Arxiv:2010.10545V2 [Cond-Mat.Str-El] 24 Feb 2021 Without Quasiparticle Excitations

Dirac Fast Scramblers

Jaewon Kim,1, ∗ Ehud Altman,1, 2 and Xiangyu Cao1, 3 1Department of Physics, University of California, Berkeley, CA 94720, USA 2Materials Science Division, Lawrence Berkeley National Laboratory, Berkeley, California 94720, USA 3Laboratoire de Physique de l’Ecole Normale Supérieure, ENS, UniversitéPSL, CNRS, Sorbonne Université,Universitéde Paris, 75005 Paris, France We introduce a family of Gross-Neveu-Yukawa models with a large number of fermion and boson flavors as higher dimensional generalizations of the Sachdev-Ye-Kitaev model. The models may be derived from local lattice couplings and give rise to Lorentz invariant critical solutions in 1+1 and 2+1 dimensions. These solutions imply anomalous dimensions of both bosons and fermions tuned by the number ratio of boson to fermion flavors. In 1+1 dimension the solution represents a stable critical phase, while in 2+1 dimension it governs a quantum phase transition. We compute the out of time order correlators in the 1+1 dimensional model, showing that it exhibits growth with the maximal Lyapunov exponent λL = 2πT in the low temperature limit.

Introduction. Since ’t Hooft’s seminal work [1], large dots, which leads to a weakly interacting fixed point, e.g., N quantum field theories — those with a large number of a Fermi liquid. On the other hand, coupling the dots local degrees of freedom — have played a pivotal role in through four-fermion interactions leads to local quantum understanding strongly coupled states of matter and their criticality with no higher dimensional scale invariance. In holographic correspondence to gravity [2–4]. Much of the attempt to avoid this fate, some authors considered field recent progress in this area was enabled by analysis of a theories with non local couplings and thus not clearly simple large N system: the Sachdev-Ye-Kitaev (SYK) realizable by a local microscopic Hamiltonian: for exam- model [5–8]. In its simplest version the SYK model can ple, interactions with a “low-momentum filter” [15] or be written as a quantum mechanical Hamiltonian of N a topological kinetic term [17]. To our knowledge the Majorana fermions with all-to-all interactions: only local higher-dimensional theory exhibiting scale invariance, Lorentz symmetry and fast scrambling is the N X supersymmetric (1+1)-d model [16, 27]. Finding such H = Jijk`χiχjχkχ` . (1) a model without SUSY or beyond (1+1)-d remains an ijk`=1 unrealized goal.

Jijk` are random coupling constants. This model is solv- In this Letter, we propose a family of solvable models able in the large N limit and has remarkable low energy in (1 + 1)-d and (2 + 1)-d that extend the SYK physics properties: as a field theory, it is (nearly) conformal in- to higher dimensions. By doing so, we also make con- variant, and related to 2d dilaton gravity [9–11]; from nection to the well-known Gross-Neveu-Yukawa (GNY) a condensed matter point of view, it exemplifies a non- theory [28, 29]. Specifically, we consider a variant of large Fermi liquid, with no well-defined quasiparticles [5, 12]. N GNY which has a large number of real bosons φa, a = Finally, from a quantum information perspective, it is a 1,...,M, and massless Dirac fermions ψi, i = 1,...,N. fast scrambler, saturating general bounds on the rate of They interact via a local random Yukawa coupling. The quantum information scrambling [6, 13]. Lagrangian in d + 1 dimensional Euclidean space-time is The SYK model is a 0 + 1 dimensional quantum field theory. A natural and much-pursued problem is to find N M 2 X ¯ X m 2 X a ¯ = ψi∂ψ/ i + φ + g ψiφaψj . (2) generalizations to nonzero spatial dimensions [14–26]. L 2 a ij One motivation for this effort is to establish concrete re- i=1 a=1 ija alizations of the AdS/CFT correspondence with richer µ ¯ † 0 0 d and more realistic gravity duals. Another motivation, Here, ∂/ = γ ∂µ, ψ = ψ γ , and γ ,... γ are gamma µ ν matrices satisfying the Clifford algebra γ , γ = δµν . from the perspective of condensed matter physics, is to { } develop controlled theories for quantum critical points The Yukawa interaction has random (but translation invariant) coefficients ga , which are zero-mean complex arXiv:2010.10545v2 [cond-mat.str-el] 24 Feb 2021 without quasiparticle excitations. ij Most of the generalizations of the SYK model that Gaussian variables satisfying have been discussed before, consist of a lattice of coupled a b 2 2 SYK quantum-dots. The problem with this approach is gijgk` = g δi`δjkδab/N (no sum) . (3) that the SYK interaction (1) between the internal degrees of freedom of the dots is irrelevant (at most marginal) Here g2 is a coupling constant, which can be positive or compared to quadratic couplings (hopping) between the negative (g iR). We work in a large N limit where the boson/fermion∈ number ratio tends to a constant:

