> 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/theory/full.md).

# Full text

The paper in full — statements, proofs, assumptions, and the correction ledger.

Converted from the canonical LaTeX source at build time. The typeset PDF is [paper/THEORY.pdf](https://github.com/ProximaCA/tonify-sims/blob/main/paper/THEORY.pdf); the Russian source of truth, which carries a longer red-team ledger, is [paper/THEORY.md](https://github.com/ProximaCA/tonify-sims/blob/main/paper/THEORY.md) and is served on the Russian variant of this site.

Machine verification of every proved claim lives in [paper/theory\_check.py](https://github.com/ProximaCA/tonify-sims/blob/main/paper/theory_check.py) — 200 random matrices for the identity, Monte Carlo for the dominance blockade, the calibrated world for the Gini decomposition.

***

A streaming platform collects listener money and delivers it to artists. There are essentially three delivery mechanisms. *Pro-rata*: all money goes into one pool and an artist is paid their share of total platform plays — how Spotify and nearly every major service works. *User-centric*: each listener’s wallet is split only across the artists that listener actually played — the perennially proposed “fairer” model. *Direct*: no pool at all; money arrives from devoted fans (donations, artist subscriptions), as on Patreon or Twitch. On top of any mechanism may sit a label contract taking the lion’s share. This paper treats the three mechanisms as operators: input — the table of “who played whom how many times” over a period; output — each artist’s income over the same period.

Four questions are answered. *First*: who gains from a pro-rata $$\to$$ user-centric switch, and is it predictable in advance (Theorem 1: audience listening intensity decides, by an exact formula)? *Second*: can a pool-rule reform close the gap between a signed and an independent artist (Lemma 2: the income gap — never; the minimum-viable-audience gap — sometimes)? *Third*: does the direct economy reduce top-end concentration (it does not: it preserves the upper tail and adds a mass of exact zeros at the bottom — inequality is relocated, not reduced)? *Fourth*: is direct income more reliable than the pool (on small audiences it is a lottery, and no parameter values make it unconditionally better — Theorem 4)?

*What counts as an answer*: a proved statement under explicitly listed assumptions, plus machine verification — every identity and bound is recomputed by a deterministic script whose numbers are quoted in the text. Every claim carries one of three statuses: proved, refuted by counterexample, or conjecture. All four lemmas survived two adversarial red-team passes that did find errors — the correction history is Section 5. *Scope*: accounting statics at a fixed play matrix; equilibrium effects (a rule change altering behavior and hence the matrix) are outside and listed as open. The italicized paragraphs (“Why this matters” / “What it says”) form a complete non-technical route; proofs are set in small type.

## Conceptual formulation

*Entities*: artist (payee); listener (paying subscriber); the artist–listener pair (active if played at least once in the period); the play (one stream — the primitive cell); the wallet (a subscriber’s net contribution); the pool (sum of wallets; exists only under pool mechanisms); the contract (a post-rule multiplier); the superfan (an audience member paying directly; direct economy only); the gift (one direct payment; a superfan makes several per period).

*Primary vs derived vs assumed*: the play table over a period is the single primary object (a platform log, or a generated matrix); listener intensity, artist plays, audience and streams are measured from it; platform averages and per-artist audience statistics are derived arithmetic; model assumptions (equal wallets, the direct-economy probability model, contract multiplicativity, tail class of audiences) are collected in Section 3.3.

*“Artist popularity”* is not a latent parameter: audience $$A\_i$$ and streams $$s\_i$$ are measured properties of one specific play matrix over one period. In the synthetic world sim4 they are induced by a generative model calibrated to external anchors — same status: properties of the generated matrix. *The fork carried through the whole paper*: Lemmas L1–L4 (Section 3.4) are statements about ANY fixed play matrix, independent of how it arose; the sim4 generator (Section 4.1) produces one concrete matrix so that identities can be machine-checked at scale. No proof uses the generator. Prior art: the type-model crossover belongs to Alaei, Makhdoumi, Malekian & Pekeč (*Mgmt. Sci.* 2022; their notation is mapped to ours in Table [1](#tab:notation)); budget balance to Bergantiños & Moreno-Ternero (2025); the tail machinery is cited from Breiman (1965), Jessen & Mikosch (2006), Faÿ et al. (2006), Robert & Segers (2008), Denisov–Foss–Korshunov (2010), Maulik–Resnick–Rootzén (2002); what is new here: the general-matrix identity with the AM–HM frame, the Jensen dispersion premium, the kernel coincidence criterion, the invariance formalization, and the distributional direct-vs-pool comparison.

