CMI_SM.pdf

Universal Spreading of Conditional Mutual Information in Noisy Random Circuits

Abstract

We study the evolution of conditional mutual information (CMI) in generic open quantum systems, focusing on one-dimensional random circuits with interspersed local noise. Unlike in noiseless circuits, where CMI spreads linearly while being bounded by the lightcone, we find that noisy random circuits with an error rate $p$ exhibit superlinear propagation of CMI, which diverges far beyond the lightcone at a critical circuit depth $t_{c}\propto p^{-1}$. We demonstrate that the underlying mechanism for such rapid spreading is the combined effect of local noise and a scrambling unitary, which selectively removes short-range correlations while preserving long-range correlations. To analytically capture the dynamics of CMI in noisy random circuits, we introduce a coarse-graining method, and we validate our theoretical results through numerical simulations. Furthermore, we identify a universal scaling law governing the spreading of CMI.

Introduction.—Understanding information spreading in non-equilibrium quantum systems is central to studying chaotic quantum systems, as well as essential for realizing quantum technologies. For these theoretical and practical purposes, random quantum circuits have served as a fruitful model to study the generic properties of quantum systems. The irreversible growth and spread of entanglement in random quantum circuits have elucidated the mechanisms underpinning quantum thermalization and chaos Deutsch (1991); Srednicki (1994); Nahum et al. (2017, 2018), provided useful models for black holes Hayden and Preskill (2007); Sekino and Susskind (2008); Hosur et al. (2016), and showcased their utility in efficient quantum error-correcting codes Brown and Fawzi (2013, 2015); Gullans et al. (2021). However, incorporating noise is essential to enhance the physical relevance and applicability to near-term quantum devices. Thus, random quantum circuits interspersed with decoherence channels have been introduced to describe generic open quantum systems Noh et al. (2020); Li et al. (2023); Zhang et al. (2019); Schuster and Yao (2023).

Among various measures characterizing information dynamics, conditional dependence holds fundamental and practical significance in quantum information and many-body physics. It is intimately connected with the classical description of quantum states Molnar et al. (2015); Hastings (2007); Leifer and Poulin (2008); Poulin and Hastings (2011), efficient quantum algorithms for state preparations Brandão and Kastoryano (2019); Kato and Brandão (2019), topological orders Kitaev and Preskill (2006); Levin and Wen (2006); Kim (2013); Kim et al. (2023), and the information-theoretic measure of mixed-state quantum entanglement Christandl and Winter (2004); Brandão et al. (2011); Seshadreesan and Wilde (2015).

Given a state $\rho$, the direct correlation between the subsystems $A$ and $C$ is quantified by mutual information,

$$I(A:C)=S(A)+S(C)-S(AC),$$ (1)

where $S(X)=-\Tr(\rho_{X}\log_{2}{\rho_{X}})$ is the von Neumann entropy of the reduced density matrix $\rho_{X}$ of the subsystem $X$. Meanwhile, the dependence of $A$ and $C$, conditioned on another subsystem $B$, is captured by conditional mutual information (CMI),

$$I(A:C|B)=S(AB)+S(BC)-S(B)-S(ABC).$$ (2)

CMI is associated with approximate recoverability Fawzi and Renner (2015); Petz (2003), thereby serving as a measure of conditional dependence. The structures of CMI in the states in thermal equilibrium or topologically ordered states have been studied extensively Kato and Brandão (2019); Kuwahara et al. (2020); Harrow et al. (2020). However, its behavior in dynamical systems remains less understood.

In this Letter, we study the dynamics of CMI in generic open quantum systems. Specifically, we consider quantum circuits that start with a product state and are subjected to random local gates, possibly with noise. We establish that in noiseless circuits, CMI spreads linearly constrained by the lightcone. Introducing noise, however, leads to the rapid spreading of CMI beyond the ligthcone. We demonstrate the underlying mechanism for such rapid spreading by showing that the local noise combined with a scrambling unitary selectively removes the short-range correlation while preserving the long-range correlation.

Equipped with the understanding of the non-local spreading of CMI, we study one-dimensional random circuits with interspersed local noise controlled by the error rate $p$. We introduce coarse-grained random circuit, a new theoretical model that admits theoretical and numerical analysis. This model reveals that the higher the error rate within the circuit, the more rapid the spreading of CMI is. Remarkably, the spreading diverges at a critical timestep $t_{c}\propto p^{-1}$, leading to the entire system becoming conditionally dependent. Furthermore, we analytically identify and numerically validate a universal scaling law governing the spreading of CMI.