M ∗ On military service for the Republic of Korea. γ (M,N ) , (4) NnS −→ −→ ∞ 2 where nS is the number of components of each spinor. on the two sub-lattices A and B with the linearized dis- x y This is the main difference from the usual large N limit persion near the Dirac point K, HK+k = kxσ + kyσ . of GNY, where the boson number remains finite. We The gamma matrices in this case are taken as will show that, in d + 1 < 4 space-time dimensions, our theory admits a family of Lorentz invariant critical solu- γ0 = σz , γ0γ1 = iσx , γ0γ2 = iσy . tions, whose critical exponents depend continuously on γ. Moreover, we show that the (1 + 1)-d critical points A boson field that couples the two bands in the same are fast scramblers. valley K is constructed from the lattice bosons as φa = A B (ϕ ϕ ) k→0, which is odd under the sub-lattice sym- Lattice models. Before analyzing the field theory, let us a − a | discuss its lattice realizations. In (0 + 1)-d, the spinor metry. Our critical theory then describes a phase tran- ψ can have only one component (γ0 = 1), and thus (2) sition to a phase with spontaneous breaking of the sub- describes a “low-rank” SYK dot [30–36] (if γ = 2, the lattice symmetry. = 1 supersymmetric SYK [37]). Indeed, integrating We note that the choice of lattice realization is not Nout the bosons leads to a Hamiltonian unique. For example, we could take the fermion fields in the (2+1)-d model to be a 4-component Nambu spinor N M A B B † A † T 1 X † † 1 X a a (ψ , ψ , (ψ ) , (ψ ) ) coupling to charge-2 bosonic H = Jij,k`c cjc c` ,Jij,k` = g g . (5) − −2 i k m2 ij k` fields. In this case the theory (2) describes a super- ij,k` a=1 conducting quantum phase transition. With a different choice of 4-component spinor and bosons carrying an an- As an N 2 N 2 matrix, the rank of J is at most M = gular momentum quantum number, the theory can de- γN N,× instead of N 2 in the standard (complex) scribe spontaneous breaking of time reversal symmetry SYK∝. The low-energy∝ states of this model differ from 4 and establishment of a quantum Hall state. that of SYK in that the fermion scaling dimension is 4 Critical solutions. We now return to the field theory in not fixed to 1/4, but rather can be tuned continuously general dimension and present its critical solutions in the by varying γ. In this sense the model is similar to SYK q IR. For generality, we shall add a boson kinetic term to (SYK with q fermion interaction), for q (2, ), q = 4. the Lagrangian (2), Like the SYK models, these are fast scramblers∈ ∞ and have6 a residual entropy. The higher-dimensional solutions we 2 2 2 2 2 m φ m φ + b(∂φa) , b 0 . (7) shall present are natural generalizations of these states. a a ≥ ; In nonzero spatial dimensions, the field theory (2) can In the large N limit, the fermion and boson Green func- be obtained as the long-wavelength limit of a lattice 1 P ¯ 1 P tions G = N i ψiψi ,F = M a φaφa , averaged of identical low-rank SYK dots connected by nearest- over the random couplings, are relatedh to theiri respec- neighbor hopping: tive self-energy Σ, Π by a set of Schwinger-Dyson (SD) equations: X † 1 X † † H = ci,x0 ci,x Jij,k`ci,xcj,xc c`,x − 2 k,x 1 n.n. xijk` G(k)−1 = k/ Σ(k) ,F (q) = . (8) 2 (6) − m2 + bq2 Π(q) H.S. X † X a † X ϕa,x − c 0 cj,x + g c cj,xϕa,x + . 2 2 ∼ i,x ij i,x 2 Σ(x) = nSγg G(x)F (x) , Π(x) = g tr[G(x)G( x)] . n.n. xija ax − − Above, q = (Ω, q) and k = (ω, k) denote (d + 1)- Here c are N flavors of lattice fermions, and ϕ are i,x a,x momenta, and “tr” is over the spinor space. Eqs. (8) introduced in a Hubbard-Stratonovich transformation. generalize the SD equations of the low-rank SYK dots, We note however that our IR solution below also ap- and can be derived following the same steps. plies to physical bosons with an additional kinetic term The main point of this Letter is that, in the IR ((∂φ )2). a limit, the SD equations admit critical solutions that O We illustrate the connection between the microscopic have the following scale-invariant and Lorentz symmetric Hamiltonian and the field theory degrees of freedom in form [38]: two examples. In (1+1)-d, the field theory can be constructed from the slowly varying chiral fermion fields G ik/ k 2∆−D−1 x/ x −2∆−1 near the two Fermi points, making up the Dirac spinor ∼ | | ∼ | | ψ = (ψL, ψR)T . The low energy bosons ϕ that couple Σ ik/ k −2∆+D−1 x/ x 2∆−2D−1 i i i a ∼ | | ∼ | | (9) between the chiral fermions carry momenta near 2kF . F q D−4∆ x 4∆−2D Finally, the gamma matrices operating in this space are ∼ | | ∼ | | Π q 4∆−D x −4∆ , γ0 = σx , γ0γ1 = iσz , ∼ | | ∼ | | for a continuous range of the fermion scaling dimension x,y,z where σ are Pauli matrices. ∆ = ∆ψ (the boson scaling dimension is ∆φ = D 2∆ψ): As a (2+1)-d example, we take N flavors of fermions − hopping on the honeycomb lattice. Then, the two compo- D D + 2 D 1 A B T min , > ∆ > − ,D := d + 1 < 4 . (10) nent spinor ψ = (ψi , ψi ) is constructed from fermions 2 4 2 3

4 1 1 (a) (b) (c) with ∆ ( 4 , 2 ) determined uniquely by γ (see Fig. 1-a). 2 This “self-tuned”∈ criticality was ﬁrst noticed in Ref. [34],