## Mathematical formulation

### Notation

| Symbol                                                                    | Type / range              | Verbal definition                                                                                             | Period / units            | Status                        |
| ------------------------------------------------------------------------- | ------------------------- | ------------------------------------------------------------------------------------------------------------- | ------------------------- | ----------------------------- |
| Symbol                                                                    | Type / range              | Verbal definition                                                                                             | Period / units            | Status                        |
| $$i$$, $$N$$                                                              | index; finite set         | artist; set (and number) of artists                                                                           | —                         | M                             |
| $$u$$, $$U$$                                                              | index; finite set         | listener; set (and number) of listeners                                                                       | —                         | M                             |
| $$p$$, $$piu$$                                                            | $$ℤ+N × U$$               | play matrix; plays of pair $$(i, u)$$                                                                         | period (year); plays      | M                             |
| $$Pu$$                                                                    | integer $$≥ 1$$ (A1)      | listener $$u$$’s intensity: $$∑ipiu$$                                                                         | year; plays               | D                             |
| $$Pi$$                                                                    | integer $$≥ 0$$           | artist $$i$$’s plays: $$∑upiu$$                                                                               | year; plays               | D                             |
| $$T$$                                                                     | integer $$> 0$$           | total platform plays: $$∑uPu = ∑iPi$$                                                                         | year; plays               | D                             |
| $$P̄$$                                                                    | real $$> 0$$              | mean intensity $$T/U$$                                                                                        | year; plays               | D                             |
| $$Ai$$                                                                    | integer $$≥ 0$$           | audience: number of $$u$$ with $$piu > 0$$                                                                    | year; listeners           | D                             |
| $$si$$                                                                    | real $$≥ 0$$              | artist streams (sim worlds: target yearly volume)                                                             | year; plays               | M / P                         |
| $$W$$, $$Wu$$                                                             | real $$> 0$$; $$≥ 0$$     | wallet (net subscriber contribution); unequal-wallet version                                                  | year; $                   | A (A2)                        |
| $$R$$                                                                     | real $$> 0$$              | the pool: $$U ⋅ W$$                                                                                           | year; $                   | D                             |
| $$wui$$                                                                   | probability vector        | listener $$u$$’s share of artist $$i$$’s plays: $$piu/Pi$$                                                    | —                         | D                             |
| $$mi$$, $$Hi$$                                                            | reals $$> 0$$             | arithmetic / harmonic mean audience intensity under $$wi$$                                                    | year; plays               | D                             |
| $$PRi$$                                                                   | real $$≥ 0$$              | artist $$i$$’s payout for the period under pro-rata, in wallet units ($)                                      | year; $                   | D                             |
| $$UCi$$                                                                   | real $$≥ 0$$              | artist $$i$$’s payout for the period under user-centric, in wallet units ($)                                  | year; $                   | D                             |
| $$Di$$, $$YA$$                                                            | random $$≥ 0$$            | artist $$i$$’s direct-economy payout for the period, $ ($$YA$$: at audience $$A$$)                            | year; $                   | model                         |
| $$Si$$                                                                    | random, $$Bin(Ai, σ)$$    | number of superfans of artist $$i$$                                                                           | year; people              | model                         |
| $$σ$$                                                                     | real $$∈ (0, 1)$$         | superfan share: probability a listener pays directly                                                          | —                         | A (0.017)                     |
| $$k$$                                                                     | integer $$≥ 1$$           | gifts per superfan per period                                                                                 | year; payments            | A (4)                         |
| $$G$$, $$Gft$$, $$ḡ$$                                                    | random $$> 0$$; real      | gift size; gift $$t$$ of fan $$f$$; mean gift $$𝔼G$$                                                         | $                         | A (lognormal; $$ḡ = 6.886$$) |
| $$τ$$                                                                     | real $$∈ (0, 1]$$         | artist’s share of a gift                                                                                      | —                         | A (0.80)                      |
| $$ρ$$                                                                     | real $$∈ (0, 1]$$         | contract pass-through to a signed artist                                                                      | —                         | D (0.06772)                   |
| $$F$$, $$Fi(p)$$                                                          | operator                  | arbitrary pool rule; its payout to artist $$i$$                                                               | year; $                   | model                         |
| $$θ$$                                                                     | real $$> 0$$              | payout threshold ($13 Stars)                                                                                  | $                         | P                             |
| $$A0$$                                                                    | integer $$> 0$$           | boost threshold in the artist-centric counterexample (15 000)                                                 | listeners                 | P                             |