Figure 1: (a) 1D noiseless quantum circuit. $\mathcal{L}(A)$ and $\mathcal{L}(C)$ are denoted by the regions enclosed by blue and red dashed lines, respectively. The gates outside of $\mathcal{L}(A)\cup\mathcal{L}(B)$ are colored grey. (b) Quantum circuit with noise. After two layers of random Clifford gates, a fraction of $p$ of the qubits in $B$ are completely depolarized. (c) Due to the gates $U_{1}$ and $U_{2}$ in (b), $A$ and $C$ effectively share $m$ Bell pairs with $B$, respectively.

Bound of lightcone in noiseless circuits.—We first show that when a quantum circuit is noiseless, the spreading of CMI is bounded by the lightcone. Consider a one-dimensional quantum circuit of $N$ qubits that begins with a product state $\rho^{\mathrm{init}}=\bigotimes_{i=1}^{N}\rho_{i}$. We choose two subsets of qubits $A$ and $C$, and denote $B$ to be ${\{1,2,\dots,N\}\setminus(A\cup C)}$ as depicted in Fig. 1(a). We say $A$ and $C$ are lightcone-separated if $\mathcal{L}(A)\cap\mathcal{L}(C)=\emptyset$, where $\mathcal{L}(X)$ is the backward lightcone of $X$.

Given the output state $\rho$, suppose $A$ and $C$ are lightcone-separated. We undo the gates outside of ${\mathcal{L}(A)\cup\mathcal{L}(C)}$ by applying the inverses of those gates successively. Since those are unitary gates applied only on $B$, they do not change $I(A:C|B)$. Then, the resultant state $\sigma$ after undoing those gates takes the form of ${\sigma=\sigma_{AB_{A}}\otimes\sigma_{B_{0}}\otimes\sigma_{CB_{C}}}$, where $B_{A}$ (or $B_{C}$) denotes the set of qubits that are in the support of $\mathcal{L}(A)$ (or $\mathcal{L}(C)$), and $B_{0}=B\setminus(B_{A}\cup B_{C})$. This results in ${I(A:C|B)=0}$ from the definition of CMI. Therefore, the spreading of CMI is bounded by the lightcone in noiseless circuits.

Selective removal of correlation.—Contrarily, noisy circuits can spread CMI into lightcone-separated regions. Here, we show that this rapid spreading stems from the fact that local noises, along with the scrambling unitary, selectively destroy the short-range correlation while preserving the long-range correlation. To understand this, we consider a quantum circuit of four blocks consisting of $m\gg 1$ qubits, as depicted in Fig. 1(b). We denote the first and last blocks by $A$ and $C$, respectively, and the two blocks in the middle by $B$. Starting from the product state $\ket{0}^{\otimes 4m}$, the two halves of the system undergo random Clifford gates $U_{1}$ and $U_{2}$, respectively. Then, we apply another random Clifford gate $U_{3}$ on $B$. Finally, we completely depolarize a fraction of $p$ of the qubits in $B$:

$$\rho_{ABC}\mapsto\Tr_{B_{2}}\rho_{ABC}\otimes\frac{1\kern-2.5pt\text{l}_{B_{2}}}{d_{B_{2}}},$$ (3)

where $B_{2}$ represents the depolarized qubits, and $B_{1}=B\setminus B_{2}$. Here, $d_{B_{2}}$ is the Hilbert space dimension of $B_{2}$. Note that $A$ and $C$ are lightcone-separated.

A key observation is that depolarizing $B_{2}$ only removes the short-range correlation ${I(A:B)}$ but not the long-range correlation ${I(A:BC)}$, as long as ${p<1/2}$. Initially, random Clifford gates $U_{1}$ and $U_{2}$ almost maximally entangle the first and last two blocks because of the large dimensionality Preskill (2018); Apel et al. (2022). This effectively forms $m$ Bell pairs shared by $A$ and $B$, and another $m$ Bell pairs shared by $B$ and $C$ [Fig. 1(c)]. Therefore, noting that each Bell pair counts two in mutual information, ${I(A:B)=I(B:C)=2m}$ and ${I(A:BC)=I(AB:C)=2m}$ before we depolarize $B_{2}$.

Then, given $p<1/2$, the complete depolarization of $B_{2}$ preceded by $U_{3}$ decreases ${I(A:B)}$ by $4mp$. To see this, note that before the depolarization, $U_{3}$ nearly maximally entangle $AB_{1}$ with $B_{2}C$ Preskill (2018); Apel et al. (2022), resulting in $S(AB_{1})=2mp+m$. Meanwhile, ${S(A)=m}$ and ${S(B_{1})=2m(1-p)}$ as both $A$ and $B_{1}$ are maximally entangled with the other subsystems. These leads to ${I(A:B_{1})=2m(1-2p)}$, and thus,

$$I(A:B)=2m(1-2p)$$ (4)

after we depolarize $B_{2}$.