γ 1 see also [41] for a Monte Carlo study. The situation of 1/2 g2 < 0 is even more special [30]: the boson self energy 1/4 diverges Π(q 0) and dominates the mass, and the IR fixed point→ has→ −∞ ∆ (0, 1 ). 0 1/4 1/2 ∆ 1 5/4 4 In (1 + 1)-d, the boson∈ mass flows to zero in the IR 2 FIG. 1. The rank (γ)-exponent (∆) relation (11) in D = 1, 2, 3 provided g > 0. To see why this must happen, let us (a,b,c) space-time dimensions. first show that the bosons cannot remain gapped in the IR. If that were the case, the fermions would be non- interacting in the IR, G ( ik/)−1 (∆ = 1/2). How- ∼ − For D 4, no ∆ satisfies these inequalities. Moreover, ever, the boson self-energy would then have a log diver- 2 2 we have≥ the “rank-exponent relation” (see Fig. 1): gence, and Π(q) g ln(1/ q ) > m at small enough q for any g2 > 0,∼ which makes| | the bosons unstable. B D C2∆ Could the bosons then condense? In this case the con- 2 −∆ γ = , (11) densate F (τ) const would generate a mass in the −BD−∆C2∆− D 2 fermion dispersion∼ leading to G ik// k (∆ = 1). How- ∼ | | where B and C are defined as ever, this would imply an inconsistent IR divergence F (x) ln ln(R/ x ). Since neither conventional state D −a D ∼ | | D 2 2 Γ( a) C 1 leads to a self consistent solution, the IR fixed point must 2 a− 2 C = (2π) 2 − ,B = . (12) a 2aΓ(a) a 1 2a be critical. [A similar argument can be applied to under- − stand the self-tuned criticality in (0+1)-d.] It should be The rank-exponent relation generalizes that in (0 + 1)-d, noted that, since the back-scattering is relevant, our crit- which was known previously [30–35, 39]. ical solutions are not Luttinger liquids. To demonstrate the above claims, suppose, as will be Another peculiar aspect of (1 + 1)-d is that the critical 1 justifed below, that we can ignore the bare terms k/ and solutions we found require a nonzero minimal rank γ > 4 . m2 + bq2 in the IR limit. Then the approximate SD Indeed, the rank-exponent relation reads (Fig. 1-b) equations become scale invariant. One may start from a 3 2∆ power-law Ansatz γ = − . (16) 8∆ 4 2∆−D−1 − G(k) = iAk/ k , (13) 1 | | The interval ∆ ( 2 , 1) permitted by (10) is mapped to 1 ∈ and check that it is compatible with the SD equations (8) γ ( 4 , ), with the limit γ 1/4 corresponding to if the exponent is fixed by (11). This can be done by using ∈ ∞ → 1 ∆ 1. The critical points for γ 4 are not included in the gamma-matrix identity a/a/ = a 2I and the Fourier our→ solutions; we speculate that the≤ fermion Green func- transform identities | | tion still has ∆ = 1, but with nontrivial log-corrections. We remark finally that, our theory at γ = 1/2 is akin to Z µ −2a iq xµ D 2a−D x e d x = Ca q , (14) the = 1 SUSY theory of Ref. [16] with q = 3: they | | | | N both have ∆φ = 1/3. Z µ x x −2a−1eiq xµ dDx = iB q q 2a−D−1 . (15) In (2 + 1)-d, the critical points are not self-tuned, but / a/ 2 2 | | − | | describe a second-order transition at g = gc > 0, between a semi-metal with gapped bosons and a marginal The value of the prefactor A depends on UV details and Fermi liquid with broken symmetry. Let us also comment is unimportant; it suffices to know that A > 0 from the on the rank-exponent relation (Fig. 1-c), which reads UV limit G(k) k→∞ 1/( ik/). Our neglect of∼ the bare− kinetic terms in the Green’s (2∆ 5)(2∆ 3) tan π∆ tan 2π∆ γ = − − . (17) function is justified if they are irrelevant in the IR, that is 8(∆ 1)(4∆ 3) if they vanish in the low energy limit compared to the self − − energy terms. The condition for the fermion kinetic term In the interval ∆ (1, 5 ) allowed by (10), ∆ is uniquely ∈ 4 k Σ(k) is ∆ > (D 1)/2. Similarly the condition determined by γ, and increases from the non-interacting ∼ | | | | − 2 4∆−D for the boson kinetic term bq Π q is ∆ < low-rank limit ∆γ→0 = 1, to the strongly coupled high- ∼ | | 5 (D + 2)/4. Thus, the inequalities (10) emerge as the rank limit ∆γ→∞ = 4 . The latter limit seems to con- condition for a consistent scale invariant solution [40]. tradict the fact that a ψψ¯ ψψ¯ interaction is irrelevant in Finally, the (renormalized) boson mass term is tuned to (2 + 1)-d, and that the high-rank (γ N ) limit ∼ → ∞ zero at criticality, or in the exceptional cases of D = 1, 2 of our model is effectively a lattice of SYK4 dots. There is even irrelevant and flows to zero in the IR, as we discuss is however a subtlety: the quartic couplings Jij,k` in (5) below. above have a variance γN −3, which is much larger 2 −3 ∼ In (0 + 1)-d, when g > 0, the mass is known to be than N in SYK4 if γ N. Hence, our theory in the ∼ irrelevant, and the system always flows to a critical point γ limit is more strongly interacting than a SYK4 → ∞ 4