| $$V$$                                                                     | real $$> 0$$              | target yearly income defining MVA ($1 200/yr)                                                                 | year; $                   | P                             |
| $$MVAs$$, $$MVAι$$                                                        | integers $$> 0$$          | minimum viable audience (signed / independent)                                                                | listeners                 | D                             |
| $$c$$, $$rate$$                                                           | reals $$> 0$$             | pool rate per listener $$c = PL ⋅ rate$$; external per-stream anchor (Duetti $$4.43/1000$$, D3, not in vault) | $/listener/yr; $/play     | D / A (calibration)           |
| $$PL$$                                                                    | real $$> 0$$              | LFM-1b lifetime mean 21.21; year transfer is calibration (CRITIC §2)                                          | year (calib.); plays/pair | M / A                         |
| $$XA$$                                                                    | real $$> 0$$              | pool income at audience $$A$$: $$cA$$ (deterministic)                                                         | year; $                   | D                             |
| $$d\*$$                                                                   | real $$> 0$$              | direct-vs-pool breakeven: $$PL ⋅ rate/(σḡτ)$$ payments/superfan/yr                                           | payments/yr               | D                             |
| $$φ(t)$$                                                                  | function, $$t > 0$$       | gift Laplace transform $$𝔼e−tτG < 1$$                                                                        | —                         | D                             |
| $$α$$, $$αG$$                                                             | reals                     | tail index of audiences; of gifts                                                                             | —                         | A (1.4) / M                   |
| $$L(x)$$                                                                  | function                  | slowly varying part of $$ℙ(A > x) = x−αL(x)$$                                                                 | —                         | A                             |
| $$Mi$$, $$m\*$$                                                           | random $$∈ (0, 1]$$; real | UC multiplier $$𝔼wi\[1/Pu]$$; its tail limit in (C)                                                          | 1/plays                   | D / A                         |
| $$q$$, $$qm$$                                                             | reals $$∈ \[0, 1)$$       | zero mass of direct $$𝔼\[(1 − σ)A]$$; measured zero share                                                    | —                         | D / M                         |
| $$Gpool$$, $$Gdirect$$, $$G+$$                                            | reals $$∈ \[0, 1]$$       | Gini of pool incomes; of direct; among positive incomes                                                       | —                         | D                             |
| $$n$$, $$n+$$, $$μ$$, $$μ+$$                                              | integers; reals           | artist counts (all / positive-income); mean incomes (Gini decomposition)                                      | —                         | D                             |
| $$δ0$$, $$F+$$, $$FX$$, $$FY$$                                            | distributions / CDFs      | point mass at zero; CDF of positive part; CDFs of $$XA$$, $$YA$$                                              | —                         | notation                      |
| $$RV(α)$$, $$op(1)$$                                                      | classes                   | regular variation of index $$α$$; vanishing in probability                                                    | —                         | notation                      |
| $$ℓi$$, $$Ãi$$                                                           | real $$≥ 1$$; integer     | sim4 target listeners $$max (1, si/PL)$$; realized audience                                                   | year; listeners           | D (sim4)                      |
| $$Ku$$                                                                    | integer $$∈ \[1, 2000]$$  | sim4 listener’s yearly playlist size                                                                          | year; artists             | A (LN($$ln 12$$, 1.0))        |
| $$γ$$, $$Āg$$                                                            | real; real $$> 0$$        | sim4 intensity–size coupling dial $${−0.3, 0, +0.3}$$; geometric-mean audience                                | —                         | P / D                         |
| $$Usil$$, $$Uact$$, $$Utot$$                                              | integers                  | silent, active, all subscribers (without A1)                                                                  | people                    | D                             |
| $$ε$$, $$a$$, $$t$$, $$x$$, $$z$$                                         | reals                     | local variables of limits and integrals                                                                       | —                         | notation                      |
| $$ρeff$$                                                                  | real $$∈ (0, 1]$$         | effective pass-through under component-heterogeneous contracts: $$∑mρmwim(F)$$                                | —                         | D                             |
| $$Δ$$                                                                     | operator                  | change of a quantity (e.g. $$Δ$$Gini pool$$→$$direct)                                                         | —                         | notation                      |
| $$Bin$$, LogNormal, $$corr$$                                              | standard                  | binomial and lognormal laws; Pearson correlation                                                              | —                         | notation                      |
| *Foreign notation (Alaei et al. 2022 type model) and dictionary to ours:* |                           |                                                                                                               |                           |                               |
| $$qi$$, $$λi$$                                                            | real; real                | mass and intensity of listener type $$i$$ $$↔︎$$ our group with $$Pu = λi$$                                   | —                         | foreign                       |
| $$πij$$                                                                   | real                      | type $$i$$’s preference share for artist $$j$$ $$↔︎$$ the type’s play distribution                            | —                         | foreign                       |
| $$λ̄$$                                                                    | real                      | $$∑iqiλi$$ $$↔︎$$ our $$P̄$$ on a type population                                                             | —                         | foreign                       |
| $$βpr$$, $$βuc$$                                                          | reals                     | their budget rates; equality holds by construction here (one pool $$R$$)                                      | —                         | foreign                       |

