A desk cannot trade on Rosenzweig's 4.21-tick boundary yet. The paper has a useful premise: reinforcement-learning agents in an order book can react to an incoming order, where tick replay cannot and a conventional impact model must assume the reaction. Rosenzweig compares historical backtesting to studying old chess games and an agentic book to playing an engine. The analogy works. The execution rule drawn from it asks more of this simulated market than the evidence can bear.
The book and its three phases
Rosenzweig uses a discrete-time continuous double auction with three trader types. One Market Agent submits random market buys and sells, one trade per timer callback, with a 50-share cap. n Liquidity Providers post and revise limit orders. They are trained to maximise PnL over a fixed horizon, see ±d ticks of the book, trade up to 50 shares, and unwind through market orders beyond an inventory limit of ±L. Scalpers use the same design but see only the touch and next level (d = 2). Each trades up to 10 shares; their count is max(3, ⌊n/3⌋). The minimum of three matters: the paper reports that with 0 to 2 Scalpers, the touch forms inconsistently and the book collapses.
Agent parameters stay identical within each class as the experiment varies n and d and scales the Scalper count with n. Each configuration gets 10,000 Monte Carlo paths of 100,000 callbacks. The outputs are mid-price volatility σ, measured in ticks, and collapse probability p: the chance that one side empties before a path ends. Below a threshold n_c, p exceeds 90% in the Collapsed phase. Below d_c(n), the Frozen phase has σ far under one tick as agents settle into queue positions. The Continuous phase between those regions allows ordinary price discovery.
For the impact experiment, Rosenzweig inserts one aggressive buy of Q = 200 at callback 10 and compares those paths with Q = 0 controls. The reported z-score is (μ_I − μ_N)/√(σ_I σ_N). Within the Continuous phase, impact fades at σ = 3.232 ticks (n = 11, d = 38), is "perfectly permanent" at σ = 4.2105 (n = 20, d = 38), and becomes a self-sustaining cascade at σ = 5.529 (n = 23, d = 30).
What did the phase diagram establish?
The phase measurements and the thermodynamic language deserve separate readings. Rosenzweig reports p > 90% at collapse and σ ≪ 1 tick in the frozen region. The paper also says n_c is constant in d and d_c ∝ n. It then treats flow as temperature, n as heat capacity, and d as volume or pressure, calling the boundaries "sharp first-order phase transitions". We did not find a discontinuous order parameter, a hysteresis check or a finite-size study to support the "first-order" label.
Read the thermodynamics as vocabulary.
The measured behaviour still matters. Depth does not shift the reported collapse boundary, and adding Liquidity Providers can make a frozen book quieter. At d = 10, both Frozen-phase impact examples are already frozen; increasing n from 14 to 29 takes σ from 0.432 to 0.0051 ticks. We did not find a numerical n_c or a slope for d_c ∝ n in the text, however. Both claims are read from Figure 2, without a stated fit or error bars.
The paper supplies a further limit on the diagram: phase behaviour "depends on the aggressive flow being drift-free". Significantly biased flow brings the book into the Collapsed phase in finite time, even when n > n_c. Here external aggressive flow comes only from one Market Agent trading at random. Rosenzweig also calls the homogeneous population "clearly unrealistic". Homogeneity lets the paper map its phase space with just n and d, but those are the same quantities its execution advice would have desks estimate.
One order size, three impact profiles
Rosenzweig writes that impact "deviates from classical square-root dynamics, exhibiting distinct dissipative, balanced, and non-dissipative regimes". Three configurations illustrate the latter distinction. Permanent impact and cascades are time profiles; square-root scaling concerns order size. Every impact run uses Q = 200, leaving size scaling unmeasured. The discussion cites Kyle (1985) for square-root impact, although Kyle's model is linear.
The proposed boundary has one illustrated configuration at σ = 4.2105. We found no finer sweep locating it. Nor does the cascading example keep depth fixed: relative to the "balanced" panel, the "non-dissipative" panel changes d from 38 to 30 and n from 20 to 23. Volatility changes alongside them. The z-score's denominator is the geometric mean of the ensemble standard deviations, which remains flat as more paths are added; z(t) therefore measures effect size. The impact section does not restate a path count (the phase sweep used 10,000 per configuration), and its figures display the first 100.
The Frozen-phase comparison is similarly sparse. "Cracking" (σ = 0.0051, n = 29) shifts a few paths, whereas "Melting" (σ = 0.432, n = 14) unfreezes many. Those two examples put the split at roughly 0.1 ticks. Rosenzweig says the profiles mirror Bouchaud et al.'s empirical impact figure; we found no quantitative comparison. The conclusion says the results validate the simulation approach, while its only live-market comparison is the resemblance to Bouchaud et al. in Figure 4.
Beyond the simulated book
We could not test these results. They depend on trained agents and a reactive matching engine, and we found no mention of released code or trained agents. Nor did we find an RL algorithm, reward details or a training procedure specified well enough to rebuild them. A simpler simulator we made would address another model. Our US equities data ends at one-minute bars; it cannot show n, d or how the book responds to an injected order.
Execution guidance would require order-level data, including queue positions, and estimates of n_c and the d_c slope with uncertainty. Until then, "execution algorithms must dynamically assess local market phase state (n, d)" asks desks for two quantities no live venue exposes.
The premise merits another test. Hold d at 38, sweep Q from 25 to 800 on both sides of σ = 4.21, and use heterogeneous agents. If the dissipative-to-cascading split survives, it becomes a result a desk can price.