Nf 2 3 4 8 10 20 are defined on a “double” Keldysh contour (Fig. 2): ∆ [42] 1.067 1.054 1.042 1.021 1.017 1.008 ∆ (17) 1.107 1.063 1.043 1.019 1.015 1.007 β β (t, x) = V †(t, x)V (t + i , x)φq(0, 0)φq(i , 0) , (18) C h 2 2 i TABLE I. Fermion scaling dimension ∆ in the D = 3 GNY with Nf fermion flavors found using conformal bootstrap [42] where φq(t) = φ(t + i) φ(t i), and (V †,V ) can be a compared to the result obtained from the rank-exponent re- spinor component and its− conjugate− ((ψA)†, ψA), or the lation (17) by setting γ = 1/(Nf nS ). boson (φ, φ), which is real. The OTOC measures the sensitivity of an observable V at (t, x) to a perturbation at (0, 0) as a quantum analogue of the butterfly effect; i − it can grow exponentially in large N systems, defining +i t a Lyapunov exponent λL. The (0 + 1)-d version of the model are known to be fast scramblers [31, 32, 35], sat- i(β/2 ) − urating the general bound on the Lyapunov exponent at i(β/2 + ) t + iβ/2 low temperatures [13]: 1 λLt (t) e , where λL = 2πT . (19) C ∼ N FIG. 2. Left: The Keldysh contour on which the OTOC (18) is defined, with field locations indicated. Right: the basic Here, we extend the result to (1+1)-d, by a method sim- rungs that generate all the ladder diagrams. Straight, wavy, ilar to that discussed in Refs. [14, 26, 47]. In the regime and dashed lines correspond to fermion, boson propagators where the OTOC has a well-defined exponential growth, and average over random couplings. it is given, to leading order in 1/N, by the sum of ladder diagrams generated by a four-point retarded kernel K(t1, x1, . . . , t4, x4) that adds a rung to the ladder (See Fig. 2). The self consistency condition (Bethe-Salpeter lattice and only the former can approach the ∆ = 5 fixed 4 equation) is then equivalent to solving for the OTOC as point (unless one goes beyond the large N limit and lets an eigenvector of the kernel with eigenvalue 1: the temperature to scale with 1/N). Z We now make connection to the standard approach p(t3, x3, t4, x4) = K p(t1, x1, t2, x2) (20) to the GNY model, i.e., Nf fermions coupled to a sin- F F gle boson. Within the new large N limit with com- parably many bosons, we found, at the classical saddle To find a Lyapunov exponent we seek exponentially grow- point, an anomalous fermion scaling dimension, which ing eigenfunctions with momentum p: is only obtained as 1/Nf corrections in the standard ap- λL(t1+t2)/2+ip(x1+x2)/2 p(x1, t1, x2, t2) = Fp e (21) proach. Moreover, we can compare quantitatively the F rank-exponent relation (17) in (2 + 1)-d to the standard where Fp = Fp(t1 t2, x1 x2) only depends on the rela- large Nf GNY exponent, by identifying γ = 1/(Nf nS). tive coordinate. As− a result,− we determine a p-dependent Solving for the scaling dimension we obtain ∆ = 1 + exponent λ = λ (p). Then the OTOC in real space can 2 2 L L 4/(3π Nf nS) + (1/N ), which matches the standard O f be expressed as a momentum integral 1/Nf result [43]. We also observe a good agreement with 1 Z ∞ state-of-the-art conformal bootstrap data on the stan- λL(p)t+ipx (t, x) ρ(p)Fp(0, 0)e dp , (22) dard GNY model [42], see Table I. These results indicate C ∼ N −∞ that the 1/N corrections of our large N theory are rather mild. where ρ(p) was shown [47] to have a pole where λL(p = is∗) = 2πT [48]. The integral (22) is analyzed by a steep- Residual entropy. A peculiar feature of SYKq is that, the entropy remains nonzero and N in the zero-T limit, est descent method for large x, t, and has maximal growth ∝ e2πT t if it is dominated by the pole in ρ(p). This is the provided the large N limit is taken first [6, 12, 44]. Such ∝ a residual entropy is also observed numerically in the case when the fields are sufficiently separated in space, x /t > v∗, where the velocity v∗ is defined as D = 1 critical points [35], and we can now understand it | | analytically, in terms of log determinants [45]. In higher v∗ := [∂sλL(is)] . (23) dimensions, only the zero-momentum component of the s=s∗ Green function is critical at zero-temperature and can Here s∗ > 0 is such that λL(is∗) = 2πT . possibly contribute to the residual entropy. Therefore, For our (1 + 1)-d critical solutions, we can calculate the residual entropy is not extensive in volume, in con- λL(p) analytically [45]. To compute the kernel we use trast to “local critical” SYK lattices [14]. the finite-T critical Green’s functions obtained by con- Scrambling. We turn to calculating the out of time order formal invariance from the zero temperature power-laws. correlations (OTOCs) [13, 46] in the low temperature The calculation is simplified by factoring of correlators limit of the 1+1 dimensional field theory. The OTOCs into functions of one chiral variable x t. Now, solving ± 5 the eigenvalue problem, we find that for any ∆ ( 1 , 1), model exhibits maximal scrambling. A natural next step, ∈ 2 λL(p) satisfies the following: beyond the critical saddle point solutions presented in this paper, is to derive a low-energy effective theory of λL(p = i2πT ) = 2πT . (24) the dominant fluctuations in analogy with the Schwarzian ± effective theory of the SYK model [54]. Moreover, we checked that v∗ := [∂sλL(is)]s=2πT < vB = The model introduced here can serve as a new start- 1 always holds (see Figure 1 of [45]). Therefore, there is a ing point for understanding strongly coupled quantum non-empty regime of fast scrambling near the light cone: critical points or phases. In (1+1)-d, we have strongly correlated gapless phases that are holographic CFTs. It x x will be interesting to study the effect of various pertur- (t, x) e2πT (t−|x|) , t | | , | | . (25) C ∼ ∈ vB v∗ bations, such as quenched disorder. The (2+1)-d solution exemplifies a strongly coupled quantum critical point Such a behavior is qualitatively similar to that in other with itinerant fermions. We may add a Gauge field to higher dimensional generalizations of the SYK model obtain a new large N limit of QED3. Finally, our confor- [14, 26, 47, 49, 50]. However, due to lack of Lorentz mal solutions also gives a new paradigm for investigating symmetry, the butterfly velocity vB in those models has how superconductivity can emerge from critical fluctua- a model dependent value (t, x) e2πT (t−|x|/vB ), while tions in absence of quasiparticles: indeed, if the coupling C ∝ a in our case vB = 1 is the speed of light. constants gij in (2) are chosen from the GOE ensemble, a a Our result here is a concrete example of the “chiral” with gij = gji R, instead of GUE as in (3), the theory scrambling modes e2πT (t±x), which are argued to appear allows superconducting∈ solutions described by Eliashberg in generic holographic CFTs [51] as a result of broken equations [33, 34]. reparametrization symmetry. It will be interesting to apply a general theory of scrambling (e.g., [52, 53]) to investigate whether the (2 + 1)-d critical points are fast- ACKNOWLEDGMENTS scrambling. Discussion. We proposed a variant of the GNY theory as We thank Sumilan Banerjee, Jordan Cotler, Yingfei a generalization of the SYK physics to higher dimensions. Gu, Igor Klebanov, Mark Mezei, Douglas Stanford, and The model is solvable in the large N limit and presents Grigory Tarnopolskiy, for comments on the manuscript. a strongly coupled Lorentz invariant critical point. Fur- We acknowledge support from a Department of Energy thermore, direct calculation shows that in (1+1)-d the grant DE-SC0019380 (EA and XC).