### The three operators: definitions, domains, edge cases

Payout operators map $$\mathbb{R}*+^{N\times U}\to\mathbb{R}*+^N$$. *Pro-rata*: $$\mathrm{PR}*i = UW,P\_i/T$$ — artist $$i$$’s payout for the period in wallet units; defined for $$T>0$$; $$P\_i=0\Rightarrow \mathrm{PR}*i=0$$ ($$A\_i=0$$ implies $$P\_i=0$$). *User-centric*: $$\mathrm{UC}*i = W\sum\_u p*{iu}/P\_u$$ — payout in wallet units; requires $$P\_u>0$$ for all paying $$u$$ (Assumption A1); $$p*{iu}/P\_u\le 1$$ always, so $$\mathrm{UC}*i\le W,A\_i$$: no listener, however light, blows up the payout. *Direct*: $$D\_i=\tau\sum*{f=1}^{S\_i}\sum*{t=1}^{k}G\_{ft}$$ with $$S\_i,|,A\_i\sim\mathrm{Bin}(A\_i,\sigma)$$, i.i.d. gifts $$G\ge 0$$, $$\mathbb{E}G=\bar g$$, all moments finite; $$\mathbb{E}\[D\_i|A\_i]=\tau\sigma k\bar g A\_i$$; at $$A\_i=0$$, $$D\_i\equiv 0$$; at $$A\_i>0$$, $$\mathbb{P}(D\_i=0)\ge(1-\sigma)^{A\_i}>0$$ — the atom at zero, the central object of Section 3.4. *Contract*: a multiplicative layer after the rule, $$\rho\cdot F\_i(p)$$; the payout threshold $$\theta$$ is a separate logistics layer (accumulation vs annual burn), treated as breaking channel B3 of Lemma 2.

