Complex Operator Growth in Dissipative Quantum Systems Authors: Hikaru Wakaura, Taiki Tanimae Published: 2026 ABSTRACT The universal operator-growth hypothesis (OGH) states that, in a closed chaotic system, the Lanczos coefficients grow linearly, bn≃αn. We ask how this structure is modified when the system is coupled to a Markovian environment, so that the generator becomes non-Hermitian. Applying the Arnoldi recursion to the vectorized Lindbladian in the infinite-temperature Wightman inner product, we organize the resulting pair of growth rates αC≡αR+iαI—defined as effective slopes of the sub-diagonal and diagonal Arnoldi coefficients over a pre-registered fit window—around two statements whose logical status we delimit precisely. First, whenever the dissipator acts as D=−2γG^ with G^, a Hermitian grading (all dephasing-type baths), the diagonal obeys the identity Rean=−2γ〈G^〉n: the imaginary rate measures how fast the growing operator accumulates weight in the dissipation channels. Second, we prove a conditional parity theorem: if the Hamiltonian, jump operators, and seeds can be made simultaneously real in some basis (an antiunitary condition), then bn is even, and Rean is odd in γ exactly, so αR is renormalized only at O(γ2), and αI=2κ0γ follows from closed-system data alone. We exhibit a one-qubit Lindbladian that satisfies the often-assumed generator symmetry G†(γ)=−G(−γ) yet violates parity (b1=|1−γ|), showing that the extra condition is essential; all models studied here satisfy it bit-exactly. For large-q SYK, these ingredients predict αC=J−2i(q−2)γ, whose imaginary part is fixed solely by the interaction range; the first ladder step is exact, and the multi-step increments approach q−2 with system size (1.92±0.04 at N=12, q=4). Under a common fit protocol, the closed-system rates saturate by N=10 (αR(0)→0.437, 2κ0→0.224). The imaginary rate is not an independent observable at leading order—its content is its sign, which resolves how the growing operator meets its environment (opposite for spin chains and SYK). 1. Introduction Operator growth—the progressive delocalization of an initially simple operator under Heisenberg evolution—underlies quantum chaos, thermalization, and the emergence of hydrodynamics [1]. The universal operator-growth hypothesis (OGH) [2] made this quantitative. Expanding the Heisenberg equation of motion in the Krylov basis generated by the Liouvillian L0=i[H,·], one obtains a semi-infinite hopping chain whose amplitudes, the Lanczos coefficients bn, grow linearly, bn≃αn (with a logarithmic correction in one dimension). The single rate α controls the exponential growth of Krylov complexity [3,4], upper-bounds a large class of complexity measures including the out-of-time-order correlator [5,6], and yields λL≤2α—a bound that is tighter than, and complementary to, the Maldacena–Shenker–Stanford bound λL≤2πT [7]. Real systems are open. Under Markovian dissipation, the generator of operator dynamics is the (adjoint) Lindbladian, which is non-Hermitian, and the symmetric Lanczos algorithm no longer applies: one must use the Arnoldi iteration [8] or a bi-Lanczos scheme [9,10,11], and the OGH has been examined for large-q SYK with weak dissipation [12]. These works establish the recursion machinery and compute the coefficients numerically; the one analytic large-q SYK rate [12] is extracted from the (real) off-diagonals and stays real, with dissipation entering as a phase. The imaginary part, however, lives in the diagonal coefficients an, for which prior open-system studies compute values but, to our knowledge, give no analytic account. What has been missing is a physical and analytic treatment: what the diagonal measures, how it depends on dissipation, and under what conditions its structure is protected. Our advance is not that the coefficients are complex—they are, in any non-Hermitian recursion—but that the imaginary rate is pinned down physically and analytically: an exact identity identifies it with the operator’s dissipation-grading weight; a conditional parity theorem—whose hypothesis we make precise, prove, and delimit with an explicit one-qubit counterexample—fixes its linear-in-γ onset and protects αR to O(γ2) for the class of models studied here; and large-q SYK yields the closed-form prediction αC=J−2i(q−2)γ. Two rigorous open-system statements survive en route to a chaos bound (a norm-decay lemma and a locality moment bound), but the chain of inequalities that would deliver λL≤2αR(γ) rests on assumptions we do not prove; we