[1] Gerard ’t Hooft, “A planar diagram theory for strong (2016), arXiv:1604.07818 [hep-th]. interactions,” Nucl. Phys. B 72, 461 (1974). [9] Kristan Jensen, “Chaos in AdS2 Holography,” Phys. Rev. [2] Juan Maldacena, “The Large N limit of superconformal Lett. 117, 111601 (2016), arXiv:1605.06098 [hep-th]. field theories and supergravity,” Int. J. Theor. Phys. 38, [10] Juan Maldacena, Douglas Stanford, and Zhenbin 1113–1133 (1999), arXiv:hep-th/9711200. Yang, “Conformal symmetry and its breaking in two- [3] Steven S. Gubser, Igor R. Klebanov, and Alexan- dimensional nearly anti-de Sitter space,” Progress of der M. Polyakov, “Gauge theory correlators from noncrit- Theoretical and Experimental Physics 2016 (2016), ical string theory,” Phys. Lett. B 428, 105–114 (1998), 10.1093/ptep/ptw124, 12C104, arXiv:1606.01857. arXiv:hep-th/9802109. [11] Julius Engelsöy, Thomas G. Mertens, and Herman [4] Ofer Aharony, Steven S. Gubser, Juan Maldacena, Hi- Verlinde, “An investigation of AdS2 backreaction and rosi Ooguri, and Yaron Oz, “Large N field theories, holography,” Journal of High Energy Physics 2016, 139 string theory and gravity,” Physics Reports 323, 183 – (2016). 386 (2000). [12] A. Georges, O. Parcollet, and S. Sachdev, “Quan- [5] Subir Sachdev and Jinwu Ye, “Gapless spin fluid ground tum fluctuations of a nearly critical Heisenberg spin state in a random, quantum Heisenberg magnet,” Phys. glass,” Phys. Rev. B 63, 134406 (2001), arXiv:cond- Rev. Lett. 70, 3339 (1993), arXiv:cond-mat/9212030 mat/0009388. [cond-mat]. [13] Juan Maldacena, Stephen H. Shenker, and Douglas [6] Alexei Kitaev, “A simple model of quantum holog- Stanford, “A bound on chaos,” JHEP 08, 106 (2016), raphy,” http://online.kitp.ucsb.edu/online/ arXiv:1503.01409 [hep-th]. entangled15/kitaev/,http://online.kitp.ucsb.edu/ [14] Yingfei Gu, Xiao-Liang Qi, and Douglas Stanford, online/entangled15/kitaev2/. Talks at KITP, April 7, “Local criticality, diffusion and chaos in generalized 2015 and May 27, 2015. Sachdev-Ye-Kitaev models,” JHEP 05, 125 (2017), [7] Subir Sachdev, “Bekenstein-Hawking entropy and arXiv:1609.07832 [hep-th]. strange metals,” Phys. Rev. X 5, 041025 (2015), [15] Micha Berkooz, Prithvi Narayan, Moshe Rozali, and arXiv:1506.05111 [hep-th]. Joan Simón,“Higher dimensional generalizations of the [8] Juan Maldacena and Douglas Stanford, “Remarks on SYK model,” Journal of High Energy Physics 2017, 138 the Sachdev-Ye-Kitaev model,” Phys. Rev. D 94, 106002 (2017), arXiv:1610.02422. 6

[16] Jeff Murugan, Douglas Stanford, and Edward Wit- Rev. Lett. 124, 017002 (2020), arXiv:1904.07240. ten, “More on supersymmetric and 2d analogs of the [35] Jaewon Kim, Xiangyu Cao, and Ehud Altman, “Low- SYK model,” Journal of High Energy Physics 2017, 146 rank Sachdev-Ye-Kitaev models,” Phys. Rev. B 101, (2017), arXiv:1706.05362. 125112 (2020), arXiv:1910.10173. [17] Gustavo J. Turiaci and Herman Verlinde, “Towards a 2d [36] Jaewon Kim, Xiangyu Cao, and Ehud Altman, “Scram- QFT analog of the SYK model,” Journal of High Energy bling versus relaxation in Fermi and non-Fermi liquids,” Physics 2017, 167 (2017), arXiv:1701.00528. Phys. Rev. B 102, 085134 (2020), arXiv:2006.02485. [18] Richard A. Davison, Wenbo Fu, Antoine Georges, Yingfei [37] Wenbo Fu, Davide Gaiotto, Juan Maldacena, and Subir Gu, Kristan Jensen, and Subir Sachdev, “Thermoelec- Sachdev, “Supersymmetric Sachdev-Ye-Kitaev models,” tric transport in disordered metals without quasiparti- Phys. Rev. D 95, 026009 (2017), arXiv:1610.08917. cles: The Sachdev-Ye-Kitaev models and holography,” [38] The anomalous form of the fermion propagator was found Phys. Rev. B 95, 155131 (2017), arXiv:1612.00849. in a GNY theory away from large N limit by an expan- [19] Xue-Yang Song, Chao-Ming Jian, and Leon Balents, sion [55]. “A strongly correlated metal built from Sachdev-Ye- [39] Aavishkar A. Patel and Subir Sachdev, “Critical strange Kitaev models,” Phys. Rev. Lett. 119, 216601 (2017), metal from fluctuating gauge fields in a solvable arXiv:1705.00117. random model,” Phys. Rev. B 98, 125134 (2018), [20] D. V. Khveshchenko, “Thickening and sickening the SYK arXiv:1807.04754. model,” SciPost Phys. 5, 12 (2018). [40] Note that the condition ∆ < D/2 is not needed for this [21] Arijit Haldar, Sumilan Banerjee, and Vijay B. Shenoy, purpose, but is necessary in D = 1 so that F (x) → 0 for “Higher-dimensional sachdev-ye-kitaev non-fermi liquids x → ∞. at lifshitz transitions,” Phys. Rev. B 97, 241106 (2018), [41] Gaopei Pan, Yuxuan Wang, and Zi Yang Meng, “Self- arXiv:1710.00842. tuned quantum criticality and non-Fermi-liquid in a [22] Debanjan Chowdhury, Yochai Werman, Erez Berg, and Yukawa-SYK Model: a Quantum Monte Carlo study,” T. Senthil, “Translationally invariant non-Fermi liquid arXiv:2001.06586 (2020). metals with critical Fermi-surfaces: Solvable models,” [42] Luca Iliesiu, Filip Kos, David Poland, Silviu S. Pufu, Phys. Rev. X 8, 031024 (2018). and David Simmons-Duffin, “Bootstrapping 3d fermions [23] Aavishkar A. Patel, John McGreevy, Daniel P. Arovas, with global symmetries,” Journal of High Energy Physics and Subir Sachdev, “Magnetotransport in a model of 2018, 36 (2018), arXiv:1705.03484. a disordered strange metal,” Phys. Rev. X 8, 021049 [43] Sergey E. Derkachov, N.A. Kivel, A.S. Stepanenko, and (2018), arXiv:1712.05026. A.N. Vasiliev, “On calculation in 1/n expansions of crit- [24] Alexander Altland, Dmitry Bagrets, and Alex ical exponents in the Gross-Neveu model with the con- Kamenev, “Quantum criticality of granular Sachdev-Ye- formal technique,” (1993), arXiv:hep-th/9302034. Kitaev matter,” Phys. Rev. Lett. 123, 106601 (2019), [44] Yingfei Gu, Alexei Kitaev, Subir Sachdev, and Grig- arXiv:1903.09491. ory Tarnopolsky, “Notes on the complex Sachdev-Ye- [25] Aavishkar A. Patel and Subir Sachdev, “Theory of a Kitaev model,” Journal of High Energy Physics 2020, Planckian metal,” Phys. Rev. Lett. 123, 066601 (2019), 157 (2020), arXiv:1910.14099. arXiv:1906.03265. [45] The Supplemental Material details the calculation of [26] Biao Lian, S. L. Sondhi, and Zhenbin Yang, “The chiral scrambling in (1+1)-d and of the residual entropy in SYK model,” Journal of High Energy Physics 2019, 67 (0+1)-d. (2019), 1906.03308. [46] A. I. Larkin and Yu. N. Ovchinnikov, “Quasiclassical [27] MárkMezei and Gábor Sárosi,“Chaos in the butterfly method in the theory of superconductivity,” Soviet Jour- cone,” Journal of High Energy Physics 2020, 186 (2020), nal of Experimental and Theoretical Physics 28, 1200 arXiv:1908.03574. (1969). [28] David J. Gross and AndréNeveu, “Dynamical symmetry [47] Yingfei Gu and Alexei Kitaev, “On the relation between breaking in asymptotically free field theories,” Phys. Rev. the magnitude and exponent of OTOCs,” Journal of High D 10, 3235–3253 (1974). Energy Physics 2019, 75 (2019), arXiv:1812.00120. [29] J. Zinn-Justin, “Four-fermion interaction near four di- [48] The argument in the aforementioned reference was for- mensions,” Nuclear Physics B 367, 105 – 122 (1991). mulated with SYKq ladder kernels, but can be extended [30] Zhen Bi, Chao-Ming Jian, Yi-Zhuang You, Kelly Ann to other ones. Pawlak, and Cenke Xu, “Instability of the non-Fermi [49] Vedika Khemani, David A. Huse, and Adam Nahum, liquid state of the Sachdev-Ye-Kitaev Model,” Phys. Rev. “Velocity-dependent lyapunov exponents in many-body B 95, 205105 (2017), arXiv:1701.07081. quantum, semiclassical, and classical chaos,” Phys. Rev. [31] Wenbo Fu, The Sachdev-Ye-Kitaev model and matter B 98, 144304 (2018), arXiv:1803.05902. without quasiparticles, Ph.D. thesis, Harvard University [50] Haoyu Guo, Yingfei Gu, and Subir Sachdev, “Transport (2018). and chaos in lattice sachdev-ye-kitaev models,” Phys. [32] Eric Marcus and Stefan Vandoren, “A new class of SYK- Rev. B 100, 045140 (2019). like models with maximal chaos,” Journal of High Energy [51] Gustavo J. Turiaci and Herman Verlinde, “On CFT and Physics 2019, 166 (2019), arXiv:1808.01190. quantum chaos,” Journal of High Energy Physics 2016, [33] Ilya Esterlis and JörgSchmalian, “Cooper pairing of in- 110 (2016), arXiv:1603.03020. coherent electrons: An electron-phonon version of the [52] Mike Blake, Hyunseok Lee, and Hong Liu, “A quantum Sachdev-Ye-Kitaev model,” Phys. Rev. B 100, 115132 hydrodynamical description for scrambling and many- (2019), arXiv:1906.04747. body chaos,” Journal of High Energy Physics 2018, 127 [34] Yuxuan Wang, “Solvable strong-coupling quantum-dot (2018), arXiv:1801.00010. model with a non-Fermi-liquid pairing transition,” Phys. [53] Felix M. Haehl, Wyatt Reeves, and Moshe Rozali, 7

