Entropic Quantum Central Limit Theorem
and Quantum Inverse Sumset Theorem

Kaifeng Bu Harvard University, Cambridge, MA 02138, USA [email protected] , Weichen Gu Dept. of Mathematics and Statistics, Univ. of New Hampshire, Durham, NH 03824, USA [email protected] and Arthur Jaffe Harvard University, Cambridge, MA 02138, USA [email protected]

Abstract.

We establish an entropic, quantum central limit theorem and quantum inverse sumset theorem in discrete-variable quantum systems describing qudits or qubits. Both results are enabled by using our recently-discovered quantum convolution. We show that the exponential rate of convergence of the entropic central limit theorem is bounded by the magic gap. We also establish an “quantum, entropic inverse sumset theorem,” by introducing a quantum doubling constant. Furthermore, we introduce a “quantum Ruzsa divergence”, and we pose a conjecture called “convolutional strong subaddivity,” which leads to the triangle inequality for the quantum Ruzsa divergence. A byproduct of this work is a magic measure to quantify the nonstabilizer nature of a state, based on the quantum Ruzsa divergence.

1. Introduction

In this paper, we generalize some central results in modern, classical Fourier analysis to a quantum framework. This new framework is based on a quantum convolution that we introduced earlier in [1, 2]. In Theorems 12 and 13 we prove an entropic, quantum, central-limit theorem (entropic q-CLT) for discrete-variable (DV) quantum systems. We introduce a quantum doubling constant in Definition 20, which enables us to prove an entropic, quantum inverse-sumset in Theorem 22. We also introduce a quantum version of the Ruzsa divergence in Definition 26. We prove a number of its properties in Theorem 28 and provide a characterization of stabilizer states via quantum Ruzsa divergence in Proposition 35.

1.1. Background

The central limit theorem (CLT) is a fundamental result in probability theory. Given independent, identically distributed random variables $\mathcal{X}_{i}$ with zero mean, and finite variance $\sigma$ , the normalized sum $\mathcal{Z}_{N}=\frac{\sum^{N}_{i=1}\mathcal{X}_{i}}{\sqrt{N}}$ converges to a Gaussian random variable $\mathcal{Z}$ with the same mean value and variance. The study of entropic central limit theorem has a long history, back to the work of Linnik [3], where the entropy $h(\mathcal{Z}_{N})=-\int f_{\mathcal{Z}_{N}}\log f_{\mathcal{Z}_{N}}$ for the probability density function $f_{\mathcal{Z}_{N}}$ . Barron later showed that $h(\mathcal{Z}_{N})$ converges to the $h(\mathcal{Z})$ as $N\to\infty$ , where $\mathcal{Z}$ is the correpsonding Gaussian random variable [4]. Furthermore, the rate of convergence in the entropic central limit theorem has attracted much attention [5, 6, 7]. In the case of discrete random variables, the entropic CLT has also been considered by Gavalakis and Kontoyiannis [8], based on the normalized sum of $n$ independent and identically distributed lattice random variables and an appropriately discretized Gaussian.

In the case of continuous variable (CV) quantum systems, various central limit theorems also have an interesting history including Cushen and Hudson [9], and related work of Hepp and Lieb [10, 11]. Many other quantum or noncommutative versions of the central limit theorem appeared later, see [12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28]. For the entropic q-CLT in CV systems, the rate of convergence was given for an $m$ -mode quantum state under some technical assumption [26].

However, little is known about the entropic q-CLT for qudits or qubits. This is one of the foci of the current paper, which is based on the quantum convolution recently proposed by the authors [1, 2, 29]. This quantum convolutional framework provides a new method to understand and study stabilizer states. Stabilizer states were first introduced by Gottesman [30], and now have many applications including quantum error correction codes [31, 32], and the classical simulation of quantum circuits, known as Gottesman-Knill theorem [33]. These applications indicate that nonstabilizer states and circuits are necessary to achieve the quantum computational advantage. Later, the extension of the Gottesman-Knill theorem beyond stabilizer circuits was further studied [34, 35, 36, 37, 38, 39, 40, 41, 42, 43]. The term “magic” was introduced by Bravyi and Kitaev [44] to express the property that a state is not a stabilizer.

One consequence of our convolutional framework is a quantum central limit theorem for DV quantum systems. This means that repeated quantum convolution with any given, mean-zero state converges to a stabilizer state. Therefore we identify the set of stabilizer states as the set of “quantum-Gaussian” states. Some convolutional inequalities for the quantum entropies and other information measures have also been proved, with stabilizer states being the only extremizers of these inequalities [1, 2, 45]. Furthermore, based on purity invariance of stabilizer states under quantum convolution, a convolution-swap test has been proposed to determine whether a state is a stabilizer [29]. This can be regarded as a quantum-state version of linearity testing.

Finally we comment that (to our knowledge) the only earlier proposal for a good convolution for qudit states was given by Audenaert, Datta, and Ozols [46] and studied by Carlen, Lieb, and Loss [47]. Their convolution is a modified convex combination of the input states $\rho$ and $\sigma$ , namely $\lambda\rho+(1-\lambda)\sigma-i\sqrt{\lambda(1-\lambda)}\,[\rho,\sigma]$ . Unfortunately, this convolution does not lead to an interesting central limit theorem.

In the setting of free probability theory, a free convolution was introduced by Voiculescu; this also leads to a free central limit theorem [48]. But this framework is different from what we consider here.

A related problem for the classical CLT is the inverse sumset theory, which is an important topic in additive combinatorics. The inverse sumset theory, also known as Freiman-Ruzsa inverse sumset theory, explores properties of sets $A$ such that the size (or the Shannon entropy) of $A+A$ is close to that of $A$ . In this theory, the doubling constant $|A+A|/|A|$ and its entropic analog $\exp(H(\mathcal{X}+\mathcal{X}^{\prime})-H(\mathcal{X}))$ with i.i.d. copies of random variables $\mathcal{X}$ and $\mathcal{X}^{\prime}$ , have been used to investigate the properties of the subset $A$ and the random variable $X$ [49, 50, 51, 23, 52, 53]. For example, by Tao’s work [50], given i.i.d. copies $\mathcal{X}$ and $\mathcal{X}^{\prime}$ of a discrete random variables, if $\exp(H(\mathcal{X}+\mathcal{X}^{\prime})-H(\mathcal{X}))$ is small, then the distribution of $\mathcal{X}$ is close to the uniform distribution on a generalized arithmetic progression. Recently, there is a breakthrough in this theory by Gowers, Green, Manners, and Tao, which proves the Marton’s conjecture (or polynomial Freiman–Ruzsa conjecture) [53]. These results involved an important measure, called Ruzsa distance [50], which was later generalized to Ruzsa divergence [54]. The properties of Ruzsa distance depends on some inequalities about the sum of sets or random variables, known as sumset inequalities. They also provide some connection between the size (or the Shannon entropy) of the sumset $A+B$ (or random variables $\mathcal{X}$ and $\mathcal{Y}$ ) and that of sets $A$ and $B$ (or $\mathcal{X}$ and $\mathcal{Y}$ ) [55, 56, 50, 57, 54].

In this paper, we generalize many of these ideas to a DV quantum setting, and find that the quantum features will lead to some new phenomenon.

1.2. Outline and summary of our main results

1.

In §3, we establish an entropic q-CLT for DV quantum systems. We show that the quantum relative entropy between the $n$ -th repetition of our quantum convolution and the mean state converges to zero at an exponential rate. This rate is bounded from below by the “magic gap” defined by the state.
2.

In §4, we establish an entropic, quantum inverse-sumset theorem. We study the quantum-doubling constant for a state $\rho$ . This is the difference between the von Neumann entropy $S(\rho\boxtimes\rho)$ of the self-convolution of $\rho$ and the entropy $S(\rho)$ of the state itself. The quantum-doubling constant and the magic gap provide an upper bound on the quantum relative entropy between the given state and its mean state.
3.

In §5, we generalize the classical Ruzsa divergence to a quantum divergence. We use quantum subset inequalities to establish some properties of the quantum Ruzsa divergence, which also provide a good characterization of stabilizer states. While sub-additivity holds for the classical convolution, i.e., $H(\mathcal{X}+\mathcal{Y})\leqslant H(\mathcal{X})+H(\mathcal{Y})$ for random variables $\mathcal{X}$ and $\mathcal{Y}$ [58, 49], we prove that quantum features prevent sub-additivity from holding in general. We conjecture a convolutional strong subaddivity of the quantum entropy, which is different from the strong sub-additivity proved by Lieb and Ruskai [59]. We prove this convolutional strong subadditivity for two special cases. Moreover, in §5.2, we define a magic measure via the quantum Ruzsa divergence to quantify the magic of states.

2. Background

Fix natural numbers $n$ (the number of qudits) and $d$ (the degree of each qudit) and study the Hilbert space $\mathcal{H}^{\otimes n}$ , where $\mathcal{H}=\mathbb{C}^{d}$ . Consider an orthonormal basis $\set{\ket{k}:k\in\mathbb{Z}_{d}}$ for the Hilbert space $\mathcal{H}$ ; here $\mathbb{Z}_{d}$ denotes the cyclic group of order $d$ . One calls these vectors the computational basis. The Pauli operators $X$ and $Z$ are defined by

(1)

X:|k\rangle\mapsto|k+1\rangle\;,\qquad Z:|k\rangle\mapsto\omega^{k}_{d}|k% \rangle\;,\qquad\forall k\in\mathbb{Z}_{d}\;,

where $\omega_{d}=\exp(2\pi i/d)$ is the primitive $d$ -th root of unity. We restrict $d$ to be prime in order to define our quantum convolution.

The local Weyl operators (or generalized Pauli operators) are

w(p,q)=\zeta^{-pq}\,Z^{p}X^{q}\;,\quad\text{where}\quad\zeta=\left\{\begin{% aligned} &\omega^{(d+1)/2}_{d}\;,&\text{for }d\text{ odd}\hfill\\ &e^{i\pi/2}\;,&\text{for }d=2\hfill\end{aligned}\right.\;.

For the $n$ -qudit system on $\mathcal{H}^{\otimes n}$ , the Weyl operators are defined as

(2)

\displaystyle w(\vec{p},\vec{q})=w(p_{1},q_{1})\otimes...\otimes w(p_{n},q_{n}% )=w(-\vec{p},-\vec{q})^{\dagger}\;,

with $\vec{p}=(p_{1},p_{2},...,p_{n})\in\mathbb{Z}^{n}_{d}$ , $\vec{q}=(q_{1},...,q_{n})\in\mathbb{Z}^{n}_{d}$ . Let us denote $\mathcal{P}_{n}$ as the group generated by the Weyl operators and phase $\zeta$ . The Weyl operators are an orthonormal basis for the space of linear operators on $\mathcal{H}^{\otimes n}$ with respect to the inner product

(3)

\langle A,B\rangle=\frac{1}{d^{n}}\operatorname{Tr}\left[A^{\dagger}B\right]\;.

Denote $V^{n}:=\mathbb{Z}^{n}_{d}\times\mathbb{Z}^{n}_{d}$ ; this represents the phase space for $n$ -qudit systems, as was studied in [60]. Let $D(\mathcal{H}^{\otimes n})$ denote the set of all quantum states on $\mathcal{H}^{\otimes n}$ , namely positive matrices with unit trace.

