NOTE-068: Tembine et al., mean-field-type games with Rosenblatt noise

Contribution

New to this paper:

• a survey-style argument that many engineered and socio-economic quantities are non-Gaussian (§II);
• the framing of Rosenblatt noise as a “baseline” for them;
• a restatement of earlier Rosenblatt-noise control and game results (Thms 1, 3, 4, 6, 7, 10 are from refs [27], [41]–[45]);
• a comparison of mean-field-type games with multi-population coalitional mean-field games in a Cournot energy market (§XIII, Table IV, Thms 12–19);
• a sketch of “Rosenblatt foundational diffusion models” (§XIV).

What is actually established:

• the ergodic LQ gain and cost under Rosenblatt noise for scalar linear feedback, and the analogous best-response and equilibrium formulas;
• the MFTG-versus-coalitional-MFG gap in a static variance-aware Cournot game (the “price of simplicity” is unbounded under scaling, Table IV).

Neither needs the noise to be non-Gaussian.

Key insight

The intended one is that model noise as it is (skewed, heavy-tailed, long-memory), because optimising against a Gaussian surrogate is suboptimal. The mathematics shows something narrower. In the linear-quadratic, linear-feedback setting the paper works in, only the noise covariance matters. So the relevant misspecification is getting H (the memory) wrong, not getting the Gaussianity wrong. Non-Gaussianity would matter only with non-linear strategies, non-quadratic costs of the state, or risk measures beyond variance. The paper names these as open or leaves them unaddressed (“the extension to non-linear strategies is an open issue”, pp. 30–31).

Assumptions

• Noise. The Rosenblatt process R_H, H ∈ (½, 1): a double Wiener–Itô integral, self-similar with R_H(ct) =ᵈ c^H R_H(t), with fBM’s covariance, and not a semimartingale (Def. 1, §III).
• Calculus. An Itô-type formula for x(t) = x₀ + ∫d₁dt + 2c̃∫d₂ dB^{H/2+1/2} + ∫d₃ dR_H, with f ∈ C² in t, C³ in x, and |f_xxx| ≤ c_t(1 + |x|^α) (Thm 1, from Čoupek–Duncan–Pasik-Duncan 2022 [27]).
• Control. Scalar linear dynamics dx = (b₁x + b₂u)dt + dR_H, ergodic cost lim sup (1/T)E∫(qx² + ru²), with the admissible set restricted to u = Kx, K ∈ ℝ (Thm 4). The same restriction to linear (and mean-field-type linear) feedback applies to Thms 5–18.
• Games. Scalar state; zero-sum requires an unexplained inequality on (b₁, b₂, b₃, q, r, s, H) (§IX). The finite-horizon MFTG (Thm 10) requires a Riccati-type ODE system with positive non-blowing-up solutions.

Key results

• Thm 4 (from [43]).
  • Optimal gain: K̂ = −[b₁ + √(b₁² + 4H(1−H)b₂²q/r)]/(2b₂(1−H)).
  • Cost: L∞(K̂) = Γ(2H+1)(q + rK̂²)/(2[−(b₁+b₂K̂)]^{2H}) = Γ(2H)(−rK̂/b₂)/[−(b₁+b₂K̂)]
  • I verified the first-order condition (1−H)b₂K² + b₁K − Hb₂q/r = 0, that the chosen root is the stabilising one, the grid minimum, and that H → ½ gives the Brownian ergodic LQR gain −(b₁ + √(b₁² + b₂²q/r))/b₂.
• Thm 5 (variance-aware). It has the same K* and a mean-field gain K̄* = (b₂+b̄₂)q̄/(r̄(b₁+b̄₁)), stabilising iff b₁ + b̄₁ < 0.
• Thms 6–7 (zero-sum, from [44]). A value exists and saddle-point linear gains solve a quadratic, with no proof of Thm 6 here.
• Thms 8–9 (non-zero-sum). Best-response gains are of Thm 4’s form with b₁ replaced by b₁ + Σ_{j≠i}b₂ⱼKⱼ. Equilibrium holds if the fixed-point system has a stabilising solution.
• Thm 10 (MFTG, from [45]). A linear-and-mean-field feedback Nash equilibrium for finite horizon with x̄^{2k̄} mean-field costs, given solvable ODEs. Thm 11 is the cooperative version.
• Thms 12–18 (Cournot MFTG). The payoff is (ηᵢ − rᵢηᵢ²/2)·Γ(2H+1)/(2[(1+Ση)/ε]^{2H}) plus a deterministic mean part. The best-response ηᵢ is the root of a quadratic, and the mean equilibrium is p̄ = (a + D + Σcᵢ/(rᵢ+r̄ᵢ))/(1 + Σ1/(rᵢ+r̄ᵢ)).
• Thm 19. The coalitional-MFG equilibrium does not depend on r̄ₖ. Table IV: the “price of simplicity” is ½[1/(rᵢ+r̄ᵢ) − (rᵢ−r̄ᵢ)/rᵢ²](∫yμ − cᵢ)², unbounded under scaling.

