The arrow of time, black holes, and quantum mixing of large N Yang-Mills theories

The part worth your time is §4. Festuccia and Liu isolate a family of planar diagrams (fig. 1) whose contributions go like (λt/h_c)^d_i with d_i = 3ⁱ − 1, so the radius of convergence in λ collapses as h_c/t at large t. That is an honest, checkable computation, and it makes a sharp point: the planar expansion for real-time correlators has zero radius of convergence in the long-time limit even though the Euclidean expansion is fine. The corollary they draw — that λ→0 and t→∞ do not commute at N=∞ — is the paper's most defensible claim, and it is the one I would cite. §5's physical picture, an O(N) window of exponentially many degenerate free states mixed by any nonzero λ so that Γ ~ O(N) and 'small λ' is not small, is heuristic but coherent, and it is the right kind of heuristic for this problem. §6 is where the arrow-of-time claim actually lives, and it is a model. G_+(ω) = e^βω/2A(ω,μ_β) follows from treating c_i^a as a random unit vector in an O(N) shell and O_ab as a banded sparse random matrix, plus the factorization in Appendix G below (G.1). The stress-test is right that §4 alone gives divergence, not a continuous spectrum. But the paper says so itself — end of §5: the diagram argument 'does not tell us directly whether the spectral functions are continuous or not.' The division of labor is stated, not hidden. The factorization assumption is the one I would push on; it is stronger than the §7 caveat about sparsity, since both factors are determined by the same Hamiltonian. One correction to the reader's report: the classifier says no limitations are stated, which is wrong. §7 says 'our arguments in this paper are qualitative in nature and far from foolproof,' and §5 flags the §4–§6 gap explicitly. The paper is unusually candid about what it has and has not shown. §7's stringy black hole is speculation and labeled as such ('it seems to us the answer is likely to be yes'); do not read it as a result. Citations are clean and standard. This is for anyone working on thermalization, ETH, or information loss in AdS/CFT; it is a good problem-statement paper. Send it to referees — the §4 computation deserves checking and the §6 assumptions deserve a referee's push.

The paper offers two arguments for breakdown of planar perturbation theory. §4 is a genuine Feynman-diagram computation: the family in fig. 1 gives contributions ~ (λt/h_c)^{d_i} with d_i = 3^i − 1, a lacunary series with radius λ_c(t) ~ h_c/t → 0. This is a real partial result, but it establishes only that the perturbative series diverges at long times; it does not establish a continuous spectral density or decay of G_+(t). A divergent perturbative series is equally consistent with a discrete spectrum reorganized non-perturbatively. The arrow-of-time conclusion therefore rests on §6, where G_+(ω) = e^{βω/2} A(ω, μ_β) is derived from the assumptions that (i) c_i^a is a random unit vector spread over an O(N) window and (ii) O_ab is a banded sparse random matrix, and crucially (iii) |c_i^a|^2 and ρ_k(a) are independent so their product factorizes (Appendix G, below eq. G.1). Assumption (iii) is stronger than the §7 caveat about c_i^a being 'very sparse': it asserts statistical independence of two quantities both determined by the same Hamiltonian. The authors state (§5, §7) that the §4–§6 connection is unclear and that the arguments are 'qualitative... far from foolproof.' So the headline claim is supported by a model, not a derivation, and the paper's own more rigorous result does not reach it.

The reader's weakest_assumption correctly identifies the §6 random-matrix modeling as the load-bearing vulnerability, and the authors acknowledge a version of it in §7. I agree with the diagnosis and the CONDITIONAL verdict. My sharpening is that the vulnerability is structural, not merely a caveat: the paper's two arguments are logically independent, and only §6 delivers the arrow of time. §4 shows the perturbative series diverges, which is necessary but not sufficient for a continuous spectrum; a divergent series can coexist with a discrete spectrum. The authors themselves say the relation between the two arguments 'is not clear at this point' (§5) and that the arguments are 'qualitative in nature and far from foolproof' (§7). The factorization step in Appendix G (independence of |c_i^a|^2 and ρ_k(a)) is a stronger assumption than the §7 caveat, since both quantities are functionals of the same Hamiltonian. None of this makes the paper wrong — it is a candid, well-motivated position paper with a real partial computation — but it means the headline claim is a model-dependent expectation rather than a derived result. CONDITIONAL remains the right verdict; the concrete test above would either support the modeling assumption or undercut it.

If correct, the central claim would give a microscopic, weak-coupling account of how an arrow of time and an effective horizon emerge in the large-N limit of a bounded gauge theory, which is directly relevant to the information-loss debate and to the interpretation of black-hole singularities in AdS/CFT. The paper is candid about its heuristic status and frames the question sharply. Its most durable contribution is the explicit diagrammatic computation of §4: the family in fig. 1 with d_i = 3^i - 1 and the recursion (4.16) is a concrete, checkable partial result showing that an otherwise convergent planar expansion can diverge in the long-time limit. The statistical modeling of §6 is a legitimate and clearly labeled tool, and the AdS/CFT parameter translation in Appendix A is standard and correctly attributed. The paper does not ship code or machine-checked proofs, and its falsifiable content is qualitative rather than numerical.