“Reparametrization modes, shadow operators, and quan- eterizations for AdS3 gravity,” Journal of High Energy tum chaos in higher-dimensional CFTs,” Journal of High Physics 2019, 79 (2019), arXiv:1808.03263. Energy Physics 2019, 102 (2019), arXiv:1909.05847. [55] Subir Sachdev, “Phase transitions of Dirac fermions,” [54] Jordan Cotler and Kristan Jensen, “A theory of reparam- in Quantum Phase Transitions (Cambridge University Press, 2011) p. 332–345, 2nd ed.

CONTENTS

Acknowledgments 5

References 5 A. Calculation of scrambling in (1+1)-d 7 B. Residual entropy 10

A. Calculation of scrambling in (1+1)-d

In this section we give more details on the calculation of Lyapunov exponent in (1+1)-d. We will first derive real-time, finte-T correlation functions, which we then use to construct the ladder kernels. Finally we exhibit the solution to the eigen-problem. Correlators. In (1 + 1)-d, it is convenient to adopt some notations from 2d conformal field theory. The Euclidean space time is identified with the complex plane z = τ ix , z¯ = τ + ix . (26) − Without loss of generality, we work with the spinor representation

T 0 x 0 1 z ψ = (ψL, ψR) , nS = 2 , γ = σ , γ γ = iσ , (27) mentioned in the main text. Then the fermion Green function in the critical solution has only two nonzero components:

L 0 G ¯ ¯ G = ,GL(z, z¯) = Az−2hz¯−2h ,GR(z, z¯) = Az−2hz¯−2h . (28) GR 0

Here, A is an undetermined prefactor (which will not aﬀect the kernel) h, h¯ are the holomorphic and anti-holomorphic scaling dimesnsions, and are related to ∆ in the main text by the following: ∆ 1 ∆ 1 h = , h¯ = + , ∆ = h + h¯ . (29) 2 − 4 2 4 Recall that the allowed interval is ∆ (1/2, 1), which corresponds to h (0, 1/4). The boson propagator has a similar form ∈ ∈ 8h2 −2hb −2hb ¯ F (z, z¯) = AF z z¯ , hb = 1 h h , AF = . (30) − − A2g2π2

(AF is obtained using the Schwinger-Dyson equations). We obtain ﬁnite-T Euclidean correlation functions using conformal symmetry, by the standard mapping between complex plane and a cylinder z = f(w) = exp(2πT iw), w = τ 0 ix, τ 0 [0, 1/T ). As a result, 1/z, 1/z¯ is replaced by πT/ sin(πT w), πT/ sin(πT w¯), respectively. For example, − ∈