Claims

id	claim	strength	support
C1	Real data in e-commerce, power, water, mango supply, hardware, AI funding, onion prices, underwater channels and EV batteries are non-Gaussian	weak	illustrative figures and prose (§II). Most numbers carry no reference; the only reported fit (H ≈ 0.8, W-error ~10⁻⁶) has no method; the data are the author’s startups’, not released
C2	These data follow “Rosenblatt” (or “two/three-mode Rosenblatt”) distributions	unsupported	no fitting procedure, parameters, or comparison with alternatives (stable, other Hermite, mixtures). “Multi-mode Rosenblatt” is undefined
C3	Itô formula for Rosenblatt-driven processes	strong (cited)	Thm 1, proved in [27] (SPA 2022), not here
C4	Optimal ergodic LQ gain and cost under Rosenblatt noise (linear feedback)	strong	proof sketch in text + [43]; I verified the gain, stability and H → ½ limit
C5	Optimising against Brownian, fBM, Gauss–Volterra or Poisson surrogates gives suboptimal control / non-equilibrium	false for fBM with matching H; true for mismatched covariance	assertion + Figs 16–17 (captions only). Contradicted by the paper’s own cost formula, which depends only on H and the covariance
C6	Optimal MSE predictor is x(t′) (Thm 2)	false as stated	contradicts the drift term and Thm 3
C7	Rosenblatt ARE link: (1−H)(b₂²/r)P² + b₁P − Hq = 0, K̂ = (b₂/r)P	wrong sign	my check: it gives the destabilising gain
C8	MFTG ≠ multi-population coalitional MFG; freezing mean-field terms can be arbitrarily suboptimal	moderate	explicit static computation (Table IV, Thm 19); a scaling argument for unboundedness
C9	Rosenblatt “super-diffusion transformers” significantly enhance generative models	unsupported	§XIV-C assertion; no model, training or experiment anywhere
C10	The paper demonstrates turnpike properties	unsupported	contributions list only; absent from the body
C11	Adaptive control under Rosenblatt noise	absent	§VII is an empty heading

Method

Direct method: restrict strategies to a parametric linear class, compute the long-run cost in closed form via the Rosenblatt Itô formula (in practice, a second-moment computation), and minimise over the gains. Games are handled by best-response fixed points in the same class. The Cournot MFTG uses an orthogonal decomposition into a zero-mean fluctuation game (stochastic) and a mean game (deterministic). There is no numerical method, no estimation and no learning.

Concepts

• Rosenblatt process R_H: a second-order Hermite process. It is non-Gaussian and self-similar with index H ∈ (½, 1), has fBM’s covariance, has stationary increments, and is not a semimartingale.
• Mean-field-type game (MFTG): payoffs and dynamics depend on the players’ own state and action distributions, possibly non-linearly (variance, quantiles). The number of players can be finite. Contrasted with MFGs, where individual players are non-atomic and payoffs are linear in the individual mean field.
• “Price of simplicity”: the payoff gap from freezing mean-field terms (Table IV).
• “Two-mode / three-mode Rosenblatt shape”: used throughout §II and never defined.

Connections

• Control results. These come from the authors’ own earlier papers: Čoupek–Duncan–Pasik-Duncan, “A stochastic calculus for Rosenblatt processes” (SPA 2022, [27]); Čoupek–Duncan–Maslowski–Pasik-Duncan (CDC 2018, [43]); Duncan–Pasik-Duncan (CDC 2020, [44]; CoDIT 2024, [41]; JSSC 2025, [42]); Duncan–Pasik-Duncan–Tembine (CoDIT 2024, [45]; CDC 2024, [46]).
• Rosenblatt process. Rosenblatt (1961); Taqqu (1975, 1979); Dobrushin–Major (1979).
• Real-data non-Gaussianity in control loops. Domański (2015, [23]).
• Diffusion models. §XIV cites no diffusion-model or transformer literature at all: no score-based, DDPM or DiT references.