state it as an open problem rather than a result. We stress at the outset what the imaginary rate does and does not add: at leading order, it is fixed by closed-system data, αI=2κ0γ, so its content is not an independent number but a mechanism diagnostic—its sign resolves how the growing operator meets the environment (accumulating dissipative weight in spin chains, shedding it in SYK)—together with the breakdown of this closed-system predictability once αI∼αR. 2. Setup: Arnoldi Recursion for the Vectorized Lindbladian We work at an infinite temperature with the Wightman inner product (A|B)=Tr(A†B)/dim, under which the commutator superoperator [H,·] is Hermitian, and the closed-system generator L0=i[H,·] is anti-Hermitian. As in all recursion-method studies of open systems, the construction is inner-product dependent; our results are for this standard choice, and their finite-temperature counterparts are examined in Appendix F. The Heisenberg generator of the Lindblad dynamics [13,14] is(1)G(O)=i[H,O]+∑μLμ†OLμ−12{Lμ†Lμ,O}=L0+D, with L0=i[H,·] being anti-Hermitian and D being the dissipator. Vectorizing O and orthonormalizing the Krylov sequence {GnO0} [15] by the Arnoldi iteration with full reorthogonalization produces an upper-Hessenberg matrix, h, whose sub-diagonal bn=hn+1,n∈R≥0 and diagonal an=hn,n∈C define(2)αC≡αR+iαI,bn≃αRn,−Rean≃αIn(n∈[n1,n2]). Three definitional caveats apply, and we state them up front. (i) At any fixed finite N, the grading expectation is bounded, 〈G^〉n≤N, so the true n→∞ slopes vanish; a nonzero rate refers to the growth window 1≪n≪N, and the thermodynamic limit must be taken first. The rates are thus preasymptotic growth-phase quantities: they characterize the regime before the O(1/γ) crossover at which dissipative oscillations and saturation set in, not a large-n asymptote. We, therefore, define αR,αI as effective slopes over a fit window that is fixed in advance and common to all system sizes (n∈[4,18] unless stated), and we report their sensitivity to the window choice in Appendix F. (ii) The two slopes are extracted from two real sequences (bn and −Rean) and bundled into one complex number for compactness; αC is not an eigenvalue, no single observable evolves directly with its magnitude or phase, and it is defined relative to this recursion and inner product (its scheme dependence is quantified in Appendix A); its operational content is the pair (αR,αI) and, on the experimental side, the t-odd component of the autocorrelation discussed in Section 5. (iii) For every model studied below the generator and seed can be represented by real matrices; the Krylov vectors then have real coefficients in that basis (their components in the computational vectorization are complex), and hence Iman=0 identically—the diagonal is its real part (Appendix B). For γ=0, G=L0 is anti-Hermitian, the Hessenberg matrix is tridiagonal, and bn reduce to the OGH Lanczos coefficients with αI=0. Numerically, we use a chaotic mixed-field Ising chain, H=−∑iZiZi+1−g∑iXi−h∑iZi with (g,h)=(1.05,0.5), under bulk dephasing Li=γZi; conventions and further models are given in Appendix A and Appendix F. Whenever we report Krylov complexity, we mean(3)K(t)=∑nn|φn(t)|2∑n|φn(t)|2,φn(t)=(vn|eGt|O0), computed from the truncated (mb=34) Hessenberg matrix; the explicit normalization is required because the non-unitary evolution does not conserve ∑n|φn|2. 3. The Complex Growth Rate 3.1. An Exact Identity for the Imaginary Rate For dephasing, the dissipator is diagonal in the Pauli-string basis, D=−2γG^, where G^ is the transverse-weight operator counting the sites on which an operator acts as X or Y. Decomposing the Arnoldi diagonal as an=(vn|L0|vn)+(vn|D|vn) and using that the first term is purely imaginary (L0 anti-Hermitian), while the second is real (D Hermitian) gives the exact identity(4)Rean=−2γ〈G^〉n, which is algebraic—it holds for every H, γ and n (Appendix B); the 10−15-level numerical check in the companion code validates the implementation, not the physics. We are explicit about its mathematical weight: Equation (4) is the diagonal matrix element of the Hermitian part of G, i.e., a one-line consequence of the decomposition G=L0+D—its value lies not in depth but in the reading it licenses: the diagonal decay profile is an exact, step-resolved record of how much of the growing operator lies in the channels the environment couples to, αI=2γσG with σG the effective slope of 〈G^〉n over the fit window (Figure 1a,b). 