### Assumptions

* (payer activity) Every paying subscriber listens: $$P\_u\ge 1$$. Silent payers are excluded from *both* pools; otherwise budget balance fails by exactly $$W\cdot U\_{\mathrm{sil}}$$ and all ratios acquire the factor $$U\_{\mathrm{act}}/U\_{\mathrm{tot}}$$ (red-team counterexample: 2 artists $$\times$$ 3 active listeners plus 2 silent give $$3/5=0.6$$). *Used by: L1 (all parts), L3 (UC bound).*
* (equal wallets) One subscription price; measured approximation. The unequal-wallet generalization preserves the identity’s structure (Section 3.4, Theorem 1 discussion). *Used by: L1, L2, L4.*
* (integer plays) $$p\in\mathbb{Z}*+^{N\times U}$$, hence $$1/P\_u\le 1$$ and $$M\_i\le 1$$; with real-valued plays the weaker bound $$p*{iu}/P\_u\le 1$$ survives. *Used by: L3 (one-sided UC tail bound).*
* (fixed matrix; accounting frame) $$p$$ is fixed: no equilibrium, pool-size, or welfare claims. *Used by: all lemmas.*
* (direct-economy model) $$S\_i|A\_i\sim\mathrm{Bin}(A\_i,\sigma)$$ with $$\sigma\perp A\_i$$; gifts i.i.d., all moments finite; $$k$$ deterministic. *Used by: L3 (direct branch), L4.*
* (multiplicative contract, single $$\rho$$) The contract multiplies the payout by a constant $$\rho$$, uniform across income components and markets; violations are channels B1, B2, B5. *Used by: L2.*
* (rule applied at artist level) $$F$$ acts at the artist level, the contract strictly downstream; domain $$F\_i(p)>0$$; violation is channel B6. *Used by: L2 (2a).*
* (heavy-tailed audiences) $$\mathbb{P}(A>x)=x^{-\alpha}L(x)$$, $$\alpha>1$$ (sim1 world: $$\alpha=1.4$$, gate T3; empirical caveat Spierdijk & Voorneveld). *Used by: L3.*
* (tail-conditional concentration; UC branch of L3 only) For every $$\varepsilon>0$$, $$\mathbb{P}(|M\_i-m^\ast|>\varepsilon,|,A\_i>a)\to 0$$ as $$a\to\infty$$, $$m^\ast>0$$. Measurable on full log datasets.

### Lemmas and theorems

The fork of Section 2 stands: everything here is about an arbitrary fixed play matrix; no proof uses the sim4 generator.

**L1: the ratio identity.**

*Uses assumptions A1, A2, A4.*

**Theorem 1** (Identity). For any artist $$i$$ with $$P\_i>0$$, under equal wallets,

$$
\frac{\mathrm{UC}*i}{\mathrm{PR}*i}
\=\mathbb{E}*{w^i}!\left\[\frac{\bar{P}}{P\_u}\right],
\qquad w^i\_u=\frac{p*{iu}}{P\_i}.
$$

*Proof.* $$\mathrm{UC}*i/\mathrm{PR}*i =\bigl(W\sum\_u p*{iu}/P\_u\bigr)\big/\bigl(UW P\_i/T\bigr) =(T/U)\sum\_u (p*{iu}/P\_i)(1/P\_u)=\bar{P},\mathbb{E}\_{w^i}\[1/P\_u]$$. ◻

With $$m\_i=\mathbb{E}*{w^i}\[P\_u]$$ (arithmetic) and $$H\_i=1/\mathbb{E}*{w^i}\[1/P\_u]$$ (harmonic mean of listener intensity under play weights),

$$
\frac{\mathrm{UC}*i}{\mathrm{PR}*i}=\frac{\bar{P}}{H\_i}
\=\underbrace{\frac{\bar{P}}{m\_i}}*{\text{level}}\cdot
\underbrace{\frac{m\_i}{H\_i}}*{\text{dispersion}\ \ge 1},
$$

the second factor $$\ge 1$$ by AM–HM, with equality iff $$P\_u$$ is constant on the support of $$w^i$$. Under unequal wallets $$W\_u$$ the identity becomes $$\mathrm{UC}\_i/\mathrm{PR}*i=\mathbb{E}*{w^i}\[W\_u/P\_u]/(\sum\_u W\_u/T)$$ — same structure, “intensity” replaced by “plays per dollar”.

**Corollary (a)** (Coincidence; the naive converse is false). $$P\_u\equiv\bar{P}$$ implies $$\mathrm{UC}=\mathrm{PR}$$ componentwise. The converse fails: with one artist and two listeners, $$p=(10,20)$$, intensities differ yet $$\mathrm{UC}\_1=\mathrm{PR}\_1=2W$$. Exact criterion: $$\mathrm{UC}\_i=\mathrm{PR}\_i$$ for all $$i$$ iff the vector $$c\_u=1/P\_u-U/T$$ lies in $$\ker p$$.

**Corollary (b)** (Zero sum). $$\sum\_i \mathrm{UC}\_i=\sum\_i\mathrm{PR}\_i=UW$$: user-centric is a pure redistribution of the same pool.