¯ πT 2h πT 2h GL = A , (31) sin(πT w) sin(πT w¯) and similarly for GR and F . To compute scrambling, we need the Wightman and retarded Green functions. They can all be obtained as an appropriate analytical continuation of the Euclidean Green function, τ it (for retarded) and τ it + β/2 (for Wightman). In terms of the light cone variables → ± → u = t x , v = t + x , (32) − 8 we have ¯ ¯ πT 2h πT 2h πT 2h πT 2h GL = A ,GL = 2 sin(2πh)A θ(u) θ(v) (33) W cosh(πT u) cosh(πT v) R sinh(πT u) sinh(πT v) ¯ ¯ πT 2h πT 2h πT 2h πT 2h GR = A ,GR = 2 sin(2πh¯)A θ(u) θ(v) (34) W cosh(πT u) cosh(πT v) R sinh(πT u) sinh(πT v) πT 2hb πT 2hb πT 2hb πT 2hb F = A ,F = 2 sin(2πh )A θ(u) θ(v) . W F cosh(πT u) cosh(πT v) R b F sinh(πT u) sinh(πT v) (35)

The additional prefactor 2 sin(... ) in front of retarded Green functions is due to analytical continuation around a branching singularity. Ladder kernel. The ladder kernel in our model acts on vector-valued functions of the following form

 φ(u1, v1, u2, v2)  F ¯(u1, v1, u2, v2) FLR  (u1, v1, u2, v2) =  ¯ (u1, v1, u2, v2) (36) F FRL   ¯ (u1, v1, u2, v2) FRL ¯ (u1, v1, u2, v2) FLR where the components are indexed the type of particles propagating on the ladder rungs (φ is short-hand for φφ, and L¯ and R¯ are the antiparticle of R and L, respectively). The kernel can then be written as follows: Z du1dv1 du2dv2 (K )(u3, v3, u4, v4) = K(u1, v1, . . . , u4, v4) (u1, u2, v1, v2) , (37) F F 2 2

(note that dudv = 2dtdx so we need to divide by the Jacobian) where K is a 5 5 matrix ×  0 KφL KφL KφR KφR KLφ 0 0 0 KRL 2   K(u1, v1, . . . , u4, v4) = g KLφ 0 0 KRL 0  (38) KRφ 0 KLR 0 0  KRφ KLR 0 0 0 whose components are

L L R KφL(u1, v1, . . . , u4, v4) = GR(u31, v31)GR(u42, v42)GW (u43, v43) , (39) R R L KφR(u1, v1, . . . , u4, v4) = GR(u31, v31)GR(u42, v42)GW (u43, v43) , (40) R KLφ(u1, v1, . . . , u4, v4) = 2γFR(u31, v31)FR(u42, v42)GW (u43, v43) , (41) L KRφ(u1, v1, . . . , u4, v4) = 2γFR(u31, v31)FR(u42, v42)GW (u43, v43) , (42) L L KRL(u1, v1, . . . , u4, v4) = 2γGR(u31, v31)GR(u42, v42)FW (u43, v43) , (43) R R KLR(u1, v1, . . . , u4, v4) = 2γGR(u31, v31)GR(u42, v42)FW (u43, v43) . (44)

Here and below, we shall use the short hand uji = uj ui and vji = vj vi. The factor 2 = nS is associated with − − every boson propagator, due to the deﬁnition of γ = M/(NnS). Eigenfunction. Denoting

λ λ sech(πT u)2 χ = u , χ = v ,S(u) = , (45) u 2πT v 2πT πT we found that the following growth Ansatz

1−h +χu/2 1−h +χv /2 fφS(u21) b S(v21) b  ¯ f S(u )1−h+χu/2S(v )1−h+χv /2  L 21 21  u + u v + v  1−h+χu/2 1−h¯+χv /2  1 2 1 2 (u1, v1, u2, v2) =  fLS(u21) S(v21)  exp λu + λv (46) 2 2 F  1−h¯+χu/2 1−h+χv /2   fRS(u21) S(v21)  1−h¯+χu/2 1−h+χv /2 fRS(v21) S(v21) 9 reduces the inﬁnite-dimensional eigen-problem above to a 3 3 eigen-problem (the simpliﬁcation from 5 components to 3 is thanks to particle hole symmetry): ×       fφ fφ 1 0 AkL AkR f = M f ,M := k 0 γA k (47)  L  L 4  b F R fR fR kb γAF kL 0

The matrix elements are

2 2 4 W (1 hb + χu/2)W (1 hb + χv/2) kb = 2γ(2 sin 2πhb) AF AR(hb) − − (48) W (hb + χu/2)W (hb + χv/2) ¯ 2 2 2 ¯ 2 W (1 h + χu/2)W (1 h + χv/2) kL = 2(2 sin 2πh) R(h) A R(h) − ¯ − (49) W (h + χu/2)W (h + χv/2) ¯ 2 2 2 ¯ 2 W (1 h + χu/2)W (1 h + χv/2) kR = 2(2 sin 2πh) R(h) A R(h) − ¯ − (50) W (h + χu/2)W (h + χv/2) where we denoted 1 W (h) = ,R(h) = 4hhΓ( 2h) . (51) Γ(h)Γ(h + 1/2) − − Let us explain why the Ansatz (46) works. First, one can check by power counting that all the temperature dependence is cancelled. (The dependence on g2 and A drops out as well.) So we can assume πT = 1. Then, note that the Fourier and Laplace transform of powers of sech and sinh are Γ functions: Z eiωusech(u)2hdu = √πΓ(h + iω/2)Γ(h iω/2)W (h) (52) − Z ∞ Γ(h + s/2 + iω/2) e−(s+iω)u sinh(u)−2hdu = R(h) . (53) Γ(1 h + s/2 + iω/2) 0 − This means that, an eigen-function that is a power of sech

χu(u1+u2) 2(1−h)+χu F (u1, u2) = e sech(u21) (54) behaves well with retarded Green functions that are a “matching” power of sinh

−2h gR(u) = θ(u) sinh(u31) . (55)

Namely, Z W (1 h + χ /2) 2 u χu(u3+u4) 2h+χu du1du2F (u1, u2)gR(u31)gR(u42) = R(h) − e sech(u43) , (56) W (h + χu/2) the result of the convolution has the same functional form as F ; only the exponent and the prefactor are altered. The Ansatz (46) is chosen such that the above formula can be applied whenever a convolution with a pair of retarded Green functions is performed. With this in mind, (47) can be veriﬁed straightforwardly. We can now study the algebraic eigensystem (47), e.g., using symbolic algebra software. In particular, we solve the following equation for (χu, χv):