Bearing on the record

• No THEORY document should cite it. The seed’s summary sentence (the “noise approximation cascade”) should not enter the record as a finding, since it is false for the fBM surrogate (C5).
• For ML practice: none. The “super-diffusion transformers” claim must not be cited as evidence for heavy-tailed or long-memory noise in diffusion models. There is no model or experiment (C9). If the Anthology of the SOTA ever files a practice or theory on non-Gaussian or fractional noise in diffusion models, it needs a paper that trains and evaluates such a model; this one does not.
• The LIT-050 summary should be rewritten to reflect the Rejected standing and the covariance-only point.

Limitations

As stated by the authors:

• “the extension of a bigger class of strategies is still an open issue” (p. 30);
• non-linear strategies are open (p. 31);
• infinite horizon is used “in order to get explicit representation” and is not favoured in practice (p. 37).

Not stated by the authors:

• restriction to linear feedback makes every result covariance-only;
• no data, code or fits are released for §II;
• conflict of interest (data from the first author’s companies);
• missing sections (§VII) and unproved theorems (Thm 6);
• the ML section has no content that could be evaluated.

Open questions

These are the questions the paper raises but does not answer.

• Does non-Gaussianity of Rosenblatt noise change optimal control when strategies may be non-linear, or costs non-quadratic in the state or risk-sensitive (e.g. exponential-of-integral)? A result showing an optimal non-linear feedback beating the best linear one under R_H, but not under fBM with the same H, would be the first evidence that the non-Gaussian part matters for control.
• Is any of the §II data actually better described by a Rosenblatt marginal than by standard heavy-tailed or mixture alternatives? A released dataset with a likelihood or information-criterion comparison would settle it.
• Do diffusion models driven by fractional or Hermite-process noise improve sample quality on a standard benchmark? That needs a trained model and a metric, neither of which is here.

Corrections to the seeded skim

• The seed’s summary sentence is the paper’s own claim, and it is wrong in the form stated. “Replacing Rosenblatt noise … with Brownian, fractional-Brownian or jump noise before optimising yields suboptimal controllers.” For the fBM part:

  • Every optimisation in the paper is over linear (state- or mean-field-type) feedback of a linear system with quadratic or mean-only costs. Under linear feedback, x(t) = e^{at}x₀ + ∫₀ᵗ e^{a(t−s)}dR_H(s) is a Wiener-type integral of a deterministic kernel, so every cost term is a second moment or a deterministic mean.
  • The paper itself says R_H “shares” fBM’s covariance ½(t^{2H} + s^{2H} − |t−s|^{2H}) (§III). Its variance formula in §XIV-C is literally the fBM one: H(2H−1)∫∫e^{θt′}e^{θt″}|t′−t″|
  • Hence Thm 4’s cost L∞(K) = Γ(2H+1)(q + rK²)/(2[−(b₁+b₂K)]^{2H}) is exactly the stationary-variance formula of an fBM-driven Ornstein–Uhlenbeck process. The optimal gain for fBM with the same H is the same.

  The “∃ε > 0: |C_surrogate − C_true| ≥ ε” claim (§VI-A, Figs 16–17) is true only for surrogates with a different covariance: Brownian motion (H = ½), a different H, or a different scale. The same objection applies to the repeated claims for the variance-aware control (§VIII-A), the zero-sum (§IX-A) and non-zero-sum (§X-A) games, the Remarks after Thm 10 (MFTG) and Table III. No figure data or derivation for the “cascade failure” is given beyond Figs 16–17’s captions.

• The seed read the ML section only in part; the full text confirms and strengthens its flag. §XIV (pp. 41–44), “Rosenblatt Foundational Diffusion Models”, contains:

  • (A) an OU forward process under Brownian motion, with a “reverse” process whose drift uses the known clean signal x₀ (a bridge to a given point, which trivially recovers x₀, not a generative model);
  • (B) an fBM version with a stated reverse-time SDE driven by a “time-reversed fractional Brownian motion”, given without derivation, and an “exact” Gaussian score valid only for a single deterministic x₀;
  • (C) the Rosenblatt OU SDE, its H → 1 limit and its variance.

