Scaling

By Stefan Thurner 4, Rudolf Hanel, Peter Klimekl. 2018

Scaling appears practically everywhere in science; it basically quantifies how the properties or shapes of an object change with the scale of the object. Scaling laws are always associated with power laws. The scaling object can be a function, a structure, a physical law, or a distribution function that describes the statistics of a system or a temporal process. We focus on scaling laws that appear in the statistical description of stochastic complex systems, where scaling appears in the distribution functions of observable quantities of dynamical systems or processes. The distribution functions exhibit power laws, approximate power laws, or fat-tailed distributions. Understanding their origin and how power law exponents can be related to the particular nature of a system, is one of the aims of the book.We comment on fitting power laws.

Subfields: Scaling & Learning Theory

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Scaling & Learning Theory

1. Guy F. de Téramond
2. Armand Joulin
3. Colin Raffel
4. S. Brodsky
5. A. Deur

Citations in the corpus

• Few-shot Learning with Retrieval Augmented Language Models (2022, methodology): “Lazaridou et al. (2022) explore NaturalQuestions in a 15-shot setup using Gopher, augmenting questions with 50 passages retrieved using Google Search.”
• The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale (2024, methodology): “WebText [20] (used to train GPT-2) involves only those non-Wikipedia webpages that were linked to from Reddit posts with at least 3 karma, with text extracted using Dragnet [39] and Newspaper1 [40] and an unspecified deduplication pipeline.”
• The QCD running coupling (2016, methodology): “In addition, the integral over xBj and global parton–hadron duality [132] amplify the dominance of non-resonant reactions.”
• The Forward Physics Facility: Sites, experiments, and physics potential (2021, background): “In the electromagnetic sector, these can be described by the quark–hadron duality phenomenon that provides a connection between the average value of interaction strengths in the quark–gluon description of the DIS formalism at high Q 2, and the average value of interaction strengths in the pion–nucle”
• Run, Don't Walk: Chasing Higher FLOPS for Faster Neural Networks (2023, background): “One notable trend is to pursue a better accuracy-latency trade-off by reducing the complexity of the attention operator [1,29,45,63,68], incorporating convolution into ViTs [6,10,57], or doing both [3,34,49,52].”
• The Mathematical Theories of Diffusion: Nonlinear and Fractional Diffusion (2017, background): “Then we introduced the renormalized Fokker-Planck equation and used a suitable entropy functional and proved stabilization of general solutions to the previous profiles that we called fractional Barenblatt profiles.”
• Spectral statistics across the many-body localization transition (2015, background): “The fractality of the wave function near the mobility edge results in a power-law enhancement of c(ω) ∝ A/ω [35, 37].”