Take a gauge theory whose gravity dual is a black hole, and ask where the black hole's one-way-in-time behaviour actually comes from. The paper argues that the answer is the large-N limit itself: at any nonzero coupling an exponentially large number of free-theory states, spread over an energy range of order $N$, mix under the interaction, so the perturbative expansion has a radius of convergence that shrinks to zero as time grows. The authors first show by an explicit planar diagram computation that at any finite order real-time correlators stay quasi-periodic and never thermalize, then argue that the series fails at late times, and finally replace perturbation theory with a statistical model of the mixing that yields a continuous spectral density and decay. If the argument holds, thermalization, information loss, and the horizon of the dual geometry are all consequences of taking $N$ to infinity rather than of strong coupling, and the bulk object dual to the hot phase deserves the name stringy black hole at weak coupling too.

What carries the argument. The load-bearing object is the Lehmann spectral decomposition $G_+(t) = Z^{-1} sum_{i,j} e^{-beta E_i + i(E_i - E_j)t} |langle i|O|jrangle|^2$, read as a statement about how exact energy eigenstates are spread over free-theory states. Two quantities carry the argument: the spread $Gamma_a^2 = sum_{b neq a} |langle a|V|b rangle|^2$, which is of order $N^2$ for any single-trace interaction and therefore gives an energy width of order $N$, and the overlap coefficients $c_i^a$ treated as a random unit vector inside that shell. Combined with the assumption that the operator matrix $O_{ab}$ is banded and sparse, this yields the averaged matrix element $rho_{E_1E_2} = e^{-S(E)} A(omega;E)$ and hence a continuous spectrum. The diagrammatic side of the argument is carried by the iterated graph of Fig. 1, whose contribution is computed exactly in its highest power of Euclidean time using the convolution identity for the thermal propagator $g_E(tau;omega_0)$, and whose recursion $F_{i+1} = F_i^3 alpha d_{i+1}(d_{i+1}-1)$ produces the $1/t$ radius of convergence.

Core claim. The central claim is that the high-temperature phase of a large-$N$ matrix quantum mechanics (with $N=4$ super-Yang-Mills on a sphere as the flagship example) is intrinsically non-perturbative in $N$ at any nonzero 't Hooft coupling $lambda$, and that this is what produces an arrow of time. The chain runs in three steps. First, a perturbative computation shows that at every finite order the thermal Wightman function is a finite sum of delta functions at multiples of the fundamental frequency, hence quasi-periodic, so nothing thermalizes. Second, a family of iterated planar diagrams (Fig. 1, appearing at orders $d_i = 3^i - 1$) produces terms growing like $(t/h_c)^{d_i}$ with $h_c = e^{3/2}(2omega_0)^2/sqrt(3 f(1+f))$, giving a convergence radius $lambda_c(t) ~ 1/t$ that vanishes as $t to infinity$; in frequency space the same diagrams grow like $d_i!$, producing an essential singularity at $lambda = 0$. Third, the physical reason is that free-theory levels at energy of order $N^2$ are degenerate to order $e^{O(N^2)}$, and the interaction mixes states across an energy window of order $N$, so the $lambda to 0$ and $t to infinity$ limits do not commute. A statistical treatment then gives $rho_{E_1 E_2} = e^{-S(E)} A(omega; E)$ with $A$ even in $omega$, falling off like $e^{-beta(E)|omega|/2}$, and at most integrably singular on the real axis, so $G_+(t) to 0$ and the spectral density is continuous for all $omega$.

The premise it rests on. The statistical argument assumes that for each exact energy eigenstate the overlaps with free-theory states form a random unit vector spread over a shell of width of order $N$, and that the operator matrix elements form a banded sparse random matrix; the authors themselves note in section 7 that if those overlaps are structured, for instance very sparse within their spread, the argument fails.

Real-time correlation functions of small gauge-invariant operators decay to zero at late times for any nonzero 't Hooft coupling, so linear response to a perturbation dies away and the system has an arrow of time in the large-$N$ limit.

There is a large-$N$ phase transition at $lambda = 0$ that is invisible in Euclidean quantities such as the partition function, but shows up as an essential singularity of real-time correlators and their Fourier transforms.

Information loss in gravitational collapse is a consequence of the classical (large-$N$) approximation rather than a property of the full quantum theory, since at finite $N$ the spectrum is discrete and evolution is unitary and recurrent.

The bulk configuration dual to the weakly coupled high-temperature phase should be called a stringy black hole: it carries entropy of order $1/g_s^2$, absorbs all probes, has a horizon, and is likely to have singularities.

Because the mechanism only uses the $N$-scaling of single-trace interactions, the same conclusion should apply to any of the matrix quantum mechanics theories in the paper's class, not only to $N=4$ super-Yang-Mills.

If the mixing picture is right, the relevant time scale for the onset of thermal behaviour should be set by the inverse width of the mixed shell, of order $1/N$ in units of the fundamental frequency, which is far shorter than the recurrence time $e^{O(N^2)}$; that gap is what makes the large-$N$ limit look irreversible.

The paper's remark that a stronger form $rho_{ij} = e^{-S(E)} A(omega;E) R_{ij}$ with random $R_{ij}$ is the hallmark of quantum chaos suggests a concrete numerical test: diagonalize a truncated two-matrix model at moderate $N$ and check whether the off-diagonal matrix elements of a simple operator follow such a distribution.

The argument implies that the weak-coupling and strong-coupling analytic structures should be continuously connected in $lambda$ absent a further phase transition, so features attributed to black hole singularities and bulk-cone singularities at strong coupling should leave traces at small $lambda$ as well.

Because the mechanism relies on an energy window of order $N$ opening up, it should not operate in the low-temperature phase, which is consistent with the smooth interpolation between free theory and strong coupling reported for the low-lying spectrum.

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review