**Corollary (c)** (Crossover). $$\mathrm{UC}\_i>\mathrm{PR}*i$$ iff $$H\_i<\bar{P}$$. On type populations this reduces to the condition of Alaei, Makhdoumi, Malekian and Pekeč (*Mgmt. Sci.* 2022, Props. 2, 8): artist $$j$$ weakly prefers pro-rata iff $$\sum\_i q\_i\pi*{ij}(\lambda\_i-\bar\lambda)\ge 0$$.

**Corollary (d)** (Dispersion premium). Fix $$m\_i=\bar{P}$$. Then $$\mathrm{UC}\_i/\mathrm{PR}\_i=m\_i/H\_i\ge 1$$, with equality iff audience intensity is degenerate: at the same mean level, *within-audience dispersion strictly raises* user-centric income.

**L2: pass-through invariance.**

*Uses assumptions A2, A4, A6, A7 (part 2a); A2, A4, A6 and a monotone income with threshold* $$V$$ *(part 2b).*

**Lemma 2** (Two-part invariance). *(2a)* For any pool operator $$F$$ *applied at the artist level*, with the multiplicative contract strictly downstream and on the domain $$F\_i(p)>0$$: $$\rho F\_i/F\_i=\rho$$ — the signed/independent *income* gap is invariant to the rule (a nonlinear $$F$$ applied to a licensor’s *catalogue* breaks this). *(2b)* The *minimum-viable-audience* gap satisfies $$\mathrm{MVA}*s/\mathrm{MVA}*\iota=1/\rho$$ iff $$F^{-1}(V/\rho)=F^{-1}(V)/\rho$$; linearity of $$F$$-income in audience is sufficient but *not* necessary (a strictly increasing log-periodic $$F$$ is nonlinear yet yields $$1/\rho$$ for all $$V$$). A Deezer-style $$\times 2$$ boost above a threshold yields $$1/(2\rho)$$ instead (counterexample in companion code) while the income ratio stays $$\rho$$ pointwise.

**L3: tail action and the atom at zero.**

*Uses assumptions A1, A3, A4, A8; condition (C) — UC branch only; A5 — direct branch only.*

Let $$\mathbb{P}(A>x)=x^{-\alpha}L(x)$$. PR is linear: index preserved. UC: $$\mathrm{UC}*i=WP\_i,\mathbb{E}*{w^i}\[1/P\_u]$$ with bounded multiplier; under tail-conditional concentration of the multiplier the index is preserved, and heavier tails are impossible for any dependence. Direct: binomial thinning plus a light-tailed compound sum preserve $$\alpha$$ (Jessen–Mikosch 2006, Lem. 3.7; Faÿ et al. 2006; Denisov–Foss–Korshunov 2010; Breiman 1965 enters only through the multiplicative contract layer), while $$\mathbb{P}(D\_i=0,|,A\_i)\ge(1-\sigma)^{A\_i}\ge e^{-1}(1-\sigma)$$ for $$A\_i\le 1/\sigma$$: an *atom at zero* on small audiences. With zero mass $$q$$, $$G\_{\mathrm{direct}}=q+(1-q)G^{+}$$: the mechanism relocates inequality from the intensive to the extensive margin; no mechanism softens the upper tail. Under A5 and a light-tailed check, direct preserves the regular-variation *index* of income. That by itself does not imply equality of top shares, tail constants, or Gini.

**L4: dominance blockade.**

*Uses assumptions A2, A4, A5; the anchor* $$c=\mathrm{PL}\cdot\mathrm{rate}$$ *(independent-artist comparison).*

Fix $$A$$ with $$(1-\sigma)^A>0$$ and pooled income $$X\_A=cA>0$$ (deterministic given $$A$$); $$Y\_A=D\_A$$ as above.

**Theorem 4** (Full blockade). (i) $$Y\_A$$ never first-order dominates $$X\_A$$; (ii) $$Y\_A$$ never second-order dominates $$X\_A$$, for any parameters; (iii) $$X\_A\succeq\_2 Y\_A$$ iff $$\mathbb{E}Y\_A\le X\_A$$, i.e. iff $$k\le d^\ast=\mathrm{PL}\cdot\mathrm{rate}/(\sigma\bar g\tau)$$; (iv) for $$k>d^\ast$$ the pair is incomparable in both orders.