Contrarily, depolarizing $B_{2}$ does not decrease ${I(A:BC)}$. Before we depolarize $B_{2}$, the decoupling inequality Preskill (2018); Abeyesinghe et al. (2009); Hayden et al. (2008); Hayden and Preskill (2007) states that due to the scrambling unitary $U_{3}$, $A$ and $B_{2}$ are almost decoupled when ${p<1/2}$:

$$\mathbb{E}_{U_{3}}\left[\left\lVert\rho_{AB_{2}}(U_{3})-\rho_{A}\otimes\frac{1\kern-2.5pt\text{l}_{B_{2}}}{d_{B_{2}}}\right\rVert_{1}\right]\leq 2^{-m(1-2p)},$$ (5)

where $\rho_{AB_{2}}(U_{3})$ is the reduced density matrix of $AB_{2}$ after applying $U_{3}$. Here, $\lVert\cdot\rVert_{1}$ denotes the trace norm, and $\mathbb{E}_{U_{3}}$ represents averaging over random Clifford gates on $B$. The decoupling of $A$ and $B_{2}$ ensures the recovery of the $m$ Bell pairs with $A$ from $B_{1}C$ with probability $1-\mathcal{O}(2^{-m(1-2p)})$ Yoshida and Kitaev (2017); Yoshida (2022). Therefore, any operation on $B_{2}$ does not decrease ${I(A:BC)}$ in large $m$ limit. As a result,

$$I(A:BC)=2m,$$ (6)

after we depolarize $B_{2}$.

Eqs. (4) and (6) confirm that the loss of $B_{2}$ preceded by $U_{3}$ selectively destroys ${I(A:B)}$ while preserving ${I(A:BC)}$. Importantly, by the chain rule ${I(A:C|B)}={I(A:BC)}-{I(A:B)}$, CMI is the discrepancy between the short-range and long-range correlations, resulting in ${I(A:C|B)=4mp}$. Consequently, the lightcone-separated regions $A$ and $C$ share a non-zero CMI after the depolarization. See Ref.  for an alternative derivation based on the stabilizer formalism.

Figure 2: (a) Coarse-grained random circuit. Beginning with a product state, we alternately apply random Clifford gates to the neighboring blocks, followed by heralded depolarizing channels. (b) At each timestep, the system is divided into $A$, $B$, and $C$ to calculate ${I^{\mathrm{norm}}(A:C|B)}$.

Figure 3: (a) Evolution of ${I^{\mathrm{norm}}(A:C|B)}$ with various circuit depths $t$ and separation lengths $x$. The first three columns present the numerical results for the finite coarse-graining factors $m$, whereas the last column shows the analytical result from Eq. (10). The first row is the result of noiseless circuits ($p=0$), with subsequent rows for $p=1.0\times 2^{-9}$ and $p=1.5\times 2^{-9}$, respectively. Red dashed lines mark the limit set by lightcone ($x=t$). For noisy cases ($p>0$), $t_{c}$ is marked with white dashed lines. White solid lines represent $x_{\mathrm{dec}}(t)$. (b) $\tilde{x}_{\mathrm{dec}}(\tilde{t})$ with various error rates $p$ and coarse-graining factors $m$. The theoretical result for $m\rightarrow\infty$ is depicted as black dashed lines. The numerical results in (a) and (b) are based on averages from $1000$ circuit realizations with $N=2^{10}$.

Coarse-grained random circuit.—With the understanding of the non-local spreading of CMI, we study the dynamics of CMI in noisy random circuits. Inspired by Ref. Choi et al. (2020), we introduce the coarse-grained random circuit, a new model of noisy random circuits [Fig. 2(a)]. This circuit comprises a one-dimensional array of $N\gg 1$ blocks, each containing $m$ qubits. Therefore, the coarse-graining factor $m$ determines the local dimensionality. Starting with a product state $\ket{0}^{\otimes mN}$, we alternately apply $2m$-qubit random Clifford gates on the even and odd pairs of the adjacent blocks. After each layer of unitary gates, we apply heralded depolarizing channel on each qubit with an error rate $p$. A heralded depolarizing channel acting on qubit $i$ with an error rate $p$ is defined by completely depolarizing qubit $i$ with probability $p$:

$$\rho\mapsto\begin{cases}\Tr_{i}\rho\otimes 1\kern-2.5pt\text{l}_{i}/2&\text{with Prob. of $p$,}\\ \rho&\text{with Prob. of $1-p$,}\end{cases}$$ (7)

which is a simplified version of the depolarizing channel—this noise model is identical to the depolarizing channel, except the spacetime location of depolarization is determined for each circuit realization . Every timestep, we calculate the normalized CMI ${I^{\mathrm{norm}}(A:C|B)}={I(A:C|B)/m}$, averaged over circuit realizations. Here $A$ is the first $N/2-x$ blocks, $B$ is the $2x$ blocks in the middle, and $C$ is the last $N/2-x$ blocks [Fig. 2(b)]