Definition 1 (Characteristic function).

The characteristic function $\Xi(\vec{p},\vec{q})$ of a state $\rho\in D(\mathcal{H}^{\otimes n})$ is the coefficient of $\rho$ in the Weyl basis,

\Xi_{\rho}(\vec{p},\vec{q})=\operatorname{Tr}\left[\rho w(-\vec{p},-\vec{q})% \right]\;,\quad\text{and}\quad\rho=\frac{1}{d^{n}}\sum_{(\vec{p},\vec{q})\in V% ^{n}}\Xi_{\rho}(\vec{p},\vec{q})\ w(\vec{p},\vec{q})\;.

The process of taking characteristic functions is the quantum Fourier transform that we consider. More details about the properties of the characteristic functions can be found in [60, 61, 2]. In this work, we also use $\Xi_{\rho}(\vec{x})$ with $\vec{x}=(\vec{p},\vec{q})\in V^{n}$ and the expectation

\mathbb{E}_{k_{i}\in\mathbb{Z}_{d}}(\ \cdot\ ):=\frac{1}{d}\sum_{k_{i}\in% \mathbb{Z}_{d}}(\ \cdot\ )\;.

Definition 2 (Stabilizer states [62, 30]).

A pure stabilizer state $|\psi\rangle\!\langle\psi|$ for an $n$ -qudit system is the projection onto a stabilizer vector $\ket{\psi}$ , namely a common unit eigenvector of an abelian (stabilizer) subgroup of the Weyl operators of size $d^{n}$ . If the generators of the stabilizer group are $\set{g_{1},...,g_{n}}_{i\in[n]}$ with $g_{i}\in\mathcal{P}_{n}$ , then

\displaystyle|\psi\rangle\!\langle\psi|=\Pi^{n}_{i=1}\mathbb{E}_{k_{i}\in% \mathbb{Z}_{d}}g^{k_{i}}_{i}\;.

A general stabilizer state is a mixed state $\rho$ obtained as a convex linear combination of pure stabilizer states; we denote the set of stabilizer states by STAB.

Definition 3 (Minimal stabilizer-projection state).

A quantum state $\rho$ is a minimal stabilizer-projection state (MSPS) associated with an abelian subgroup generated by $\set{g_{i}}_{i\in[r]}$ with $g_{i}\in\mathcal{P}_{n}$ , if it has the following form

\displaystyle\rho=\frac{1}{d^{n-r}}\Pi^{r}_{i=1}\mathbb{E}_{k_{i}\in\mathbb{Z}% _{d}}g_{i}^{k_{i}}\;.

For example, in an $n$ -qudit system and with the abelian group with the generators $\set{Z_{1},...,Z_{n-1}}$ , the states $\set{\frac{1}{d}|\vec{j}\rangle\!\langle\vec{j}|\otimes I}_{\vec{j}\in\mathbb{% Z}^{n-1}_{d}}$ are MSPS.

Definition 4 (Clifford unitary).

An $n$ -qudit unitary $U$ is a Clifford unitary, if $Uw(\vec{x})U^{\dagger}$ is also a Weyl operator up to a phase for any $\vec{x}\in V^{n}$ .

It is easy to see that Clifford unitaries map stabilizer states to stabilizer states.

Definition 5 (Mean state (MS) [1, 2]).

Given an $n$ -qudit state $\rho$ , the mean state $\mathcal{M}(\rho)$ is the operator with the characteristic function:

(4)

\displaystyle\Xi_{\mathcal{M}(\rho)}(\vec{x}):=\left\{\begin{aligned} &\Xi_{% \rho}(\vec{x}),&&|\Xi_{\rho}(\vec{x})|=1,\\ &0,&&|\Xi_{\rho}(\vec{x})|<1.\end{aligned}\right.

The mean state $\mathcal{M}(\rho)$ is a stabilizer state.

In addition, $\mathcal{M}(\rho)$ has a stabilizer group, i.e., the abelian group generated by the Pauli operator $w(\vec{x})$ such that $w(\vec{x})\mathcal{M}(\rho)w(\vec{x})^{\dagger}=\mathcal{M}(\rho)$ . For simplicity, we denote it as $G_{\rho}$ .

Definition 6 (Zero-mean state [1, 2]).

A given $n$ -qudit state $\rho$ has zero-mean, if the characteristic function of $\mathcal{M}(\rho)$ takes values in $\set{0,1}$ .

Note that, if $\rho$ is not a zero-mean state, there exists a Weyl operator $w(\vec{x})$ such that $w(\vec{x})\rho w(\vec{x})^{\dagger}$ is a zero-mean state [2].

Definition 7 (Magic gap [1, 2]).

Given an $n$ -qudit state $\rho\in\mathcal{D}(\mathcal{H}^{\otimes n})$ for any integer $d\geqslant 2$ , the magic gap of $\rho$ is

\displaystyle MG(\rho)=1-\max_{\vec{x}\in\text{Supp}(\Xi_{\rho}):|\Xi_{\rho}(% \vec{x})|\neq 1}|\Xi_{\rho}(\vec{x})|\;.

If $\set{\vec{x}\in\text{Supp}(\Xi_{\rho}):}{\Xi_{\rho}(\vec{x})|\neq 1}=\emptyset$ , define $MG(\rho)=0$ , i.e., there is no gap on the support of the characteristic function.

Definition 8 (Quantum convoltuion [1, 2]).

Let $s^{2}+t^{2}\equiv 1\mod d$ , with $s,t\neq 0$ , and let $U_{s,t}$ be the unitary

(5)

\displaystyle U_{s,t}=\sum_{\vec{i},\vec{j}\in\mathbb{Z}^{n}_{d}}|s\vec{i}+t% \vec{j}\rangle\langle\vec{i}|\otimes|-t\vec{i}+s\vec{j}\rangle\langle\vec{j}|\;,

acting on the $2n$ -qudit sytems $\mathcal{H}_{A}\otimes\mathcal{H}_{B}$ with $\mathcal{H}_{A}=\mathcal{H}_{B}=\mathcal{H}^{\otimes n}$ , and the vector $|\vec{i}\rangle=|i_{1}\rangle\otimes\cdots\otimes|i_{n}\rangle\in\mathcal{H}^{% \otimes n}$ . The convolution of two $n$ -qudit states $\rho$ and $\sigma$ is

(6)

\displaystyle\rho\boxtimes_{s,t}\sigma=\operatorname{Tr}_{B}\left[U_{s,t}(\rho% \otimes\sigma)U^{\dagger}_{s,t}\right].

2.1. Basic properties of entropy

We review some basic properties of quantum entropy and relative entropy.

Definition 9 (Quantum entropy).

Given a quantum state $\rho$ , the von Neuman entropy is

(7)

\displaystyle S(\rho):=-\operatorname{Tr}\left[\rho\log\rho\right].

Given a parameter $\alpha\in[0,+\infty]$ , the quantum Rényi entropy is

(8)

\displaystyle S_{\alpha}(\rho):=\frac{1}{1-\alpha}\log\operatorname{Tr}\left[% \rho^{\alpha}\right].

Note that $\lim_{\alpha\to 1}S_{\alpha}(\rho)=S(\rho)$ . Also $S_{\infty}(\rho)=\lim_{\alpha\to\infty}S_{\alpha}(\rho)=-\log\lambda_{\max}$ , where $\lambda_{\max}$ is the largest eigenvalue of $\rho$ .

Definition 10 (Quantum relative entropy).

The relative entropy of $\rho$ with respect to $\sigma$ is

(9)

\displaystyle D(\rho||\sigma):=\operatorname{Tr}\left[\rho(\log\rho-\log\sigma% )\right]\;.

Given a parameter $\alpha\in[0,+\infty]$ , the quantum Rényi relative entropy $D_{\alpha}$ is

\displaystyle D_{\alpha}(\rho||\sigma):=\frac{1}{\alpha-1}\log\operatorname{Tr% }\left[\left(\sigma^{\frac{1-\alpha}{2\alpha}}\rho\sigma^{\frac{1-\alpha}{2% \alpha}}\right)^{\alpha}\right].

Note that $\lim_{\alpha\to 1}D_{\alpha}(\rho||\sigma)=D(\rho||\sigma)$ , and

D_{\infty}(\rho||\sigma)=\lim_{\alpha\to+\infty}D_{\alpha}(\rho||\sigma)=\min% \log\set{\lambda:\rho\leqslant\lambda\sigma}\;.

One fundamental result is that the mean state $\mathcal{M}(\rho)$ is the closest MSPS to the given state $\rho$ by using the quantum relative entropy as a distance measure.

Lemma 11 ([1, 2]).

Given an $n$ -qudit state $\rho$ , we have

\displaystyle\min_{\sigma\in MSPS}D_{\alpha}(\rho||\sigma)=D_{\alpha}(\rho||% \mathcal{M}(\rho))=S_{\alpha}(\mathcal{M}(\rho))-S_{\alpha}(\rho)\;,

for any $\alpha\geqslant 1$ .

3. The entropic, quantum, central-limit theorem

Here we give our main results on the entropic q-CLT. In the earlier work we established and studied a quantum, central limit theorem [1, 2, 29], where the exponential rate of convergence in $L_{2}$ norm is bounded below by the magic gap. Here let us consider the entropic q-CLT. There are two entropic measures in the study of the entropic q-CLT, one is the quantum relative entropy $D(\rho||\mathcal{M}(\rho))$ , and the other is the difference of quantum entropies $S(\mathcal{M}(\rho))-S(\rho)$ . These two measures are shown to be equivalent by Lemma 11. Hence, we will use these entropic measures to establish an entropic q-CLT.

Theorem 12 (Entropic q-CLT).

Given an $n$ -qudit state $\rho$ , the quantum entropy of the $N^{\rm th}$ repetition of the quantum convolution $S(\boxtimes^{N}\rho)$ increases monotonically to $S(\mathcal{M}(\rho))$ , namely

(10)

\displaystyle S(\boxtimes^{N}\rho)\nearrow S(\mathcal{M}(\rho))\;,\text{ as }~% {}N\to\infty\;.

This is equivalent to

(11)

\displaystyle D(\boxtimes^{N}\rho||\mathcal{M}(\boxtimes^{N}\rho))\to 0\;,% \text{ as }~{}N\to\infty\;.

For a state $\rho$ with zero-mean,

(12)

\displaystyle D(\boxtimes^{N}\rho||\mathcal{M}(\rho))\to 0\;,\text{ as }~{}N% \to\infty\;.

Proof.

The monotonicity of the quantum entropy $S(\boxtimes^{N}\rho)$ with respect to $N$ in the q-CLT has been proved by the authors; see Theorem 57 in [1] and also [2, 29], namely

\displaystyle S(\boxtimes^{N}\rho)\leqslant S(\boxtimes^{N+1}\rho)\;,\ \forall N% \geqslant 1\;.

Moreover, as $\boxtimes^{N}\rho\to\mathcal{M}(\rho)\text{ for}~{}~{}N\to\infty$ , by the continuity of quantum entropy, we have

\displaystyle S(\boxtimes^{N}\rho)\to S(\mathcal{M}(\rho))\;,\text{ as }~{}~{}% N\to\infty\;.