*Proof.* (i) For $$x\in(0,X\_A)$$: $$\mathbb{P}(X\_A>x)=1$$ while $$\mathbb{P}(Y\_A>x)\le 1-(1-\sigma)^A<1$$. (ii) SSD against degenerate $$X\_A$$ requires $$\int\_0^t F\_Y\le\max(0,t-X\_A)$$; at $$t=X\_A$$ the left side is $$\ge(1-\sigma)^A X\_A>0$$. (iii) ($$\Leftarrow$$) $$\int\_0^t F\_Y=t-\mathbb{E}\min(Y,t)\ge t-\mathbb{E}Y\ge t-X\_A$$ and $$\ge 0$$. ($$\Rightarrow$$) $$t\to\infty$$ gives $$\mathbb{E}Y\le X\_A$$. (iv) From (i)–(iii). ◻

*Remark 1*. The correct comparison is mean versus risk: for $$k>d^\ast$$ direct pays an expectation premium priced by small-audience lottery risk; $$\mathbb{P}(Y\_A\<X\_A)\le\inf\_{t>0}e^{tX\_A}\bigl(1-\sigma+\sigma\varphi(t)^k \bigr)^A$$ with $$\varphi(t)=\mathbb{E}e^{-t\tau G}<1$$ — exponential decay in $$A$$. In the companion sim1 calibration at $$A{=}30$$, $$k{=}4$$: mean advantage $$\times 4$$, yet $$\mathbb{P}(Y\<X)\approx 0.60$$ — and on the full 2M-sample run $$\mathbb{P}(Y\<X)-\mathbb{P}(Y=0)=0$$ to machine precision: losing to the pool is exactly the zero outcome. Under the contract layer $$c=\mathrm{PL}\cdot\mathrm{rate}\cdot\rho$$ the threshold scales to $$\rho,d^\ast\approx 0.068$$: for a signed artist, direct wins in expectation already at $$k\ge 1$$.

## Numerical implementation

*Generative model (sim4).* Artist popularity: piecewise $$s\_i\sim$$ (lognormal body $$/$$ log-bridge $$/$$ Pareto($$\alpha{=}1.4$$) tail), calibrated to anchors $$T\_1{:},87%$$ of artists below $$10^3$$ streams/yr, $$T\_2{:},2.6%$$ above $$225{,}734$$, $$T\_3{:}$$ top $$0.28%$$ holding $$40$$–$$55%$$ of streams (gate band; the seed-42 matrix realizes $$40.3%$$, sim1’s larger world $$44.5%$$); target audiences $$\ell\_i=s\_i/\mathrm{PL}$$, $$\mathrm{PL}=21.21$$ plays per pair (LFM-1b). Playlist sizes $$K\_u\sim\mathrm{LogNormal}(\ln 12,,1.0)$$ (assumed; an MVP-measured dial). Artist selection: exact weighted sampling without replacement, $$\mathbb{P}(i\in S\_u)\propto \ell\_i$$ (Gumbel top-$$K\_u$$), i.e. preferential attachment. Pair intensity, three named regimes on one graph:

$$
\text{(a) ergodic: integer } p\_{iu},\ P\_u\equiv 2048;\qquad
\text{(b) } p\_{iu}=\lceil\mathrm{LN}(\ln 5.16,,1.676)\rceil
\text{ rescaled to } s\_i;\qquad
\text{(c) } p\_{iu}\propto \mathrm{LN}\cdot(A\_i/\bar A\_g)^{\gamma}.
$$

Regime (a) has $$P\_u\equiv 2048$$ exactly (a power of two: every $$p\_{iu}/P\_u$$ is a dyadic rational, so the control identity is exact in floating point) and is the control (Corollary (a)); $$\gamma$$ dials the intensity–size coupling.