3.2. A Conditional Parity Theorem in the Dissipation Strength Writing G(γ)=L0+γD˜ with D˜ Hermitian, the generator obeys(5)G†(γ)=−G(−γ)⇒μk(γ)∗=(−1)kμk(−γ), for the moments μk=(O0|Gk|O0). It is tempting to conclude that Equation (5) alone forces bn to be even and Rean to be odd in γ—an earlier version of this work did so—but the inference is false: for a non-normal generator, the Arnoldi coefficients are functions of the mixed Gram matrix Γjk=(GjO0|GkO0), which involves (G†)jGk and is not determined by the scalar moment sequence. A one-qubit Lindbladian makes this explicit: for H=−Y/2, L=γZ and seed O0=(X+Z)/2, the generator symmetry in Equation (5) holds exactly, yet the first Arnoldi coefficient is(6)b1(γ)=|1−γ|, manifestly containing a linear term (Appendix C). Parity is restored by one additional hypothesis. Theorem (conditional parity). Suppose there exists an antiunitary, T, with TG(γ)T−1=G(γ)† and TO0∝O0—equivalently, a basis in which Hamiltonian, jump operators, and seeds are represented by simultaneously real matrices. Then, vn(−γ)=(−1)nTvn(γ), the Gram matrices at ±γ are complex conjugates, and(7)bn(−γ)=bn(γ),an(−γ)=−an(γ)∗,exactly (proof in Appendix C). Consequently, for this class, αR(γ)=αR(0)+c2γ2+O(γ4), and αI(γ)=2κ0γ+O(γ3). The hypothesis is not exotic: the mixed-field Ising chain with Z-dephasing, Majorana-dephased SYK, and the number-dephased Bose–Hubbard trimer all satisfy it, and their parity is verified at machine precision (0.0 for Ising, 4.4×10−16 for SYK; twelve configurations, including deliberately broken ones, in Appendix C). It is, however, a genuine restriction: adding a Y-field to the Ising chain with the same seed breaks parity at the 10−1 level, and the linear-in-γ renormalization of αR returns. The former universality claim is hereby withdrawn. There are three quantitative consequences, all under the common fit protocol of Section 2. (a) Identity versus prediction. Within any single fit window, the identity in Equation (4) makes αI(γ)=2γκ(γ) exact by construction (we verify the ratio equals 1 to ten digits); the nontrivial content of Figure 2a is, therefore, basis continuity, κ(γ)=κ0+O(γ)—numerically κ(γ)≃0.0869+0.0115γ in the common window, a 5.1% drift across γ≤0.4, with same-window leading-order agreement αI/(2γκ0)−1=+0.09% at γ=0.05. We no longer describe this as a parameter-free prediction. (b) Quadratic protection. In the common window, fitting αR(γ)=αR(0)+c1γ+c2γ2 over γ≤0.4 gives c1=0.002 (consistent with parity) and c2=0.008; fixing c1=0 as parity dictates gives c2=0.012. The value of c2 is strongly window- and estimator-dependent (0.01–0.06 across the scan of Appendix F), but its sign is stable: c2>0, so weak dephasing increases the hopping (Figure 2b). (c) Finite size. Under the single pre-registered window, for every N, the closed-system invariants saturate: αR(0)=0.178,0.323,0.402,0.422,0.436,0.437 and κ0=0.067,0.087,0.105,0.109,0.112,0.112 for N=5–10; the N=6 used in Figure 1 and Figure 2 sits 26% below the plateau, so those figures illustrate structure, while headline rates should be read from the plateau (αR(0)→0.437, 2κ0→0.224; Appendix F). In parallel, the Krylov complexity saturates and the Wightman autocorrelation decays with γ (Figure 1e,f), while the Liouvillian spectrum broadens from the imaginary axis into the left half-plane (Figure 1d). 3.3. Model Dependence and Large-q SYK The grading G^—and hence the sign and magnitude of αI—is model-dependent, which is the non-universal content of the imaginary rate. For the SYKq model [16,17] with Majorana dephasing Li=2γψi, the dissipator reduces, in the size-parity sector of a single-Majorana seed, to the operator co-size: Rean=−2γ(N−〈s〉n), so a growing operator becomes less dissipation-sensitive, opposite to the spin chain. The large-q size ladder is known—each interaction adds q−2 Majoranas, so 〈s〉n=1+(q−2)n [18]—giving σG=−(q−2). With αR=J [2], we obtain the closed-form arge-q prediction (8)αCSYK=J−2i(q−2)γ, whose imaginary part is fixed solely by the interaction range q, independent of the coupling J. Its logical status is mixed, and we state it plainly: the first ladder step s1−s0=q−2 is an operator-algebra identity (exact at any N, verified to machine precision), while the identification of subsequent Krylov steps with ladder generations is a large-q, large-N statement. Disorder-averaged finite-size data support the approach: at q = 4 the second increment s2 − s1 rises from 1.58 ± 0.22 (N = 8) through 1.86 ± 0.03 (N = 10) to 1.92 ± 0.04 (N = 12), approaching q − 2 = 2 as finite-size saturation recedes (Appendix D). The exact statements of this paper are Equations (4) and (7) (the latter under its stated hypothesis); Equation (8) is a prediction they jointly motivate. 