Self-Similar Fragmentation Class

• Self-similar fragmentation class is defined as a Markovian process where fragment breakups scale by a power-law using a self-similarity index α.
• The process is characterized by a triplet (α, c, ν) and analyzed through non-local PDEs, cumulant functions, and martingale techniques, highlighting its analytical structure.
• It has practical applications in random tree theory, statistical mechanics, and astrophysics, providing insights into fractal dimensions and long-term scaling behavior.

A self-similar fragmentation class comprises stochastic processes that model the evolution of systems in which entities break up (fragment) into smaller entities over time, in such a way that the fragmentation dynamics are invariant under scaling—the rates and statistics of breakups depend on the size or “mass” of fragments in a power-law (self-similar) form. These processes are central in probability, statistical physics, and applied mathematics, with connections to random trees, stochastic PDEs, statistical mechanics, and even astrophysics.

1. Mathematical Definition and Classification

A self-similar fragmentation process is rigorously defined as a Markovian process whose state at time $t$ is a (possibly infinite) mass partition $s = (s_1 \geq s_2 \geq \cdots \geq 0)$, often with total mass $\sum_i s_i \leq 1$. The key structural datum is:

• Self-similarity index $\alpha \in \mathbb{R}$: Each fragment of mass $m$ evolves independently, and its fragmentation clock is sped up by a factor $m^\alpha$ (Stephenson, 2013).

If $\mathcal{P}$ is the mass-partition space, the process $X(t)$ satisfies the branching property: conditionally on $X(t) = (x_i)$, the sub-processes tracking sizes of the descendants of each $x_i$ are independent and distributed as scaled copies of the original.

The law of a self-similar fragmentation is fully determined by a triple $(\alpha, c, \nu)$ where:

• $\alpha$ is the self-similarity index.
• $c \geq 0$ is an erosion rate (continuous mass loss).
• $\nu$ is a $\sigma$-finite dislocation measure on decreasing sequences $s = (s_1, s_2, ...)$ with $\sum_i s_i \leq 1$, satisfying $\nu(1,0,0,\ldots) = 0$ and $\int (1 - s_1) \nu(ds) < \infty$ (Stephenson, 2013).

Generalizations incorporate types (multi-type fragmentations (Stephenson, 2017)) or fragment "marks" with self-similar branching speeds (Duchamps, 2019).

2. Fundamental Analytical and Probabilistic Structure

Self-similar fragmentation equations can be expressed as non-local PDEs or as measure-valued equations for distributions $u(t,x)$ of sizes, typically: $$\partial_t u(t,x) + \partial_x[\tau(x)u(t,x)] + B(x) u(t,x) = \int_x^\infty k(y,x) B(y) u(t,y) dy$$ with scaling choices $\tau(x) = c x^{\alpha+1}$, $B(x) = x^\alpha$ (Bertoin et al., 2015, Mischler et al., 2013). Crucial analytical quantities include:

• The cumulant function $\kappa(q)$, often with structure:

$$\kappa(q) = a q^2 + (b-a) q + \int_{\mathcal{S}} \left(1 - \sum_{i} s_i^q\right) \nu(ds)$$

whose convexity and zeros govern existence, uniqueness, and long-term behavior (Bertoin et al., 2015, Shi, 2016, Bertoin et al., 2016).

• Mellin transforms, generating exponential martingales $\sum_i s_i^q$ (Bertoin et al., 2015).

Self-similar fragmentation can be realized as an interacting family of positive self-similar Markov processes (pssMp), via Lamperti transforms of Lévy processes. Negative jumps correspond to splits, and time-changes encode the self-similarity (Shi, 2016, Dadoun, 2016).

3. Canonical Examples and Classification Results

Fragmentation Regimes and Classification Theorems

• Homogeneous class ($\alpha=0$): Standard compensated fragmentations or branching Lévy processes; law determined by $\kappa$ (Shi, 2016).
• Critical fragmentation class: In coagulation–fragmentation models with homogeneous kernels, mass-conserving self-similar solutions exist if and only if the fragmentation exponent is the “borderline” value $\gamma = \lambda-1$, where $\lambda$ is the degree of homogeneity of the coagulation kernel. Existence requires mass below a threshold $\rho_*$ (Laurençot, 2019).

Explosion / Non-explosion dichotomy: The existence of a (Malthusian) $q$ such that $\kappa(q) = 0$ is both necessary and sufficient for the absence of local explosion (i.e., that all fragment sizes can be listed in a decreasing sequence tending to zero). If $\kappa(q) > 0$ for all $q$ and $\alpha \neq 0$, the process explodes locally: at some time, infinitely many fragments occupy any fixed size interval (Bertoin et al., 2016).