*Machine verification* (deterministic, seed 42; PASS/FAIL gates printed before any conclusion, FAIL exits with code 1): the Theorem-1 identity and AM–HM decomposition on 200 random matrices (max error $$8.9\times 10^{-16}$$), on the full 21.8M-pair sim4 matrix ($$7.1\times 10^{-15}$$), and — since August 2026 — on *real* Last.fm logs (31,040 artists, 1.02M pairs: $$5.8\times 10^{-15}$$, with AM $$\ge$$ HM holding artist by artist); zero-sum ($$1.8\times 10^{-14}$$); world anchors T1–T3; the exact-zero control gate ($$\max|\mathrm{UC}/\mathrm{PR}-1|=0$$); the L4 atom at $$A{=}30$$ (0.598 vs theory 0.598, 2M samples) with $$\mathbb{P}(Y\<X)-\mathbb{P}(Y=0)=0$$; the calibrated-world Gini identity $$q\_m+(1-q\_m)G^{+}=G\_{\mathrm{direct}}=0.973844$$ to machine zero ($$G\_{\mathrm{pool}}=0.972$$; robust to lognormal gifts); both L2 counterexamples ($$\mathrm{MVA}$$ 12 772 / 94 296, ratio $$1/(2\rho)$$; the log-periodic $$F$$ giving exactly $$1/\rho$$); crossover-sign gate: zero violations of Corollary (c) across all artists and regimes.

*Reproducibility*: python3/numpy/scipy/matplotlib, no external services. `git clone https://github.com/ProximaCA/tonify-sims`; `python3 paper/theory_check.py` (seconds; CHECK 1–6); `python3 sim4/bipartite_gen.py` ($$\sim$$7 min; gates G1–G3, exports, figures 15–16); `python3 run_all.py` (everything). All numbers quoted above are outputs of these scripts.

## Red team: two adversarial passes

Two independent red-team passes attacked every lemma (recompute every inequality, reproduce every derivation, hunt counterexamples). Pass one: five attackers, verdict “sound with fixes” in all five zones — nine confirmed lemma findings and ten consistency findings. Pass two: five fresh attackers filed 25 accusations, all 25 confirmed by independent verification. No lemma broke in either pass. Retracted: one inequality (the atom bound misprint, corrected to $$e^{-1}(1-\sigma)$$), one concentration bound (Hoeffding, violated $$\sim$$45$$\times$$ by random gifts; replaced by the Chernoff bound via $$\varphi(t)$$), the necessity claim “only if linear” in (2b) (log-periodic counterexample; exact criterion $$F^{-1}(V/\rho)=F^{-1}(V)/\rho$$ installed), and one simulation narrative (head saturation 5.00 — circular sampling; the real head has $$\tilde A/\ell=1.08$$). Introduced: load-bearing assumption A1 (silent subscribers), breaking channels B5 and B6, the artist-level assumption A7. Full protocols live in the working-session journals outside the repo; the Russian THEORY.md carries the episode-by-episode narrative.

## Open questions

(1) Equilibrium versions of L1/L2 (endogenous $$p$$); does the Jensen premium survive? (2) The sign of $$\Delta$$Gini pool$$\to$$direct in general (conjecture: a threshold $$\sigma^\ast(\alpha,k)$$ exists for the positive-income Gini, never for the full one). (3) Exact characterization of when UC thins the tail (Maulik–Resnick–Rootzén). (4) L4 for recurrent $$k$$ with churn (dominance on income flows). (5) An axiomatic characterization of direct alongside the B\&MT axiomatizations of PR and UC.

## Empirical bridge

Each lemma rests on a measurable quantity: $$k$$ (payments per superfan per year; benchmarks Twitch 12, Patreon 12, breakeven median 1.25); $$\sigma(A)$$ (superfan conversion vs audience size — tests A5); intensity–size coupling across artist sizes (closes condition (C)) — **now measured**: on Last.fm logs the reduced pair-intensity slope is $$+0.101\pm0.004$$ overall, $$+0.232$$ at the head, $$+0.023$$ among beyond-mainstream listeners, decomposing into retention $$+0.254$$ and an *anti*-coupled per-day rate $$-0.154$$ (companion `emp1`); the gift tail $$\alpha\_G$$ (whales vs attention regime); $$\rho(F)$$ (pass-through endogeneity, channel B1 — the Deezer artist-centric switch is a natural experiment); the threshold $$\theta$$ and its censoring mode (channel B3).


---

# 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/theory/full.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.