  There is no transformer, no network, no training objective, no dataset, no sample, no metric and no comparison. The sentence “This approach incorporates longer-term context-awareness and non-Gaussianity which significantly enhances the performance of generative machine intelligence models” (§XIV-C) is an assertion. The variance formula offered in support is the fBM one, so it cannot show non-Gaussian benefit. The Index Terms list “Rosenblatt super diffusion transformers”, and the contributions paragraph says the R-OU framework “underpins the design of super diffusion transformers, offering a rigorous foundation for advancements in machine intelligence” (p. 3). Neither is supported.

• The seed says the data section shows no formal fit “in the parts I read”. There is none anywhere. §II (pp. 3–21) calls distributions “two-mode Rosenblatt”, “three-mode Rosenblatt” or “Rosenblatt-like” throughout without any estimation procedure, parameters or goodness-of-fit, with one exception. The underwater ROV experiment on the Niger river (p. 19) reports “estimated Hurst parameter H ≈ 0.8” and “a Wasserstein error on the order of 10⁻⁶”. That comes from 600 samples (1 s over 10 min), with no units, scale, method or comparison model.

  • The marginal of the Rosenblatt process is a single one-parameter (H) law up to scale. I believe it is unimodal, since it is self-decomposable (Maejima–Tudor 2013; I did not re-check this here), so “two-mode” and “three-mode Rosenblatt shape” can only mean unspecified mixtures.
  • Several quantitative claims carry no reference:
    • “Gaussian models underestimate tail risk by 60%” (Fig. 1);
    • “Traditional Gaussian-based models mispredicted grid stress events by 40%” (Fig. 4);
    • Punjab demand “20 BCM” vs “35 BCM”, “12,000 farmers” (2021), “$500 million in losses in 2022”;
    • “A 2023 and 2024 analysis of Germany’s power grid found that net load … followed a two-mode Rosenblatt distribution”.
  • The mango, shallot/onion, underwater and SoH data come from Guinaga and Timadie, which the author biography says the first author founded or co-founded. The data are not released.

• Other errors the seed did not catch:

  • Theorem 2 is false as stated. It says the optimal mean-square predictor of x(t), for dx = b₁x dt + dR_H, given the past up to t′, is x(t′). That fails whenever b₁ ≠ 0 (even for martingale noise the drift gives e^{b₁(t−t′)}x(t′)), and R_H is not a martingale. Theorem 3, on the next page, gives a different (memory-dependent) optimal linear predictor.
  • The Riccati link after Thm 4 has a sign error. “(1−H)(b₂²/r)P² + b₁P − Hq = 0, K̂ = (b₂/r)P” with P the positive root gives the destabilising root: for b₁ = 0.7, b₂ = 1.3, q = 2, r = 0.5, H = 0.75 it gives K = 2.55 and a closed-loop pole at +4.02. The correct form is (1−H)(b₂²/r)P² − b₁P − Hq = 0 with K̂ = −(b₂/r)P, which reproduces Thm 4’s K̂ = −4.705 (my check). The printed identity “H(q + rK̂²) − (b₁ + b₂K̂) = −rK̂/b₂” should read H(q + rK̂²) = (rK̂/b₂)(b₁ + b₂K̂).
  • Thm 5’s mean gain K̄ = (b₂+b̄₂)q̄/(r̄(b₁+b̄₁)) is stabilising only if b₁ + b̄₁ < 0.* This condition is not stated, though the admissible set requires b₁ + b̄₁ + K̄(b₂+b̄₂) < 0. (I verified K̄* and the mean cost b̄₀²q̄r̄/((b₁+b̄₁)²r̄ + (b₂+b̄₂)²q̄) for the stable case.)
  • Unfinished material:
    • §VII “Adaptive control driven by Rosenblatt noise” is an empty heading;
    • §VI-B’s two extensions state problems ending “(independent ? correlated ?)” with no result;
    • the contributions claim “Turnpike properties”, which appear nowhere in the body;
    • Thm 6 (minimax = maximin) has no proof.
  • Wrong cross-reference: the proof of Thm 4 cites “Theorem 1 and Equation (1)” for the Itô formula, but (1) is the underwater path-loss equation; (3) is meant.
  • Wrong date: Einstein’s Brownian-motion paper is dated 1908 in text and reference [3], but Ann. Phys. 322, 549–560 is 1905.