The choice of unitary gates and noise channel is due to their classical simulability and analytical convenience, as Clifford gates and complete depolarization are efficiently described in the stabilizer formalism Gottesman (1998); Aaronson and Gottesman (2004). Importantly, since the Clifford group forms a unitary 3-design Webb (2016); Zhu (2017), the averaged quantities are expected to be similar to those from the Haar random circuit or generic chaotic models Nahum et al. (2017, 2018). Moreover, despite the simplification, the heralded depolarizing channel showcases qualitative similarity to the depolarizing channel when combined with random gates Noh et al. (2020); Li et al. (2023).

Since the qubits are kept depolarized, the system’s entropy increases until the state becomes maximally mixed. Roughly, each qubit is scrambled with $\sim m/p$ qubits before being depolarized. Thus, given a sufficiently small error rate, each qubit becomes maximally entangled with others before its depolarization. This results in two depolarized qubits for each depolarization event . Therefore, at timestep $t$, $~{}m(1-2pt)$ qubits remain un-depolarized in each block, and the system becomes almost maximally mixed at the critical timestep $t_{c}=1/2p$.

Fig. 3(a) presents the evolution of ${I^{\mathrm{norm}}(A:C|B)}$ for finite $m$ (numerical) and $m\rightarrow\infty$ (analytical) with various error rates. In noiseless circuits, ${I(A:C|B)=0}$ for $x>t$, since $A$ and $C$ are lightcone-separated. We observe that the spreading of $I(A:C|B)$ becomes faster as we increase the coarse-graining factor $m$. This is because there are non-negligible portions of Clifford gates that do not maximally entangle the two blocks when $m$ is small, and that portion shrinks as $m$ grows Preskill (2018); Apel et al. (2022).

Contrarily, introducing noise results in ${I(A:C|B)}$ attaining non-zero values in regions well beyond the lightcone, attributed to the selective removal of short-range correlations. This phenomenon can be analytically captured in the limit $m\rightarrow\infty$ . As seen in Eq. (6), depolarization channels cannot decrease the long-range correlation ${I(A:BC)}$ as long as less than half of the qubits are depolarized. Meanwhile, at every two timesteps, a $2m$-qubit random Clifford gate is applied between $A$ and $B$, thereby increasing ${I(A:BC)}$. Since the random Clifford gate maximally generates entanglement in the large $m$ limit and that there are $2m(1-2pt)$ non-depolarized qubits on its support, ${I(A:BC)}$ increases by ${4m(1-2pt)}$ at every two timesteps. Therefore, by integrating this increment,

$$I(A:BC)=2mt(1-pt),$$ (8)

for $t<t_{c}$. In contrast, using a similar method for deriving Eq. (4), we show in Ref.  that the short-range correlation ${I(A:B)}$ is upper bounded by ${2m(1-2pt)x}$, resulting in

$$I(A:B)=\min\{2m(1-2pt)x,I(A:BC)\}.$$ (9)

Therefore, by the chain rule,

$$I^{\mathrm{norm}}(A:C|B)=\max\{2t(1-pt)-2(1-2pt)x,0\},$$ (10)

given $t<t_{c}$. Note that $I(A:C|B)=0$ when $t>t_{c}$ as the system becomes maximally mixed.

As we increase the separation $x$, ${I^{\mathrm{norm}}(A:C|B)}$ decays until it becomes zero. We define $x_{\mathrm{dec}}(t)$ as the separation $x$ at which ${I(A:C|B)}$ begins to be zero for $t<t_{c}$, indicating the spreading front of CMI. Then, Eq. (10) demonstrates that

$$x_{\mathrm{dec}}(t)=\frac{t(1-pt)}{1-2pt}.$$ (11)

One can see that the spreading becomes more rapid as we increase the error rate $p$. Notably, the spreading front $x_{\mathrm{dec}}(t)$ diverges as ${t\rightarrow t_{c}}$. Therefore, a non-negligible amount of conditional dependence spreads out to the entire system right before the state becomes maximally mixed.