D(χu, χv) := det(M I) = 0 , (57) − so that the eigenvalue problem (47) has a nontrivial solution. We checked that, for any h (0, 1/4), χu = 0, χv = 1 ∈ and χv = 0, χu = 1 are always two solutions; these are equivalent to

(λu = 2πT, λv = 0) or (λu = 0, λv = 2πT ) . (58)

Since λuu + λvv = (λu + λv)t + (λu λv)t, we may identify −

λL = λu + λv , ip = λu λv , (59) − and thus the solutions (58) imply λL(p = i2πT ) = 2πT , as claimed in the main text. ± 10

1.0 1.00 0.5 0.8 0.95 0.4 0.6 0.90

* 0.3 v v χ 0.85 0.4 0.2 T ) 2 π ( p = 0 )/( L

0.80 λ 0.2 0.1 0.75 0.0 0.0 0.0 0.2 0.4 0.6 0.8 1.0 0.00 0.05 0.10 0.15 0.20 0.25 0.00 0.05 0.10 0.15 0.20 0.25

χu h h

FIG. 3. Left: Values of (χu, χv) such that the kernel eigensystem (47) has a nontrivial solution, for h = 0.01 (blue dashed), h = 0.15 (red), and h = 0.245 (black). For any h,(χu, χv) = (0, 1) and (1, 0) lie on the curve. This implies fast scrambling for the (1+1)-d critical solutions. The curve approaches the axes {χu = 0} ∪ {χv = 0} as h → 0 and h → 1/4. Middle: v∗ as a function of h ∈ (0, 1/4). Right: λL(p = 0) as as a function of h ∈ (0, 1/4).

The two solutions (58) are connected by a continuous curve (χu, χv) of solutions in the quadrant χu > 0, χv > 0 ; { } see Fig. 3 (left) for some plots. These solutions determine the relation between λL and p, for p ( 2iπT, 2iπT ). In ∈ − particular, the velocity v∗ in the main text can be evaluated as

∂λL(is) ∂χu D ∂χv D v∗ = = − . (60) ∂s ∂ D + ∂ D s=2πT χu χv χu=0,χv =1

Its dependence on h is plotted in Fig. 3 (middle). v∗ < 1 for any h (0, 1/4), and approaches 1 at both limits. As ∈ a consequence, at those limits, the fast scrambling regime x /t (v∗, 1) becomes empty. This is expected as both limits are free-fermionic. | | ∈ Let us also look at λL(p = 0), which determines the growth rate of the OTOC at x = 0. It is given by λL(p = 0)/(2πT ) = 2χ where D(χu = χ, χv = χ) = 0. We plot its dependence on h in Fig. 3 (right). We observe that this growth rate is signiﬁcantly below the fast scrambling bound 2πT , and tends to 0 at the free-fermionic limits. In this sense our model is less chaotic than the = 4 supersymmetric Yang-Mills theory, for which λL(p = 0) 2πT in the strong coupling limit (we thank DouglasN Stanford for pointing out this). →

B. Residual entropy

In this appendix we discuss brieﬂy the thermodynamics of the critical points, focusing on the residual entropy in (0+1)-d. Recall that the free energy in the large N limit is given by the saddle-point action, which can be obtained by standard methods. It reads as follows in D dimensional spacetime:

n γ Z βF/N = ln det(G) + S ln det(F ) β tr[G(x)Σ(x)]dDx . (61) − − 2 − where V is the spatial volume, such that βV = R 1dDx, and G,F and Σ are the saddle point solutions. In a low-T expansion,

βF = βE0 + S0 + o(1) , (62) − − where o(1) denotes terms that vanishes in the zero temperature limit. The leading, divergent, term is a UV-dependent ground state energy. We are interested in the constant term, which is the residual entropy. By extending the argument of [44], Appendix E, one can show that the βV term in (61) does not contribute to S0. Hence, S0 is given by the fermion and boson log-determinants: ∝

1 S0/N = sf + nSγsb , where sf = ln det(G) , sb = ln det(F ) , (63) − 2 which should be properly UV regularized. 11

0.6 0.5 0.4 / N 0

S 0.3 0.2 0.1 0.0 0 1 2 3 4 5 γ

FIG. 4. The residual entropy per fermion in the (0 + 1)-d critical points, as a function of the rank γ, for both g2 > 0 (bottom branch) and g2 < 0 (top branch). The solid curve is the analytical prediction, and the circles are from numerical solutions of the SD equations.

We now specialize to (0 + 1)-d, with a single-component spinor. We again follow [44]. Up to a divergent UV term, both sb and sf are sum over log-gamma functions of Matsubara frequencies,

∞ 1 X Γ(n + 2 + ∆) sf = ln , (64) − Γ(n + 3 ∆) n=0 2 − ∞ ∞ 1 X Γ(n + ∆b) 1 X Γ(n + 1 + ∆b) sb = ln + ln . (65) 2 Γ(n + 1 ∆b) 2 Γ(n + 2 ∆b) n=0 − n=0 −

Using Eq. (E7) of [44], we ﬁnd that sf and sb satisfy the following diﬀerential equations as a function of their respective scaling dimension ∆ and ∆b = 1 2∆: − ∂sf = π(2∆ 1) tan(π∆) , sf (∆ = 0) = ln 2 , (66) ∂∆ − ∂sb 1 = π(2∆b 1) cot(π∆b) , sf (∆b = 1/2) = 0 . (67) ∂∆b 2 −

Here, the boundary condition sf (∆ = 0) = ln 2 is fixed by the trivial limit. sb(∆b = 1/2) is fixed by the high-rank limit. Indeed, if sb(∆b = 1/2) = 0, S0/N γsb would be infinite since γ = at that point. So sb must vanish at that 6 ≥ ∞ point. Solving these equations for sf , sb, and using the rank-exponent relation, we obtain an analytical prediction for 2 2 S0, which we plotted in Fig. 4 as a function of γ (for both g > 0 and g < 0). It reproduces well the numerical value obtained from solving the SD equations and using thermodynamic relations.