Since $\mathcal{M}(\boxtimes^{N}\rho)$ is equivalent to $\mathcal{M}(\rho)$ up to a Clifford unitary, then $S(\mathcal{M}(\boxtimes^{N}\rho))=S(\mathcal{M}(\rho))$ . Thus, by the equality $D(\rho||\mathcal{M}(\rho))=S(\mathcal{M}(\rho))-S(\rho)$ in Lemma 11, we have

	$\displaystyle D(\boxtimes^{N}\rho\|\|\mathcal{M}(\boxtimes^{N}\rho))=$	$\displaystyle S(\mathcal{M}(\boxtimes^{N}\rho))-S(\boxtimes^{N}\rho)$
	$\displaystyle=$	$\displaystyle S(\mathcal{M}(\rho))-S(\boxtimes^{N}\rho)$
	$\displaystyle\to$	$\displaystyle 0,~{}~{}\text{as}~{}~{}N\to\infty.$

For a zero-mean state $\rho$ , $\mathcal{M}(\boxtimes^{N}\rho)=\mathcal{M}(\rho)$ . Hence, we have

\displaystyle D(\boxtimes^{N}\rho||\mathcal{M}(\rho))=D(\boxtimes^{N}\rho||% \mathcal{M}(\boxtimes^{N}\rho))\to 0,~{}~{}\text{as}~{}~{}N\to\infty.

∎

We can also provide an upper bound on the rate of convergence in the entropic q-CLT in terms of magic gap, where the rate of convergence is exponentially small with respect to the number of convolutions.

Theorem 13 (Rate of convergence).

Given an $n$ -qudit state $\rho$ , the quantum relative entropy of $\boxtimes^{N}\rho$ with respect to the mean state $\mathcal{M}(\boxtimes^{N}\rho)$ has the following bound,

	$\displaystyle D(\boxtimes^{N}\rho\|\|\mathcal{M}(\boxtimes^{N}\rho))=$	$\displaystyle S(\mathcal{M}(\rho))-S(\boxtimes^{N}\rho)$
(13)		$\displaystyle\leqslant$	$\displaystyle\log\left[1+(1-MG(\rho))^{2N-2}\left(\operatorname{Tr}\left[\rho^% {2}\right]R(\rho)-1\right)\right]$
	$\displaystyle\leqslant$	$\displaystyle(1-MG(\rho))^{2N-2}\left(\operatorname{Tr}\left[\rho^{2}\right]R(% \rho)-1\right)\;,$
	$\displaystyle\to$	$\displaystyle 0,~{}\text{as}~{}N\to\infty,$

where $R(\rho)$ is the rank of the state $\mathcal{M}(\rho)$ .

Proof.

Let us assume that $G_{\rho}$ is the stabilizer group of $\mathcal{M}(\rho)$ , then $R(\rho)=\frac{d^{n}}{|G_{\rho}|}$ , and thus $S(\mathcal{M}(\rho))=\log R(\rho)=\log\frac{d^{n}}{|G_{\rho}|}$ . For simplicity, let us take $\lambda=(1-MG(\rho))^{2}=\max_{\vec{x}\in\text{Supp}(\Xi_{\rho}):|\Xi_{\rho}(% \vec{x})|\neq 1}|\Xi_{\rho}(\vec{x})|^{2}$ . Hence

		$\displaystyle D(\boxtimes^{N}\rho\|\|\mathcal{M}(\boxtimes^{N}\rho))$
	$\displaystyle=$	$\displaystyle S(\mathcal{M}(\rho))-S(\boxtimes^{N}\rho)$
	$\displaystyle\leqslant$	$\displaystyle S(\mathcal{M}(\rho))-S_{2}(\boxtimes^{N}\rho)$
	$\displaystyle=$	$\displaystyle\log\frac{d^{n}}{\|G_{\rho}\|}+\log\left(\frac{\|G_{\rho}\|}{d^{n}}+% \frac{1}{d^{n}}\sum_{\vec{x}\notin G}\|\Xi_{\boxtimes^{N}\rho}(\vec{x})\|^{2}\right)$
	$\displaystyle\leqslant$	$\displaystyle\log\frac{d^{n}}{\|G_{\rho}\|}+\log\left(\frac{\|G_{\rho}\|}{d^{n}}+% \frac{\lambda^{N-1}}{d^{n}}\sum_{\vec{x}\notin G}\|\Xi_{\rho}(\vec{x})\|^{2}\right)$
	$\displaystyle=$	$\displaystyle\log\frac{d^{n}}{\|G_{\rho}\|}+\log\left[\frac{\|G_{\rho}\|}{d^{n}}+% \lambda^{N-1}\left(\operatorname{Tr}\left[\rho^{2}\right]-\frac{\|G_{\rho}\|}{d^% {n}}\right)\right]$
	$\displaystyle=$	$\displaystyle\log\left[1+\lambda^{N-1}\left(\operatorname{Tr}\left[\rho^{2}% \right]\frac{d^{n}}{\|G_{\rho}\|}-1\right)\right],$

where the second line comes from Lemma 11, the third line comes from the monotonicity of Rényi entropy $S(\rho)\geqslant S_{2}(\rho)$ , the fifth line comes from the definition of $\lambda$ , and the sixth line comes from the fact that

\displaystyle\operatorname{Tr}\left[\rho^{2}\right]=\frac{1}{d^{n}}\sum_{\vec{% x}}|\Xi_{\rho}(\vec{x})|^{2}=\frac{|G_{\rho}|}{d^{n}}+\frac{1}{d^{n}}\sum_{% \vec{x}\notin G_{\rho}}|\Xi_{\rho}(\vec{x})|^{2}.

∎

By the entropic q-CLT, we have the following q-CLT based on trace distance as a corollary.

Corollary 14.

Given an $n$ -qudit state $\rho$ with zero mean, we have

(14)

\displaystyle\left\lVert\boxtimes^{N}\rho-\mathcal{M}(\rho)\right\rVert_{1}% \leqslant\sqrt{2}(1-MG(\rho))^{N-1}\sqrt{\left(\operatorname{Tr}\left[\rho^{2}% \right]R(\rho)-1\right)},\quad\text{as}~{}~{}N\to\infty.

Proof.

This is a direct corollary of the quantum Pinsker inequality $\frac{1}{2}\left\lVert\rho-\sigma\right\rVert^{2}_{1}\leqslant D(\rho||\sigma)$ , and the Theorem 13. ∎

Refer to caption — Figure 1. A diagram of the repeated quantum convolution in the q-CLT.

Now, let us generalize the entropic q-CLT from von Neumann entropy to Rényi entropy.

Theorem 15 (Rényi entropic q-CLT via magic gap).

Given an $n$ -qudit state $\rho$ and any parameter $\alpha\in[1,\infty]$ , the $\alpha$ -quantum Rényi relative entropy of $\boxtimes^{N}\rho$ with respect to the mean state $\mathcal{M}(\boxtimes^{N}\rho)$ satisfies the bound,

	$\displaystyle D_{\alpha}(\boxtimes^{N}\rho\|\|\mathcal{M}(\boxtimes^{N}\rho))=$	$\displaystyle S_{\alpha}(\mathcal{M}(\rho))-S_{\alpha}(\boxtimes^{N}\rho)$
(15)		$\displaystyle\leqslant$	$\displaystyle\log\left(1+(1-MG(\rho))^{N-1}R(\rho)\sqrt{\operatorname{Tr}\left% [\rho^{2}\right]-\frac{1}{R(\rho)}}\right)$
	$\displaystyle\leqslant$	$\displaystyle(1-MG(\rho))^{N-1}R(\rho)\sqrt{\operatorname{Tr}\left[\rho^{2}% \right]-\frac{1}{R(\rho)}}$
	$\displaystyle\to$	$\displaystyle 0\;,~{}\text{ as}~{}~{}N\to\infty.$

Proof.

Based on the monotonicity of Rényi relative entropy $D_{\alpha}\leqslant D_{\infty}$ for any $\alpha\geqslant 0$ , we only need to prove the statement for the max-relative entropy $D_{\infty}$ . For simplicity, let us take $\lambda=(1-MG(\rho))^{2}=\max_{\vec{x}\in\text{Supp}(\Xi_{\rho}):|\Xi_{\rho}(% \vec{x})|\neq 1}|\Xi_{\rho}(\vec{x})|^{2}$ . First, we have

		$\displaystyle\|\bra{\psi}\rho\ket{\psi}\|$
	$\displaystyle\leqslant$	$\displaystyle\|\bra{\psi}\rho-\mathcal{M}(\rho)\ket{\psi}\|+\|\bra{\psi}\mathcal{% M}(\rho)\ket{\psi}\|$
	$\displaystyle\leqslant$	$\displaystyle\sqrt{\operatorname{Tr}\left[(\rho-\mathcal{M}(\rho))^{2}\right]}% +\frac{1}{R(\rho)}$
	$\displaystyle=$	$\displaystyle\sqrt{\frac{1}{d^{n}}\sum_{\vec{x}\notin G_{\rho}}\|\Xi_{\rho}(% \vec{x})\|^{2}}+\frac{1}{R(\rho)},$

where the second line comes from the triangle inequality, the third line comes from the Cauchy-Schwarz inequality and the last line comes from the fact that $\rho-\mathcal{M}(\rho)=\frac{1}{d^{n}}\sum_{\vec{x}\notin G_{\rho}}\Xi_{\rho}(% \vec{x})w(\vec{x})$ . Hence, we have

		$\displaystyle\lambda_{\max}(\boxtimes^{N}\rho)$
	$\displaystyle\leqslant$	$\displaystyle\sqrt{\frac{1}{d^{n}}\sum_{\vec{x}\notin G_{\rho}}\|\Xi_{\boxtimes% ^{N}\rho}(\vec{x})\|^{2}}+\frac{1}{R(\rho)}$
(16)		$\displaystyle\leqslant$	$\displaystyle\sqrt{\frac{\lambda^{N-1}}{d^{n}}\sum_{\vec{x}\notin G_{\rho}}\|% \Xi_{\rho}(\vec{x})\|^{2}}+\frac{1}{R(\rho)}$
	$\displaystyle=$	$\displaystyle\lambda^{(N-1)/2}\sqrt{\operatorname{Tr}\left[\rho^{2}\right]-% \frac{1}{R(\rho)}}+\frac{1}{R(\rho)},$

where the third line comes from the definition of $\lambda$ , and the last line comes from the fact that

\displaystyle\operatorname{Tr}\left[\rho^{2}\right]=\frac{1}{d^{n}}\sum_{\vec{% x}}|\Xi_{\rho}(\vec{x})|^{2}=\frac{|G_{\rho}|}{d^{n}}+\frac{1}{d^{n}}\sum_{% \vec{x}\notin G_{\rho}}|\Xi_{\rho}(\vec{x})|^{2}.