Table: Summary of Key Structural Data

Attribute	Notation	Role
Self-similarity index	$\alpha$	Controls scaling of fragmentation rates
Dislocation measure	$\nu$	Governs breakup statistics
Erosion (drift)	$c$	Continuous mass loss
Cumulant function	$\kappa(q)$	Governs moments/martingales, explosion
Malthusian exponent	$q_*$ or $p^*$	Root of $\kappa(q) = 0$, critical for non-explosion
Genealogical tree	$(T,d,\mu)$	Encodes the splitting genealogy with natural measure

4. Profiles, Martingales, and Fractality

Fragmentation processes canonically generate continuum trees with intrinsic measure structures. Given the genealogy tree (constructed via split times and death times), there is a “natural” measure $\mu$ or $A_T$ supported on the leaves. The “profile” $A(t)$ is the measure of leaves at height at most $t$. Regularity properties of this profile are sharply characterized in terms of $\alpha$ and $\kappa$:

• Absolute continuity: $dA$ is a.s. singular w.r.t. Lebesgue measure iff $\alpha \leq -\omega_-$, and absolutely continuous (has $L^2$-density) iff $\alpha > -\omega_-$, where $\omega_-$ is the unique positive solution to $\kappa(\omega_-) = 0$ (Ged, 2017).
• Hausdorff dimension: In the singular regime $-\omega_+ < \alpha \leq -\omega_-$, $\dim_H(dA(t)) = \omega_-/(-\alpha)$ almost surely (Ged, 2017, Stephenson, 2013, Stephenson, 2017).
• Scaling and martingales: The additive martingale $M_q(t) = e^{-t \kappa(q)} \sum_{u} \chi_u(t)^q$ is a.s. convergent for appropriate $q$ and critical for the limiting genealogy and dimensions (Dadoun, 2016, Ged, 2017).

Self-similar growth-fragmentation, an extension where fragments may grow or shrink between splitting, is uniquely determined by the pair $(\alpha, \kappa)$, with the classification result proved via the notion of “bifurcator” couplings (Shi, 2016).

5. Long-time Asymptotics, Profiles, and Extremal Statistics

As $t \to \infty$, fragmentation processes display statistical scaling—self-similar profiles emerge, and extremal behavior can be analyzed:

• Empirical measures: In the positive-index case, the empirical measure $$\rho_t^{(\alpha)}(dy) = \sum_{i} X_i(t)^{\omega_-} \delta_{t^{1/\alpha} X_i(t)}(dy)$$ converges to a deterministic profile (Dadoun, 2016).
• Largest fragments: The size of the largest fragment decays as $X_1(t) = t^{-1/\alpha + o(1)}$ almost surely (Dadoun, 2016, Dyszewski et al., 2024). Recent sharp results identify sublogarithmic corrections and refined clustering at the extreme edge:

$$m_t = \frac{\log t}{\alpha} - \frac{(1-\theta)\log\log t}{\alpha} + o(\log\log t)$$

for fragmentation with infinite activity (tail exponent $\theta$), where $m_t = -\log X_1(t)$ (Dyszewski et al., 2024).

Convergence to self-similar profile: In classical and critical fragmentation equations, necessary and sufficient conditions (often a log-moment criteria for the fragmentation kernel) guarantee the solution converges in norm to a unique self-similar profile as $t \to \infty$ (Biedrzycka et al., 2017, Laurencot et al., 2014).

6. Structural Variations and Applications

• Multi-type fragmentations: Incorporate types (or marks) for each block, with rates and dislocation laws possibly depending on type. The genealogy is encoded in a colored/marked $\mathbb{R}$-tree, and the (multi-dimensional) Hausdorff dimension is $p^*/|\alpha|$, where $p^*$ is the solution to a matrix-valued Malthusian equation (Stephenson, 2017).
• Self-similar branching speeds: Each fragment carries a mark (e.g., a clock), evolving in a self-similar fashion, and determining its fragmentation rate. The law is characterized by $(\alpha, c, d, B, A)$, incorporating both mass and mark parameters, generalizing the classical class (Duchamps, 2019).
• Physical and astrophysical regimes: In magnetically regulated star-forming filaments, force-balanced, self-similar fragmentation arises, characterized by observable scaling laws (e.g., $B \propto n^{0.41}$) and morphologies repeating self-similarly over several orders of magnitude (Li et al., 2015).

7. Practical and Theoretical Implications

Self-similar fragmentation models have direct connections to random tree theory (the genealogy tree is a random measured $\mathbb{R}$-tree), random planar maps and statistical geometry (volume profiles correspond to self-similar growth-fragmentation (Ged, 2017)), fractal geometry (dimension results and invariant measures), aging and lifetimes in population models, and critical balance in cluster kinetics. Spectral analysis via semigroup methods reveals exponential convergence to scaling profiles for a broad class of fragmentation rates (Mischler et al., 2013).

Recent work has refined the classification boundaries (e.g., via bifurcator uniqueness proofs (Shi, 2016)), lifted log-moment criteria, and demonstrated the universality of scaling exponents and fractal characteristics.

---

References

Key results on the definition, classification, genealogy, martingales, fractal structure, and extremal asymptotics are proven or summarized in (Stephenson, 2013, Bertoin et al., 2015, Bertoin et al., 2016, Shi, 2016, Dadoun, 2016, Biedrzycka et al., 2017, Stephenson, 2017, Ged, 2017, Laurençot, 2019, Duchamps, 2019, Dyszewski et al., 2024, Mischler et al., 2013, Li et al., 2015, Laurencot et al., 2014).