence times should yield distinct Pareto exponents.