> For the complete documentation index, see [llms.txt](https://tonify-1.gitbook.io/tonify-docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://tonify-1.gitbook.io/tonify-docs/measurement/gamma.md).

# The coupling

The intensity–size coupling measured on public Last.fm logs, with its gates and its limits.

## The question

The lemmas hold for any play matrix. But one quantity decides *which side of the crossover* real artists fall on: how strongly the intensity of an artist's audience scales with the artist's size. In the simulation this is a dial, `γ`. On real data it is a slope: how `log(plays per pair)` moves with `log(audience)`. Call the measured version **b** — it is a proxy for `γ`, not the same object, because it also absorbs graph selection and retention.

## What was measured

Panel of Last.fm listening logs, 2005–2014. All gates passed before any number was read.

| Cut                                   | b (user fixed effects, two-way clustered SE) |
| ------------------------------------- | -------------------------------------------- |
| Full panel (\~10 years), 50/50 blend  | **+0.101 ± 0.004**                           |
| — mainstream listeners                | +0.160                                       |
| — beyond-mainstream listeners         | +0.023                                       |
| — window: calendar year               | +0.068                                       |
| — window: one month                   | +0.048                                       |
| — head of the distribution (A ≥ 200)  | +0.232                                       |
| Last.fm-360K (top-50 per user)        | +0.044                                       |
| BeyMS truncated to top-50 (falsifier) | +0.041                                       |

**The sign is solid inside this domain.** It does not flip in any cut: two datasets, two estimators, 27 threshold combinations, both accounting windows, both listener groups, every tail slice; the placebo (permuting audiences) gives ≈ 0. It is *not* a universal law — see [Prior art](https://tonify-1.gitbook.io/tonify-docs/measurement/prior-art.md) for a domain where the sign reverses.

**The level is not one number, and that is the result.** It depends on who you count and over what window.

## Three structural facts

**The slope is retention, not intensity.** Decomposed by timestamps: +0.101 = pair duration +0.254 plus per-day rate **−0.154**. Listeners of big artists stay with them *longer* but play them *less often per day*. Since real user-centric pools settle monthly, the policy-relevant slope is closer to +0.048.

**The 50/50 blend does not identify the level.** The dataset is constructed half mainstream, half beyond-mainstream; those groups give +0.160 and +0.023. A population figure would need reweighting by the real mainstreaminess distribution.

**The slope is convex — steeper at the head.** Locally at A ≥ 200 it is +0.232. For the question "does the top gain from user-centric" the head slope is the relevant one, not the global average. This is not an anomaly of our data: the Dirichlet model predicts exactly this shape, `b/(1−b)` as a function of penetration.

## The falsifier that worked

Truncating the full-history dataset to the top 50 artists per user moves b from +0.101 to +0.041 — almost exactly the value of the truncated dataset (+0.044). The gap between the two datasets is therefore an artefact of truncation, not a difference in the world. Our pre-registered guess about the *direction* of that bias was wrong, and the gate caught it.

## Feeding it back into the model

The estimator was run on the simulation's own data-generating process to measure its attenuation (ceiling discretisation eats \~16%, factor 0.84). Inverted, the panel b = +0.101 corresponds to a model **γ ≈ +0.12**; the head corresponds to ≈ +0.28. Those are the values the [simulation](https://tonify-1.gitbook.io/tonify-docs/simulations/sim4.md) now runs on — the model no longer turns a free dial.

![Slope on both datasets](https://raw.githubusercontent.com/ProximaCA/tonify-sims/main/figures/fig17_gamma_empirical.png)

## Limits

Last.fm is scrobbling, not a cash register; its demography is skewed; the curated dataset is not a random sample; b is a reduced form that absorbs graph selection as well as intensity; the epoch is pre-algorithmic. The precision `± 0.004` is sampling error only — the systematic uncertainty of sample composition and window choice is larger. All of this is an argument for the platform measuring its own value, not against measuring.

## Run it

```bash
python3 emp1/gamma_measure.py    # gates G1–G4, sensitivity grid, fig17
python3 emp1/moreau_check.py     # cross-check against published streaming estimates
```

Both need the external datasets (\~4 GB, not in git); sources and checksums are in [emp1/README.md](https://github.com/ProximaCA/tonify-sims/blob/main/emp1/README.md).


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation by asking a question.

Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters:

```
GET https://tonify-1.gitbook.io/tonify-docs/measurement/gamma.md?ask=<question>&goal=<user_goal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.