4. Rigorous Partial Results and an Open Problem Two open-system statements can be made rigorously. First, a norm-decay lemma: ddt(O|O)=2(O|D|O)=−4γ〈G^〉(O|O)≤0, so the Wightman norm of any operator is non-increasing under dephasing. Second, a locality moment bound: the generator G=L0+D is a sum of geometrically local, bounded superoperators with per-site strength w≤2J+2γ, and a generalized Lieb–Robinson counting for local Markovian generators [19] gives(9)∥GnO∥≤(cw)nn!∥O∥,w=2J+2γ, with c being a lattice-connectivity constant (Appendix E). Dissipation, therefore, enters the moment growth only through the O(γ) shift of w, not through the factorial structure. In an earlier version, we combined Equation (9) with regularity assumptions to conjecture a dissipative chaos bound λLopen≤2αR(γ). We now refrain: (i) the step from factorial moments to linearly growing bn requires a quantitative lower bound on the Krylov Gram determinants that we cannot supply; (ii) the closed-system chain of inequalities relating Krylov growth to the OTOC has no proven non-Hermitian counterpart; and (iii) an operational definition of λLopen, together with an actual open-system OTOC computation, would be needed to give the statement empirical content. In addition, since parity gives c2>0 for our models, any such bound would be weaker than the closed-system one at small γ. Formulating and testing a dissipative chaos bound is, in our assessment, a well-posed open problem, and Equation (9) fixes the one ingredient that dissipation demonstrably does not spoil. 5. Discussion Markovian dissipation splits the operator-growth data into a pair of effective rates, αC=αR+iαI, with the imaginary part tied by the exact identity (4) to the operator’s dissipation-grading weight, its γ-dependence fixed—for the antiunitary-symmetric class—by the conditional parity theorem (7), and a closed-form large-q SYK prediction (8). The imaginary rate is an environment-resolved diagnostic: αI is the gradient along the Krylov chain of the on-site decay rate −Rean. Its leading order carries no information beyond the closed-system slope κ0 and the known dissipation strength—that is precisely the content of the parity theorem—so the physics it resolves is its sign and the departure from the rigid-rotation regime. For dephasing spins, the decay rate grows with depth (transverse weight accumulates, αI>0); for Majorana SYK, it decreases (a growing operator’s co-size shrinks, αI<0)—this is not gain but a negative sensitivity gradient. In neither case does the total norm grow (the norm-decay lemma of Section 4 holds). Two honest limitations frame these results. First, αI is scheme-specific: the corresponding diagonal slope extracted from a bi-Lanczos recursion differs from the Arnoldi value by 50–150% over γ∈[0.05,0.4] (Appendix A), so αC is defined relative to the Arnoldi–Wightman construction, not as a recursion-independent invariant. Second, all rates are window-defined effective slopes (Section 2), with sensitivity quantified in Appendix F. The construction extends to bosons: in a Bose–Hubbard trimer [20] at fixed total particle number—an energy-bounded, chaotic, truncation-free model—the same complex-rate structure appears, with αI∝γ (αI/γ drifting by 6% over γ∈[0.05,0.4]), αR even in γ, and the grading identity Reak=−γ2∑i〈(Δni)2〉k holding to 10−14; both rates converge with Ntot (Appendix F). A single driven Kerr mode, by contrast, is Fock-truncation-limited (Appendix F), so energy boundedness is essential—as it is in finite-energy encodings such as GKP [21]. Experimentally, the coefficients follow from moments of the autocorrelation C(t)=(O|eGt|O) via the recursion method; the diagonal an—and hence αI—is encoded in the component of C(t) that is not even in t, absent in the closed system, so the imaginary rate is accessible wherever Lanczos coefficients are already extracted [2]. We caution that this reconstruction is an inference problem in its own right: high-order moments amplify measurement noise rapidly with n, so a practical extraction of an from experimental C(t) will require regularized or otherwise noise-robust estimation; developing such schemes is beyond the scope of this work. Finally, the broadening of the Liouvillian spectrum (Figure 1d) motivates—but we emphasize does not by itself establish—a complex-plane spectral form factor [22] as a dissipative chaos diagnostic.