Hence,

		$\displaystyle D_{\infty}(\boxtimes^{N}\rho\|\|\mathcal{M}(\rho))$
	$\displaystyle=$	$\displaystyle S_{\infty}(\mathcal{M}(\rho))-S_{\infty}(\boxtimes^{N}\rho)$
	$\displaystyle=$	$\displaystyle\log R(\rho)-\log\frac{1}{\lambda_{\max}(\boxtimes^{N}\rho)}$
	$\displaystyle=$	$\displaystyle\log\left[R(\rho)\lambda_{\max}(\boxtimes^{N}\rho)\right]$
	$\displaystyle\leqslant$	$\displaystyle\log\left[1+\lambda^{(N-1)/2}R(\rho)\sqrt{\operatorname{Tr}\left[% \rho^{2}\right]-\frac{1}{R(\rho)}}\right]$
	$\displaystyle\leqslant$	$\displaystyle\lambda^{(N-1)/2}R(\rho)\sqrt{\operatorname{Tr}\left[\rho^{2}% \right]-\frac{1}{R(\rho)}},$
	$\displaystyle\to$	$\displaystyle 0,~{}~{}\text{as}~{}~{}N\to\infty.$

where the second line comes from the Lemma 11, and the fifth line comes from the inequality (16). ∎

3.1. Entropic q-CLT for qubit systems

Now, let us consider the entropic q-CLT for the qubit systems, where we need to change the definition of quantum convolution to the one used in [29].

Definition 16 (Key Unitary).

The key unitary $V$ for $K$ quantum systems, with each system containing $n$ qubits, is

(17)

\displaystyle V:=U^{\otimes n}=U_{1,n+1,...,(K-1)n+1}\otimes U_{2,n+2,...,(K-1% )n+2}\otimes...\otimes U_{n,2n,...,Kn}.

Here $U$ is a $K$ -qubit unitary constructed using CNOT gates:

(18)

\displaystyle U:=\left(\prod^{K}_{j=2}CNOT_{j\to 1}\right)\left(\prod^{K}_{i=2% }CNOT_{1\to i}\right)\;,

and $CNOT_{2\to 1}\ket{x}\ket{y}=\ket{x+y}\ket{y}$ for any $x,y\in\mathbb{Z}_{2}$ .

Definition 17 (Convolution of multiple states).

Given $K$ states $\rho_{1},\rho_{2},...,\rho_{K}$ , each with $n$ -qubits, the multiple convolution $\boxtimes_{K}$ of $\rho_{1},\rho_{2},...,\rho_{K}$ maps to an $n$ -qubit state:

(19)

\displaystyle\boxtimes_{K}(\rho_{1},\rho_{2},...,\rho_{K})=\boxtimes_{K}(% \otimes^{K}_{i=1}\rho_{i})=\operatorname{Tr}_{1^{c}}\left[V\otimes^{K}_{i=1}% \rho_{i}V^{\dagger}\right]\;.

Here $V$ is the key unitary in Definition 16, and $\operatorname{Tr}_{1^{c}}\left[\cdot\right]$ denotes the partial trace taken on the subsystem $2,3...,K$ , i.e., $\operatorname{Tr}_{1^{c}}\left[\cdot\right]=\operatorname{Tr}_{2,3,...,K}\left% [\cdot\right]$ .

Theorem 18 (Entropic q-CLT for qubits).

Given an $n$ -qubit state $\rho$ , and $N=2K+1$ for any integer $K\geqslant 1$ ,

	$\displaystyle D(\boxtimes_{N}\rho\|\|\mathcal{M}(\boxtimes_{N}\rho))=$	$\displaystyle S(\mathcal{M}(\rho))-S(\boxtimes_{N}\rho)$
(20)		$\displaystyle\leqslant$	$\displaystyle\log\left[1+(1-MG(\rho))^{2N-2}\left(\operatorname{Tr}\left[\rho^% {2}\right]R(\rho)-1\right)\right]$
	$\displaystyle\leqslant$	$\displaystyle(1-MG(\rho))^{2N-2}\left(\operatorname{Tr}\left[\rho^{2}\right]R(% \rho)-1\right)$
	$\displaystyle\to$	$\displaystyle 0,~{}~{}\text{as}~{}~{}N\to\infty.$

Proof.

The proof is the same as the qudit case. ∎

Theorem 19 (Rényi entropic q-CLT via magic gap).

Given an $n$ -qubit state $\rho$ , $N=2K+1$ for any integer $K\geqslant 1$ , and any parameter $\alpha\in[1,\infty]$ , we have

	$\displaystyle D_{\alpha}(\boxtimes_{N}\rho\|\|\mathcal{M}(\boxtimes_{N}\rho))=$	$\displaystyle S_{\alpha}(\mathcal{M}(\rho))-S_{\alpha}(\boxtimes_{N}\rho)$
(21)		$\displaystyle\leqslant$	$\displaystyle\log\left(1+(1-MG(\rho))^{N-1}R(\rho)\sqrt{\operatorname{Tr}\left% [\rho^{2}\right]-\frac{1}{R(\rho)}}\right)$
	$\displaystyle\leqslant$	$\displaystyle(1-MG(\rho))^{N-1}R(\rho)\sqrt{\operatorname{Tr}\left[\rho^{2}% \right]-\frac{1}{R(\rho)}},$
	$\displaystyle\to$	$\displaystyle 0,~{}~{}\text{as}~{}~{}N\to\infty.$

Proof.

The proof is the same as the qudit case. ∎

4. Entropic quantum inverse sumset theorem

In this section, we study the quantum inverse sumset theory in DV quantum system. For which, we need to define a quantum doubling constant, and then use it to quantify the distance between the given state and the stabilizer states.

Definition 20 (Quantum-doubling constant).

Given an $n$ -qudit state $\rho$ , the quantum-doubling constant is

(22)

\displaystyle\delta_{q}[\rho]=\exp(S(\rho\boxtimes\rho)-S(\rho)).

In general, the $\alpha$ -order quantum-doubling constant is

(23)

\displaystyle\delta_{q,\alpha}[\rho]=\exp(S_{\alpha}(\rho\boxtimes\rho)-S_{% \alpha}(\rho)).

Based on the definition, the quantum doubling constant is the entropy difference of the first step in the q-CLT. Moreover, for a pure state $\psi$ , the quantum-doubling constant $\delta[\psi]$ is equal to the magic entropy $ME(\psi)=S(\psi\boxtimes\psi)$ defined in [29] up to a logarithm.

Remark 21.

Similar to the quantum-doubling constant, we can also define the quantum-difference constant is

(24)

\displaystyle\delta^{-}_{q}[\rho]=\exp(S(\rho\boxminus\rho)-S(\rho))\;,

where $\rho\boxminus\rho=\operatorname{Tr}_{A}\left[U_{s,t}(\rho\otimes\rho)U^{% \dagger}_{s,t}\right]$ , i.e., the complementary channel of the quantum channel $\boxtimes$ .

In this work, we consider the following problem: given a quantum state $\rho$ , how could the quantum-doubling constant $\delta_{q}[\rho]$ tell the structure of the state $\rho$ , that is, how close the state is to the set of MSPS. We call this the quantum inverse sumset problem. Here, we focus on the pure state case, for which we have the following result.

Theorem 22 (Quantum inverse sumset theorem using magic gap).

Given an $n$ -qudit pure state $\psi$ ,

(1) $\delta_{q}[\psi]\geqslant 1$ , with equality iff $\psi\in\text{STAB}$ .

(2) If $1<\delta_{q}[\psi]\leqslant C$ , then

(25)

\displaystyle D(\psi||\mathcal{M}(\psi))\leqslant\frac{\log R(\psi)}{\log R(% \psi)-\log[1+\lambda(R(\psi)-1)]}\log C.

Proof.

(1) It comes directly from the entropy inequality for quantum convolution in [1, 2].

(2) First, by the monotonicity of relative entropy under the quantum channel, there exists some factor $\kappa_{\psi}\leqslant 1$ such that

\displaystyle D(\psi\boxtimes\psi||\mathcal{M}(\psi)\boxtimes\psi)=D(\psi||% \mathcal{M}(\psi))\kappa_{\psi}.

By Lemma 11 and the fact that $S(\mathcal{M}(\psi)\boxtimes\psi)=S(\mathcal{M}(\psi))$ , it can be rewritten as

\displaystyle S(\mathcal{M}(\psi))-S(\psi\boxtimes\psi)=(S(\mathcal{M}(\psi))-% S(\psi))\kappa_{\psi}.

Since $\delta_{q}[\psi]>1$ , $\psi$ is not a stabilizer state, and thus $\kappa_{\psi}<1$ . Hence

\displaystyle S(\psi\boxtimes\psi)-S(\psi)=(1-\kappa_{\psi})(S(\mathcal{M}(% \psi))-S(\psi))\;,

which implies that

\displaystyle D(\psi||\mathcal{M}(\psi))=\frac{1}{1-\kappa_{\psi}}\left[S(\psi% \boxtimes\psi)-S(\psi)\right]\leqslant\frac{1}{1-\kappa_{\psi}}\log C.

Now, let us provide an upper bound on the factor $\kappa_{\psi}$ using magic gap.

	$\displaystyle\kappa_{\psi}=\frac{D(\psi\boxtimes\psi\|\|\mathcal{M}(\psi))}{D(% \psi\|\|\mathcal{M}(\psi))}=\frac{S(\mathcal{M}(\psi))-S(\psi\boxtimes\psi)}{S(% \mathcal{M}(\psi))}$
	$\displaystyle\leqslant\frac{S(\mathcal{M}(\psi))-S_{2}(\psi\boxtimes\psi)}{S(% \mathcal{M}(\psi))},$

where the inequality comes from the fact that $S_{\alpha}$ is nonincreasing with respect to $\alpha$ .

Let us assume that $G_{\psi}$ is the stabilizer group of $\psi$ , then $R(\psi)=\frac{d^{n}}{|G_{\psi}|}$ , and thus $S(\mathcal{M}(\psi))=\log R(\psi)=\log\frac{d^{n}}{|G_{\psi}|}$ . Hence

		$\displaystyle S(\mathcal{M}(\psi))-S_{2}(\psi\boxtimes\psi)$
	$\displaystyle=$	$\displaystyle\log\frac{d^{n}}{\|G_{\psi}\|}+\log\left(\frac{\|G_{\psi}\|}{d^{n}}+% \frac{1}{d^{n}}\sum_{\vec{x}\notin G}\|\Xi_{\psi}(\vec{x})\|^{4}\right)$
	$\displaystyle\leqslant$	$\displaystyle\log\frac{d^{n}}{\|G_{\psi}\|}+\log\left(\frac{\|G_{\psi}\|}{d^{n}}+% \frac{\lambda}{d^{n}}\sum_{\vec{x}\notin G}\|\Xi_{\psi}(\vec{x})\|^{2}\right)$
	$\displaystyle=$	$\displaystyle\log\frac{d^{n}}{\|G_{\psi}\|}+\log\left[\frac{\|G_{\psi}\|}{d^{n}}+% \lambda\left(1-\frac{\|G_{\psi}\|}{d^{n}}\right)\right]$
	$\displaystyle=$	$\displaystyle\log\left[1+\lambda\left(\frac{d^{n}}{\|G_{\psi}\|}-1\right)\right],$

where the third line comes the from definition of $\lambda$ , and the fourth line comes from the fact that

(26)

\displaystyle 1=\operatorname{Tr}\left[\psi^{2}\right]=\frac{1}{d^{n}}\sum_{% \vec{x}}|\Xi_{\psi}(\vec{x})|^{2}=\frac{|G_{\psi}|}{d^{n}}+\frac{1}{d^{n}}\sum% _{\vec{x}\notin G_{\psi}}|\Xi_{\psi}(\vec{x})|^{2}

∎

Here we consider the inverse sumset theorem for the pure states. For mixed states, the above method cannot provide a good estimate for the property (2), which may require new techniques. We leave it for a future study. However, we can still prove the following result for the $n$ -qudit mixed states.