Our derivations of the rapid spreading and critical behavior of CMI are based on the limit of $m\rightarrow\infty$, where the long-range correlation is perfectly protected from the local noise while the short-range correlation keeps decreasing. However, we emphasize that this selective removal of short-range correlation also emerges in the cases of finite $m$ as long as the error rate is sufficiently low. As seen in Fig. 3(a), those features of rapid spreading and divergence at $t_{c}$ are observed in the numerical results of the finite $m$.

Universal scaling law.—Remarkably, the dynamics of CMI exhibit a universal scaling law. Specifically, we consider the following rescaling:

$$\tilde{t} =2pt,$$ (12)

$$\tilde{x} =2px,$$ (13)

$$\tilde{I}^{\mathrm{norm}}(A:C|B) =pI^{\mathrm{norm}}(A:C|B).$$ (14)

This rescaling makes Eqs. (10) and (11) error-rate-independent, demonstrating the universal scaling law:

$$\tilde{I}^{\mathrm{norm}}(A:C|B)=\max\{0,\tilde{t}(1-\tilde{t}/2)-(1-\tilde{t})\tilde{x}\},$$ (15)

$$\tilde{x}_{\mathrm{dec}}(\tilde{t})=\frac{\tilde{t}(1-\tilde{t}/2)}{1-\tilde{t}},$$ (16)

for $\tilde{t}<1$, where $\tilde{x}_{\mathrm{dec}}(\tilde{t})=2px_{\mathrm{dec}}(\tilde{t}/2p)$. Fig. 3(b) presents rescaled spreading fronts $\tilde{x}_{\mathrm{dec}}(\tilde{t})$ from the numerical results of finite $m$ along with the analytical result of $m\rightarrow\infty$ in Eq. (16). For a fixed coarse-graining factor $m$, the spreading fronts across varying error rates collapse into a single curve, affirming the applicability of the universal scaling law even for the circuit with finite $m$. As we increase $m$, the collapsed curve $\tilde{x}_{\mathrm{dec}}(\tilde{t})$ approaches the analytical result.

Note that slight deviations exist in the rescaled spreading fronts $\tilde{x}_{\mathrm{dec}}(\tilde{t})$ for large $p$ values when $m=1$. In the circuits with finite $m$, local noise still selectively removes short-range correlations ${I(A:B)}$. However, it can reduce ${I(A:BC)}$ for a small amount since the correlation is scrambled into a finite number of qubits. This introduces higher-order effects of $p$ on ${I(A:BC)}$, leading to small deviations in $\tilde{x}_{\mathrm{dec}}(\tilde{t})$. Fig. 3(b) shows that such higher-order effects are suppressed as we decrease the $p$ or increase $m$.

Discussion.—Rapid spreading of CMI shows that causally separated regions become conditionally dependent. However, it is important to remark that although the non-trivial correlation extends far beyond the lightcone, our findings do not conflict with the Lieb-Robinson bound or the principles of causality Lieb and Robinson (1972); Bravyi et al. (2006); Chen et al. (2023), as local observers in $A$ and $C$ cannot detect $I(A:C|B)$.

Meanwhile, the divergence of $x_{\mathrm{dec}}(t)$ as $t\rightarrow t_{c}$ indicates that CMI spreads into the entire system. We emphasize that this pervasive conditional dependence implies that the multipartite quantum correlation extends throughout the entire system. Specifically, at the critical timestep, we show in Ref.  that no matter how largely separated $A$ and $C$ are, measuring $B$ and heralding the measurement outcome to $A$ and $C$ generate Bell pairs shared by $A$ and $C$. Here, CMI acts as an upper bound of the number of generatable Bell pairs.

This result bears similarities to entanglement swap** Bennett et al. (1993), as it also generates entanglement between two distant regions through measurement and heralding the outcome. However, a fundamental difference exists. In the entanglement-swap** protocol, two separate Bell pairs between $A$, $B$, and $B$, $C$ are prepared, and the measurement on $B$ generates entanglement between $A$ and $C$. In our case, however, we have inseparable multipartite correlations among $A$, $B$, and $C$, and the measurement on $B$ reduces them into entanglement between $A$ and $C$ .

Finally, we remark that prior studies have shown that there exist constant depth circuits acting on a product state that generate CMI between arbitrarily large distant regions Zou and Haah (2016); Williamson et al. (2019); Kato and Brandão (2020); Kim et al. (2023). Our findings further extend this understanding by demonstrating that the formation of long-distance conditional dependence also emerges within noisy random circuits without designing a sophisticated circuit to achieve it.

in 1,…,0 See pages \x, of CMI_SM.pdf