Proposition 23.

Given an $n$ -qudit state $\rho$ , the quantum-doubling constant satisfies the following properties:

(1) Positivity: $\delta_{q}[\rho]\geqslant 1$ , with equality iff $\rho\in\text{MSPS}$ .

(2) Additivity under tensor product: $\delta_{q}[\rho_{1}\otimes\rho_{2}]=\delta_{q}[\rho_{1}]\delta_{q}[\rho_{2}]$ .

(3) Invariance under Clifford unitary: $\delta_{q}\left[U\rho U^{\dagger}\right]=\delta_{q}[\rho]$ for any Clifford unitary $U$ .

(4) Monotonicity under partial trace: $\delta_{q}\left[\operatorname{Tr}_{i}\left[\rho\right]\right]\leqslant\delta_{% q}[\rho]$ , where $\operatorname{Tr}_{i}\left[\cdot\right]$ denotes the partial trace on the $i$ -th qudit for any $i\in[n]$ .

Proof.

These results come directly from the properties of quantum Ruzsa divergence in Theorem 28. ∎

For the qubit case, we can generalize the quantum doubling constant to the quantum tripling constant as follows.

Definition 24 (Quantum tripling constant).

Given an $n$ -qubit state $\rho$ , the quantum tripling constant is

(27)

\displaystyle\tilde{\delta}_{q}[\rho]=S(\boxtimes_{3}\rho)-S(\rho).

where the quantum convolution $\boxtimes_{3}$ is defined in (19).

Similar to the qudit case, we also have the following result for the $n$ -qubit pure state.

Proposition 25 (Quantum inverse sumset theorem for qubits).

Given an $n$ -qubit pure state $\psi$ ,

(1) $\tilde{\delta}_{q}[\psi]\geqslant 1$ , with equality iff $\psi\in\text{STAB}$ .

(2) If $1<\tilde{\delta}_{q}[\psi]\leqslant C$ , then

(28)

\displaystyle D(\psi||\mathcal{M}(\psi))\leqslant\frac{\log R(\psi)}{\log R(% \psi)-\log[1+\lambda^{2}(R(\psi)-1)]}\log C.

Proof.

The proof is similar to that of qudit case. ∎

Note that, the properties of the quantum-doubling constant in qudits in Proposition 23 also hold for the quantum-tripling constant in qubits by using the properties of $\boxtimes_{3}$ in [29].

5. Quantum Ruzsa divergence

In this section, we introduce a quantum Ruzsa divergence. To study its properties, we discuss the quantum sumset inequalities from the point of view of our quantum convolutional framework.

5.1. Definition and properties of the quantum Ruzsa divergence

Definition 26 (Quantum Ruzsa Divergence).

Given a quantum convolution $\boxtimes$ , the quantum Ruzsa divergence of a state $\rho$ with respect to the state $\sigma$ is

(29)

\displaystyle D_{Rz}(\rho||\sigma):=S(\rho\boxtimes\sigma)-S(\rho).

The $\alpha$ -order quantum Ruzsa divergence of $\rho$ with respect to $\sigma$ is

(30)

\displaystyle D_{\alpha,Rz}(\rho||\sigma):=S_{\alpha}(\rho\boxtimes\sigma)-S_{% \alpha}(\rho).

Definition 27 (Symmertized Quantum Ruzsa Divergence).

Given two quantum states $\rho$ and $\sigma$ , the symmertized quantum Ruzsa divergence between $\rho$ and $\sigma$ is

(31)

\displaystyle d_{Rz}(\rho,\sigma):=\frac{1}{2}\left(S(\rho\boxtimes\sigma)+S(% \sigma\boxtimes\rho)-S(\rho)-S(\sigma)\right).

The $\alpha$ -order quantum Ruzsa divergence between $\rho$ and $\sigma$ is

(32)

\displaystyle d_{\alpha,Rz}(\rho,\sigma):=\frac{1}{2}\left(S_{\alpha}(\rho% \boxtimes\sigma)+S_{\alpha}(\sigma\boxtimes\rho)-S_{\alpha}(\rho)-S_{\alpha}(% \sigma)\right).

Theorem 28.

The quantum Ruzsa divergence satisfies the following properties:

(1) Positivity: $D_{Rz}(\rho||\sigma)\geqslant 0$ , . Also $D_{Rz}(\rho||\sigma)=0$ iff the state $\rho$ is in the abelian C*-algebra generated by stabilizer group $S$ of $\sigma$ , i.e., $\rho$ is a convex sum of MSPSs associated with $S$ . From this we infer that $D_{Rz}(\rho||\rho)=0$ , iff $\rho$ is an MSPS.

(2) Additivity under tensor product: $D_{Rz}(\rho_{1}\otimes\rho_{2}||\sigma_{1}\otimes\sigma_{2})=D_{Rz}(\rho_{1}||% \sigma_{1})+D_{Rz}(\rho_{2}||\sigma_{2})$ .

(3) Invariance under Clifford unitary: $D_{Rz}(U\rho U^{\dagger}||U\sigma U^{\dagger})=D_{Rz}(\rho||\sigma)$ for any Clifford unitary $U$ .

(4) Monotonicity under partial trace: $D_{Rz}(\operatorname{Tr}_{i}\left[\rho\right]||\operatorname{Tr}_{i}\left[% \sigma\right])\leqslant D_{Rz}(\rho||\sigma)$ , where $\operatorname{Tr}_{i}\left[\cdot\right]$ denotes the partial trace on the $i$ -th qudit for any $i\in[n]$ .

(5) Convexity in the first term and concavity in the second: $D_{Rz}(\sum_{i}p_{i}\rho_{i}||\sigma)\leqslant\sum_{i}p_{i}D_{Rz}(\rho_{i}||\sigma)$ , and $D_{Rz}(\rho||\sum_{i}q_{i}\sigma_{i})\geqslant\sum_{i}q_{i}D_{Rz}(\rho||\sigma% _{i})$ , where $\set{p_{i}}_{i}$ and $\set{q_{i}}_{i}$ are classical probability distributions.

Proof.

(1) $D_{Rz}(\rho||\sigma)\geqslant 0$ comes from the entropy inequality under convolution, i.e., $S(\rho\boxtimes\sigma)\geqslant\max\set{S(\rho),S(\sigma)}$ [1, 2], and the condition for $D_{Rz}(\rho||\sigma)=0$ is the condition for the equality $S(\rho\boxtimes\sigma)=S(\rho)$ . (See Theorem 58 in [2].)

(2) This follows from the fact that $(\rho_{1}\otimes\rho_{2})\boxtimes(\sigma_{1}\otimes\sigma_{2})=(\rho_{1}% \boxtimes\sigma_{1})\otimes(\rho_{2}\boxtimes\sigma_{2})$ .

(3) This is the commutativity of Clifford unitaries and quantum convolution; see Lemma 85 in [2]. In other words, there always exists some Clifford unitary $U^{\prime}$ such that $(U\rho U^{\dagger})\boxtimes(U\sigma U^{\dagger})=U^{\prime}(\rho\boxtimes% \sigma)U^{\prime{\dagger}}$ for Clifford unitary $U$ .

(4) First, we have

		$\displaystyle S\left((\mathbb{E}_{\vec{x}}w(\vec{x})\rho w(\vec{x})^{\dagger})% \boxtimes\sigma\right)-S\left(\mathbb{E}_{\vec{x}}w(\vec{x})\rho w(\vec{x})^{% \dagger}\right)$
	$\displaystyle\leqslant$	$\displaystyle\mathbb{E}_{\vec{x}}\left[S\left((w(\vec{x})\rho w(\vec{x})^{% \dagger})\boxtimes\sigma\right)-S\left(w(\vec{x})\rho w(\vec{x})^{\dagger}% \right)\right]$
	$\displaystyle=$	$\displaystyle\mathbb{E}_{\vec{x}}\left[S\left(w(s\vec{x})(\rho\boxtimes\sigma)% w(s\vec{x})^{\dagger}\right)-S\left(w(\vec{x})\rho w(\vec{x})^{\dagger}\right)\right]$
	$\displaystyle=$	$\displaystyle S(\rho\boxtimes\sigma)-S(\rho).$

Here the second line comes from the joint convexity of the quantum relative entropy $D(U_{s,t}\rho\otimes\sigma U^{\dagger}_{s,t}||\rho\boxtimes\sigma\otimes\frac{% I}{d^{n}})$ , i.e.,

	$\displaystyle D\left(U_{s,t}\mathbb{E}_{\vec{x}}w(\vec{x})\rho w(\vec{x})^{% \dagger}\otimes\sigma U^{\dagger}_{s,t}\|\|\mathbb{E}_{\vec{x}}w(\vec{x})\rho w(% \vec{x})^{\dagger}\boxtimes\sigma\otimes\frac{I}{d^{n}}\right)$
	$\displaystyle\leqslant\mathbb{E}_{\vec{x}}D\left(U_{s,t}w(\vec{x})\rho w(\vec{% x})^{\dagger}\otimes\sigma U^{\dagger}_{s,t}\|\|w(\vec{x})\rho w(\vec{x})^{% \dagger}\boxtimes\sigma\otimes\frac{I}{d^{n}}\right),$

where $D(U_{s,t}\rho\otimes\sigma U^{\dagger}_{s,t}||\rho\boxtimes\sigma\otimes\frac{% I}{d^{n}})=S(\rho\boxtimes\sigma)+n\log d-S(\rho)-S(\sigma)$ . And the third line is a consequence of Proposition 41 in [2]. In fact,

\displaystyle\mathcal{E}\left(w(\vec{x})\otimes w(\vec{y})\rho_{AB}w(\vec{x})^% {\dagger}\otimes w(\vec{y})^{\dagger}\right)=w(s\vec{x}+t\vec{y})\rho_{AB}w(s% \vec{x}+t\vec{y})^{\dagger},

where $\mathcal{E}(\rho_{AB})=\operatorname{Tr}_{B}\left[U_{s,t}\rho_{AB}U^{\dagger}_% {s,t}\right]$ . Consequently, we only need to prove that

(33)

\displaystyle D_{Rz}\left(\operatorname{Tr}_{i}\left[\rho\right]||% \operatorname{Tr}_{i}\left[\sigma\right]\right)=S\left((\mathbb{E}_{\vec{x}}w(% \vec{x})\rho w(\vec{x})^{\dagger})\boxtimes\sigma\right)-S\left(\mathbb{E}_{% \vec{x}}w(\vec{x})\rho w(\vec{x})^{\dagger}\right)\;.

To prove this statement, note that

		$\displaystyle\left(\operatorname{Tr}_{i}\left[\rho\right]\otimes\frac{I_{i}}{d% }\right)\boxtimes\sigma$
	$\displaystyle=$	$\displaystyle\left(\mathbb{E}_{\vec{x}_{i}\in V}w(\vec{x}_{i})\rho w(\vec{x}_{% i})^{\dagger}\right)\boxtimes\sigma$
	$\displaystyle=$	$\displaystyle\mathbb{E}_{\vec{x}_{i}\in V}w(s\vec{x}_{i})(\rho\boxtimes\sigma)% w(s\vec{x}_{i})^{\dagger}$
	$\displaystyle=$	$\displaystyle\operatorname{Tr}_{i}\left[\rho\boxtimes\sigma\right]\otimes\frac% {I_{i}}{d},$

where the first and last lines come from the fact that $\mathbb{E}_{\vec{x}_{i}\in V}w(\vec{x}_{i})(\cdot)w(\vec{x}_{i})^{\dagger}=% \operatorname{Tr}_{i}\left[\cdot\right]\otimes\frac{I_{i}}{d}$ , and the second line comes from the following property (See Proposition 41 in [2])

\displaystyle\mathcal{E}(w(\vec{x})\otimes w(\vec{y})\rho_{AB}w(\vec{x})^{% \dagger}\otimes w(\vec{y})^{\dagger})=w(s\vec{x}+t\vec{y})\rho_{AB}w(s\vec{x}+% t\vec{y})^{\dagger},

where $\mathcal{E}(\rho_{AB})=\operatorname{Tr}_{B}\left[U_{s,t}\rho_{AB}U^{\dagger}_% {s,t}\right]$ . Repeating the above process for $\sigma$ , we obtain the following result

(34)

\displaystyle\left(\operatorname{Tr}_{i}\left[\rho\right]\otimes\frac{I_{i}}{d% }\right)\boxtimes\sigma=\operatorname{Tr}_{i}\left[\rho\boxtimes\sigma\right]% \otimes\frac{I_{i}}{d}=\operatorname{Tr}_{i}\left[\rho\right]\boxtimes% \operatorname{Tr}_{i}\left[\sigma\right]\otimes\frac{I_{i}}{d}.

Hence, we have

	$\displaystyle S\left((\mathbb{E}_{\vec{x}}w(\vec{x})\rho w(\vec{x})^{\dagger})% \boxtimes\sigma\right)=$	$\displaystyle S\left(\operatorname{Tr}_{i}\left[\rho\right]\boxtimes% \operatorname{Tr}_{i}\left[\sigma\right]\otimes\frac{I_{i}}{d}\right)=S\left(% \operatorname{Tr}_{i}\left[\rho\right]\boxtimes\operatorname{Tr}_{i}\left[% \sigma\right]\right)+\log d,$
	$\displaystyle S\left(\mathbb{E}_{\vec{x}}w(\vec{x})\rho w(\vec{x})^{\dagger}% \right)=$	$\displaystyle S\left(\operatorname{Tr}_{i}\left[\rho\right]\otimes\frac{I_{i}}% {d}\right)=S\left(\operatorname{Tr}_{i}\left[\rho\right]\right)+\log d.$

Hence, we have

\displaystyle S\left((\mathbb{E}_{\vec{x}}w(\vec{x})\rho w(\vec{x})^{\dagger})% \boxtimes\sigma\right)-S\left(\mathbb{E}_{\vec{x}}w(\vec{x})\rho w(\vec{x})^{% \dagger}\right)=D_{Rz}\left(\operatorname{Tr}_{i}\left[\rho\right]||% \operatorname{Tr}_{i}\left[\sigma\right]\right).

Therefore, we obtain the result.

(5) The convexity of $D_{Rz}(\rho||\sigma)$ with respect to the state $\rho$ comes directly from the joint convexity of the quantum relative entropy $D(U_{s,t}\rho\otimes\sigma U^{\dagger}_{s,t}||\rho\boxtimes\sigma\otimes\frac{% I}{d^{n}})$ . That is,

\displaystyle D(\sum_{i}p_{i}U_{s,t}\rho_{i}\otimes\sigma U^{\dagger}_{s,t}||% \sum_{i}p_{i}\rho_{i}\boxtimes\sigma\otimes\frac{I}{d^{n}})\leqslant\sum_{i}p_% {i}D(U_{s,t}\rho_{i}\otimes\sigma U^{\dagger}_{s,t}||\rho_{i}\boxtimes\sigma% \otimes\frac{I}{d^{n}}).

And, the concavity of $D_{Rz}(\rho||\sigma)$ with respect to the state $\sigma$ comes directly from the concavity of the von Neuman entropy $S(\cdot)$ . That is,

\displaystyle S(\sum_{i}q_{i}\rho\boxtimes\sigma_{i})\geqslant\sum_{i}q_{i}S(% \rho\boxtimes\sigma_{i}).

∎

Note that the quantum Ruzsa divergence differs significantly from the quantum relative entropy. For instance, the quantum relative entropy for identical states is always zero, i.e., $D(\rho||\rho)=0$ , which does not hold for quantum Ruzsa divergence.

The symmetrized quantum Ruzsa divergence also satisfies the properties of Theorem 28; one can repeat the proof above. Moreover, the properties (1)-(3) also hold for the $\alpha$ -order quantum Ruzsa divergence for $1\leqslant\alpha<+\infty$ ; this is a consequence of the property of quantum Rényi entropy and Lemma 57 in [2].

Conjecture 29 (Triangle inequality for the quantum Ruzsa divergence).

Given a balanced convolution $\boxtimes_{s,t}$ , i.e., $s\equiv t\mod d$ , the quantum Ruzsa divergence satisfies the triangle inequality

(35)

\displaystyle D_{Rz}(\rho||\tau)\leqslant D_{Rz}(\rho||\sigma)+D_{Rz}(\sigma||% \tau)\;,

which is equivalent to

(36)

\displaystyle S(\rho\boxtimes\tau)+S(\sigma)\leqslant S(\rho\boxtimes\sigma)+S% (\sigma\boxtimes\tau)\;.

Note that the inequality (36) is not balanced; the state $\sigma$ appears once on the left-hand side but twice on the right-hand side. Hence, we proposed the following balanced version, which we call ”convolutional strong sub-additivity”.

Conjecture 30 (Convolutional strong subadditivity).

Given three $n$ -qudit quantum states $\rho,\sigma,\tau$ , we have

(37)

\displaystyle S(\rho\boxtimes\tau\boxtimes\sigma)+S(\sigma)\leqslant S(\rho% \boxtimes\sigma)+S(\sigma\boxtimes\tau)\;,

where $\rho\boxtimes\tau\boxtimes\sigma:=(\rho\boxtimes_{s,t}\tau)\boxtimes_{l,m}\sigma$ , $\rho\boxtimes\sigma:=\rho\boxtimes_{s,t}\sigma$ , $\sigma\boxtimes\tau:=\sigma\boxtimes_{s,t}\tau$ , $s^{2}+t^{2}\equiv 1\mod s$ , $s\equiv t\mod d$ , $m\equiv ls$ , and $l^{2}+m^{2}\equiv 1\mod d$ .

The reason that we choose the parameters in this way is to generalize the classical balanced convolution on three random variables $\frac{\mathcal{X}+\mathcal{Y}+\mathcal{Z}}{\sqrt{3}}=\sqrt{\frac{2}{3}}\left(% \frac{\mathcal{X}+\mathcal{Y}}{\sqrt{2}}\right)+\frac{1}{\sqrt{3}}\mathcal{Z}$ to the quantum case.

Lemma 31.

If the convolutional strong subadditivity holds, then the triangle inequality of quantum Ruzsa divergence also holds.

Proof.

This comes from the entropy inequality for quantum convolution $S(\rho\boxtimes\tau\boxtimes\sigma)\geqslant\max\set{S(\rho\boxtimes\tau)\;,S(% \sigma)}$ [1, 2]. ∎

Here, let us show that the convolutional strong subadditivity holds with some additional assumption on the states.

Proposition 32.

The convolutional strong subadditivity holds for the following two cases,

(1) the quantum states $\rho,\sigma$ , and $\tau$ are all stabilizer states;

(2) the quantum states $\rho,\sigma$ , and $\tau$ are diagonal states in the computational basis.

Proof.

(1) Let us assume that quantum states $\rho,\sigma$ , and $\tau$ are all stabilizer states with stabilizer group $G_{\rho}$ , $G_{\sigma}$ and $G_{\tau}$ . then the absolute value of the characteristic $\Xi_{\rho}$ is either to $1$ oder $0$ , and $\Xi_{\rho}(\vec{x})$ equals $1$ , iff the Weyl operator $w(\vec{x})$ belongs to the stabilizer group. Similar arguments also work for $\sigma$ and $\tau$ .

Since

\displaystyle\Xi_{\rho\boxtimes\tau\boxtimes\sigma}(\vec{x})=\Xi_{\rho}(ls\vec% {x})\,\Xi_{\tau}(lt\vec{x})\,\Xi_{\sigma}(m\vec{x})\;,

with $s^{2}+t^{2}\equiv 1\mod d,l^{2}+m^{2}\equiv 1\mod d$ , then $|\Xi_{\rho\boxtimes\tau\boxtimes\sigma}(\vec{x})|$ is equal to $1$ oder $0$ , and

\displaystyle|\Xi_{\rho\boxtimes\tau\boxtimes\sigma}(\vec{x})|=1\;,~{}~{}\text% {iff}~{}~{}w(\vec{x})\in G_{\rho}\cap G_{\sigma}\cap G_{\sigma}\;.

That is, $\rho\boxtimes\tau\boxtimes\sigma$ is a stabilizer state with $G_{\rho}\cap G_{\sigma}\cap G_{\tau}$ as the stabilizer group, and the quantum entropy is

\displaystyle S(\rho\boxtimes\tau\boxtimes\sigma)=\log\frac{d^{n}}{|G_{\rho}% \cap G_{\sigma}\cap G_{\tau}|}\;.

Using the same reasoning, we can also prove that $\rho\boxtimes\sigma$ is a stabilizer state with $G_{\rho}\cap G_{\sigma}$ as the stabilizer group, and the quantum entropy is

\displaystyle S(\rho\boxtimes\sigma)=\log\frac{d^{n}}{|G_{\rho}\cap G_{\sigma}% |}\;.

$\sigma\boxtimes\tau$ is a stabilizer state with $G_{\sigma}\cap G_{\tau}$ as the stabilizer group, and the quantum entropy is

\displaystyle S(\sigma\boxtimes\tau)=\log\frac{d^{n}}{|G_{\sigma}\cap G_{\tau}% |}\;.

Besides, the quantum entropy of the stabilizer state $\sigma$ is

\displaystyle S(\sigma)=\log\frac{d^{n}}{|G_{\sigma}|}.

Hence, to prove the convolutional strong sub-additivity, it is equivalent to proving the following statement,

\displaystyle\frac{1}{|G_{\rho}\cap G_{\tau}\cap G_{\sigma}||G_{\sigma}|}% \leqslant\frac{1}{|G_{\rho}\cap G_{\sigma}||G_{\sigma}\cap G_{\tau}|}\;.

That is

\displaystyle\frac{|G_{\rho}\cap G_{\sigma}|}{|G_{\rho}\cap G_{\sigma}\cap G_{% \tau}|}\leqslant\frac{|G_{\sigma}|}{|G_{\sigma}\cap G_{\tau}|}\;.

Since $G_{\rho},G_{\sigma},G_{\tau}$ are all finte abelian group, then above statement is equivalent to the

\displaystyle|G_{\rho}\cap G_{\sigma}:G_{\rho}\cap G_{\sigma}\cap G_{\tau}|% \leqslant|G_{\sigma}:G_{\sigma}\cap G_{\tau}|\;,

where the $[G:H]$ is the index of a subgroup H in a group G. This comes from the following property of the index of the group: if $H,K$ are subgroups of $G$ , then

\displaystyle|H:H\cap K|\leqslant|G:K|.

(2) Let us consider that $2s^{2}\equiv 1\mod d$ , $m\equiv ls$ , and $l^{2}+m^{2}\equiv 1\mod d$ . Since $\rho,\sigma,\tau$ are diagonal in the computational basis, they can be written as follows

(38)

\displaystyle\rho=\sum_{\vec{x}\in\mathbb{Z}^{n}_{d}}p(\vec{x})|\vec{x}\rangle% \!\langle\vec{x}|,\quad\sigma=\sum_{\vec{x}\in\mathbb{Z}^{n}_{d}}q(\vec{x})|% \vec{x}\rangle\!\langle\vec{x}|,\quad\tau=\sum_{\vec{x}\in\mathbb{Z}^{n}_{d}}r% (\vec{x})|\vec{x}\rangle\!\langle\vec{x}|\;,

where $\set{p(\vec{x})},\set{q(\vec{x})},\set{r(\vec{x})}$ are probability distributions on $\mathbb{Z}^{n}_{d}$ .

Let us consider three random variables $\mathcal{X},\mathcal{Y},\mathcal{Z}$ , which takes values in $\mathbb{Z}^{n}_{d}$ as follows

\displaystyle\text{Pr}[\mathcal{X}=\vec{x}]=p(\vec{x})\;,\quad\text{Pr}[% \mathcal{Y}=\vec{x}]=q(\vec{x})\;,\quad\text{Pr}[\mathcal{Z}=\vec{x}]=r(\vec{x% }).

Then, after some calculation, we find that

	$\displaystyle S(\rho\boxtimes\tau\boxtimes\sigma)=H(\mathcal{X}+\mathcal{Y}+% \mathcal{Z})\;,\quad S(\sigma)=H(\mathcal{Y});$
	$\displaystyle S(\rho\boxtimes\sigma)=H(\mathcal{X}+\mathcal{Y})\;,\quad S(% \sigma\boxtimes\tau)=H(\mathcal{Y}+\mathcal{Z})\;,$

where $H(\mathcal{X})$ is the Shannon entropy of the discrete random variable $\mathcal{X}$ , and $\mathcal{X}+\mathcal{Y}$ is the sum of random variables $\mod d$ . Hence, the convolutional strong-subadditivity in this case is reduced to the classical case

(39)

\displaystyle H(\mathcal{X}+\mathcal{Y}+\mathcal{Z})+H(\mathcal{Y})\leqslant H% (\mathcal{X}+\mathcal{Y})+H(\mathcal{Y}+\mathcal{Z})

which comes directly from the data-processing inequality $I(\mathcal{X}:\mathcal{X}+\mathcal{Y}+\mathcal{Z})\leqslant I(\mathcal{X}:% \mathcal{X}+\mathcal{Y})$ .

∎

Note that, in the convolutional strong subadditivity (37), we consider the quantum convolution of three states as $(\rho\boxtimes\tau)\boxtimes\sigma$ , where $s^{2}+t^{2}\equiv 1\mod s$ , $s\equiv t\mod d$ , $m\equiv ls$ , and $l^{2}+m^{2}\equiv 1\mod d$ . We may also consider other possible quantum convolutions on three input states, like the one in (19) defined on qubit-systems.

Moreover, the usual strong subadditivity in quantum information theory is: given a tripartite state $\rho_{ABC}$ , it holds that $S(\rho_{ABC})+S(\rho_{C})\leqslant S(\rho_{AC})+S(\rho_{BC})$ , which was conjectured by Robinson, Ruelle [63] and Lanford, Robinson [64] and later proved by Lieb and Ruskai [59]. Our inequality (37) shares a similar form, so that we call (37) as ”Convolutional strong subadditivity”. Besides the strong subadditivty, there is a subadditivity for bipartite states, i.e., $S(\rho_{AB})\leqslant S(\rho_{A})+S(\rho_{B})$ . It is natural to consider the convolutional subaddivity or supaddivity. However, neither of them holds, as we can give some counterexamples in the following theorem.

Theorem 33 (No subaddivity or supaddivity for the quantum convolution).

(1) There exist quantum states $\rho$ and $\sigma$ such that

(40)

\displaystyle S(\rho\boxtimes\sigma)>S(\rho)+S(\sigma).

(2) There also exist quantum states $\rho,\sigma$ such that

(41)

\displaystyle S(\rho\boxtimes\sigma)<S(\rho)+S(\sigma).

Proof.

(1) Let us take $\rho$ to be the eigenstate of Pauli $Z$ operator corresponding to $+1$ eigenvalue, i.e.,

\displaystyle\rho=\frac{1}{d^{n}}\sum_{\vec{a}\in\mathbb{Z}^{n}_{d}}Z^{\vec{a}% }\;,

and $\sigma$ to be the eigenstate of Pauli $X$ operator corresponding to $+1$ eigenvalue, i.e.,

\displaystyle\sigma=\frac{1}{d^{n}}\sum_{\vec{a}\in\mathbb{Z}^{n}_{d}}X^{\vec{% a}}\;.

Then $S(\rho)=S(\sigma)=0$ and $\rho,\sigma$ are not commuting with each other. However, $\rho\boxtimes\sigma=\frac{I}{d^{n}}$ , as

\displaystyle\Xi_{\rho\boxtimes\sigma}(\vec{x})=\Xi_{\rho}(s\vec{x})\Xi_{% \sigma}(t\vec{x})=0,

for any $\vec{x}\neq 0$ . Hence

S(\rho\boxtimes\sigma)=n\log d>S(\rho)+S(\sigma)=0.

(2) Let us take both $\rho$ and $\sigma$ to be the maximally mixed state $I/d^{n}$ , then

S(\rho)=S(\sigma)=n\log d.

Moreover, $\rho\boxtimes\sigma=I/d$ , then

S(\rho\boxtimes\sigma)=n\log d<S(\rho)+S(\sigma)=2n\log d.

∎

Remark 34.

In the classical case, the convolution satisfies the fundamental inequality: $H(\mathcal{X}+\mathcal{Y})\leqslant H(\mathcal{X})+H(\mathcal{Y})$ [58, 49]. However, in the quantum case such an inequality does not hold, as shown in the Theorem 33. Therefore, new results and phenomena emerge in the quantum setting due to the quantum features, e.g., the non-commutativity of quantum states.

For the quantum Ruzsa divergence, we have the following result to characterize the stabilizer states.

Proposition 35.

Let $\psi_{1}$ and $\psi_{2}$ be two pure $n$ -qudit states for which $d_{Rz}(\psi_{1},\psi_{2})=0$ . Then there exists a pure stabilizer state $\phi_{stab}$ such that

d_{Rz}(\psi_{1},\phi_{stab})=d_{Rz}(\psi_{2},\phi_{stab})=0\;.

Here $d_{Rz}$ is the symmetric Ruzsa divergence defined in (31).

Proof.

Based on the definition of $d_{Rz}$ and the assumption that $\psi_{1}$ and $\psi_{2}$ are pure states, we infer that $S(\psi_{1}\boxtimes\psi_{2})=0$ , i.e., $\psi_{1}\boxtimes\psi_{2}$ is a pure state. Hence

(42)

\displaystyle 1=\frac{1}{d^{n}}\sum_{\vec{x}\in V^{n}}|\Xi_{\psi_{1}\boxtimes% \psi_{2}}(\vec{x})|^{2}=\frac{1}{d^{n}}\sum_{\vec{x}\in V^{n}}|\Xi_{\psi_{1}}(% s\vec{x})|^{2}\Xi_{\psi_{2}}(t\vec{x})|^{2}.

In addition,

\displaystyle 1=\frac{1}{d^{n}}\sum_{\vec{x}\in V^{n}}|\Xi_{\psi_{1}}(\vec{x})% |^{2}=\frac{1}{d^{n}}\sum_{\vec{x}\in V^{n}}|\Xi_{\psi_{2}}(\vec{x})|^{2}\;,

and

\displaystyle|\Xi_{\psi_{1}}(\vec{x})|\leqslant 1,|\Xi_{\psi_{2}}(\vec{x})|% \leqslant 1,\forall\vec{x}\;.

Hence, for any $\vec{x}$ ,

(43)

\displaystyle|\Xi_{\psi_{1}}(s\vec{x})|^{2}=|\Xi_{\psi_{2}}(t\vec{x})|^{2}=0~{% }\text{or }~{}1.

Thus, the size of support $\text{Supp}(\Xi_{\psi_{1}})$ and $\text{Supp}(\Xi_{\psi_{2}})$ are $d^{n}$ . From this we infer that both $\psi_{1}$ and $\psi_{2}$ are pure stabilizer states, based on the result that a pure state is a stabilizer state, iff the size of the support of its characteristic function is $d^{n}$ [41]. Moreover, both $\psi_{1}$ and $\psi_{2}$ have the same stabilizer group as $\psi_{1}$ up to some phase, based on (42). Therefore, we obtain the result.

∎

Remark 36.

A natural generalization of Proposition 35 is the following conjecture: if $d_{Rz}(\psi_{1},\psi_{2})\leqslant\epsilon$ , then there exist some pure stabilizer state $\phi_{stab}$ such that both

d_{Rz}(\psi_{1},\phi_{stab})\leqslant c\epsilon\;,\quad\text{and}\quad d_{Rz}(% \psi_{2},\phi_{stab})\leqslant c\epsilon\;,

where $c$ is some constant independent of $n$ . This question is a quantum generalization of Theorem 1.8 in [53].

5.2. Magic measure via quantum Ruzsa divergence

In this section, let us define a magic measure via the quantum Ruzsa divergence.

Definition 37.

The quantum Ruzsa divergence of magic $M_{Rz}(\rho)$ of a state $\rho$ is:

(44)

\displaystyle M_{Rz}(\rho):=\min_{\sigma\in STAB}D_{Rz}(\rho||\sigma).

Definition 38 (Stabilizer channel).

A quantum channel $\Phi:D(\mathcal{H}_{A})\to D(\mathcal{H}_{A})$ on a $n$ -qudit system $\mathcal{H}_{A}=\mathcal{H}^{\otimes n}$ is a stabilizer channel if it has the Stinespring representation

(45)

\displaystyle\Phi(\rho)=\operatorname{Tr}_{B}\left[U(\rho\otimes\sigma)U^{% \dagger}\right],

where $\sigma$ is a stabilizer state, and $U:\mathcal{H}_{A}\otimes\mathcal{H}_{B}\to\mathcal{H}_{A}\otimes\mathcal{H}_{B}$ is a Clifford unitary.

Proposition 39.

Given an $n$ -qudit state $\rho$ , the quantum Ruzsa divergence of magic $M_{Rz}(\rho)$ satisfies the following properties:

(1) $M_{Rz}(\rho)\geqslant 0$ ; also $M_{Rz}(\rho)=0$ , iff $\rho$ is an MSPS.

(2) $M_{Rz}$ is monotone under a stabilizer channel. That is, $M_{Rz}(\Phi(\rho))\leqslant M_{Rz}(\rho)$ for any stabilizer channel $\Phi$ .

Proof.

Property (1) follows from the positivity in Theorem 28.

(2) To prove the monotonicity of $M_{Rz}$ under a stabilizer channel, we only need to prove the monotonicity of $M_{Rz}$ for three cases, (2a) monotonicity under tensor product of a stabilizer state; (2b) monotonicity under Clifford unitary ; (2c) monotonicity under partial trace.

(2a) comes from the fact that additivity under tensor product in Theorem 28.

(2b) comes from the invariance of the quantum Ruzsa divergence under Clifford unitary in Theorem 28, i.e., $D_{Rz}(U\rho U^{\dagger}||U\sigma U^{\dagger})=D_{Rz}(\rho||\sigma)$ for any Clifford unitary, and the closedness of the stabilizer states under Clifford unitary, i.e., $U\sigma U^{\dagger}$ is a stabilizer state, for any Clifford unitary $U$ and stabilizer state $\sigma$ .

(2c) By the monotonicity of quantum Ruzsa divergence under partial trace in Proposition 28, we have

\displaystyle D_{Rz}(\rho||\sigma)\geqslant D_{Rz}\left(\operatorname{Tr}_{i}% \left[\rho\right]||\operatorname{Tr}_{i}\left[\sigma\right]\right)\geqslant% \min_{\sigma^{\prime}\in\text{STAB}}D_{Rz}\left(\operatorname{Tr}_{i}\left[% \rho\right]||\sigma^{\prime}\right)\;,

for any $\sigma\in\text{STAB}$ , where the last inequality comes from the closedness of stabilizer states under partial trace.

∎

Although the quantum Ruzsa divergence is differen from the quantum relative entropy in general, we find that they have the following relation by taking minimization over stabilizer states.

Proposition 40.

For any quantum state $\rho$ , we have

(46)

\displaystyle M_{Rz}(\rho)=\min_{\sigma\in\text{MSPS}}D(\rho||\rho\boxtimes% \sigma)\;,

where $D(\rho||\rho\boxtimes\sigma)$ is the quantum relative entropy of $\rho$ with respect to $\rho\boxtimes\sigma$ .

Proof.

First, by the concavity of quantum entropy, the minimization in $\min_{\sigma\in\text{STAB}}S(\rho\boxtimes\sigma)-S(\rho)$ is taken on pure stabilizer states. Hence, we only need to consider the case where $\sigma$ is a pure stabilizer state. Then $\rho\boxtimes\sigma$ is diagonal in the basis of the stabilizer states. This is because

	$\displaystyle\rho\boxtimes\sigma=$	$\displaystyle\rho\boxtimes(\mathbb{E}_{w(\vec{x})\in G_{\sigma}}w(\vec{x})% \sigma w(\vec{x})^{\dagger})$
	$\displaystyle=$	$\displaystyle\mathbb{E}_{w(\vec{x})\in G_{\sigma}}\rho\boxtimes(w(\vec{x})% \sigma w(\vec{x})^{\dagger})$
	$\displaystyle=$	$\displaystyle\mathbb{E}_{w(\vec{x})\in G_{\sigma}}w(t\vec{x})(\rho\boxtimes% \sigma)w(t\vec{x})^{\dagger},$

where the first line comes from the fact that $w(\vec{x})\sigma w(\vec{x})^{\dagger}=\sigma$ for any $w(\vec{x})\in G_{\sigma}$ , and the last the line comes from the following property (See Proposition 41 in[2])

\displaystyle\mathcal{E}(w(\vec{x})\otimes w(\vec{y})\rho_{AB}w(\vec{x})^{% \dagger}\otimes w(\vec{y})^{\dagger})=w(s\vec{x}+t\vec{y})\rho_{AB}w(s\vec{x}+% t\vec{y})^{\dagger},

where $\mathcal{E}(\rho_{AB})=\operatorname{Tr}_{B}\left[U_{s,t}\rho_{AB}U^{\dagger}_% {s,t}\right]$ .

Without loss of generality, let us assume that the basis is the computational basis $\set{\ket{k}}$ . Let us denote $\Delta$ as the fully-dephasing channel with respect to this basis, then we have

(47)

\displaystyle\rho\boxtimes\sigma=\Delta(\rho)\boxtimes\sigma=\Pi\Delta(\rho)% \Pi=\sum_{i}p_{i}|\Pi(i)\rangle\!\langle\Pi(i)|,

where $\Delta(\rho)=\sum_{i}p_{i}|i\rangle\!\langle i|$ , and $\Pi$ is a permutation on the basis. Thus,

(48)

\displaystyle\min_{\sigma\in\text{STAB}}S(\rho\boxtimes\sigma)-S(\rho)=\min_{B% _{STAB}}S(\Delta_{B_{STAB}}(\rho))-S(\rho),

where $\min_{B_{STAB}}$ denotes the minimization over all the orthonormal basis generated by the stabilizer states, and $\Delta_{B_{STAB}}$ is the corresponding fully-dephasing channel. Moreover,

$\displaystyle-\operatorname{Tr}\left[\rho\log\rho\boxtimes\sigma\right]$	$\displaystyle=$	$\displaystyle-\operatorname{Tr}\left[\rho\log\Delta(\rho\boxtimes\sigma)\right]$
	$\displaystyle=$	$\displaystyle-\operatorname{Tr}\left[\Delta(\rho)\log\Delta(\rho\boxtimes% \sigma)\right]$
	$\displaystyle\geqslant$	$\displaystyle-\operatorname{Tr}\left[\Delta(\rho)\log\Delta(\rho)\right],$

where the last line comes from the non-negativity of the relative entropy $D(\Delta(\rho)||\Delta(\rho\boxtimes\sigma))$ . Hence, we have

\min_{\sigma\in\text{MSPS}}D(\rho||\rho\boxtimes\sigma)=\min_{B_{STAB}}S(% \Delta_{B_{STAB}}(\rho))-S(\rho).

This is the desired result. ∎

6. Conclusion

We provide an entropic q-CLT, as well as a quantitive bound on the rate of its convergence. Moreover, we propose a framework to study quantum sumset and quantum inverse sumset problems using our convolution on DV quantum systems. Furthermore, we introduce a quantum divergence, that we call the quantum Ruzsa divergence, and show it can serve as a magic measure.

There are still many interesting problems to solve with in the framework of quantum sumset and inverse sumset theory. We list some of them here:

(1) Can one generalize other classical sumset and inverse sumset results to our quantum convolutional framework? For example, what is the quantum version of the polynomial Freiman–Ruzsa conjecture in this setting? The interesting recent work of Gowers and collaborators [53] may be helpful.

(2) For example, one can also generalize the quantum inverse sumset theory to CV quantum systems (e.g., bosonic quantum systems) by using the CV beam splitter as the quantum convolution. Define the CV quantum-doubling constant as $\delta_{CV}(\rho)=S(\rho\boxplus_{\lambda}\rho)-S(\rho)$ in a similar way as Definition 20, where $\boxplus_{\lambda}$ is the CV quantum convolution defined via beam splitter [65, 66, 67]. Then the CV quantum inverse sumset problem will characterize the relative entropy $D(\rho||\rho_{G})$ , by using the CV quantum doubling constant $\delta_{CV}$ , where $\rho_{G}$ is the Gaussian state with same mean value and covariance matrices. Moreover, we can also define the CV Ruzsa divergence as $S(\rho\boxplus_{\lambda}\sigma)-S(\rho)$ . We plan to develop these ideas within the CV setting in future work.

(3) Finally, can we find some physical interpretations and applications of these mathematical results?

7. Acknowledgement

We thank Michael Freedman, Yichen Hu, Bryna Kra, Yves Hon Kwan, Elliott Lieb, Freddie Manners, Graeme Smith, and Yufei Zhao for the helpful discussion. This work was supported in part by the ARO Grant W911NF-19-1-0302 and the ARO MURI Grant W911NF-20-1-0082.

8. Appendix

8.1. Properties of quantum convolution

For completeness, we list some useful properties of quantum convolution that we may use throughout.

Lemma 41 ([1, 2]).

The quantum convolution $\boxtimes_{s,t}$ satisfies the following properties:

(1)

Convolution-multiplication duality: $\Xi_{\rho\boxtimes_{s,t}\sigma}(\vec{x})=\Xi_{\rho}(s\vec{x})\Xi_{\sigma}(t% \vec{x})$ , for any $\vec{x}\in V^{n}$
(2)

Convolutional stability: If both $\rho$ and $\sigma$ are stabilizer states, then $\rho\boxtimes_{s,t}\sigma$ is a stabilizer state.
(3)

Quantum central limit theorem: The iterated convolution $\boxtimes^{N}\rho$ of a zero-mean state $\rho$ converges to $\mathcal{M}(\rho)$ as $N\to\infty$ .
(4)

Quantum maximal entropy principle: $S(\rho)\leqslant S(\mathcal{M}(\rho))$ .
(5)

Commutativity with Clifford unitaries: for any Clifford unitary $U$ , there exists some Clifford unitary $V$ such that $(U\rho U^{\dagger})\boxtimes(U\sigma U^{\dagger})=V(\rho\boxtimes\sigma)V^{\dagger}$ for any input states $\rho$ and $\sigma$ .

Note that the quantum convolution on qubit-system also satisfy the properties in Lemma 41. The details can be found in [29].

References

[1] Kaifeng Bu, Weichen Gu, and Arthur Jaffe. Quantum entropy and central limit theorem. Proceedings of the National Academy of Sciences, 120(25):e2304589120, 2023.
[2] Kaifeng Bu, Weichen Gu, and Arthur Jaffe. Discrete quantum Gaussians and central limit theorem. arXiv:2302.08423, 2023b.
[3] Ju. V. Linnik. An information-theoretic proof of the central limit theorem with Lindeberg conditions. Theory of Probability & Its Applications, 4(3):288–299, 1959.
[4] Andrew Barron. Entropy and the central limit theorem. Ann. Probab., 14(1):336–342, Sep 1986.
[5] Shiri Artstein, Keith Ball, Franck Barthe, and Assaf Naor. Solution of Shannon’s problem on the monotonicity of entropy. J. Amer. Math. Soc., 17(4):975–982, 2004.
[6] Shiri Artstein, Keith Ball, Franck Barthe, and Assaf Naor. Solution of Shannon’s problem on the monotonicity of entropy. Probability Theory and Related Fields, 129(3):381–390, 2004.
[7] Oliver Johnson and Andrew Barron. Entropy and the central limit theorem. Probability Theory and Related Fields, 129(3):391–409, Sep 2004.
[8] Lampros Gavalakis and Ioannis Kontoyiannis. The entropic central limit theorem for discrete random variables. In 2022 IEEE International Symposium on Information Theory (ISIT), pages 708–713, 2022.
[9] C. Cushen and R. Hudson. A quantum-mechanical central limit theorem. J. Appl. Probab., 8(3):454–469, Feb 1971.
[10] K. Hepp and E.H. Lieb. Phase-transitions in reservoir-driven open systems with applications to lasers and superconductors. Helv. Phys. Acta, 46(5):573–603, Feb 1973.
[11] K. Hepp and E.H. Lieb. On the superradiant phase transition for molecules in a quantized radiation field: the Dicke maser model. Ann. Phys., 76(2):360–404, Feb 1973.
[12] N. Giri and W. von Waldenfels. An algebraic version of the central limit theorem. Probab. Theory Relat. Fields, 42(2):129–134, June 1978.
[13] D. Goderis and P. Vets. Central limit theorem for mixing quantum systems and the ccr-algebra of fluctuations. Commun. Math. Phys., 122(2):249–265, June 1978.
[14] T. Matsui. Bosonic central limit theorem for the one-dimensional xy model. Rev. Math. Phys., 14(07n08):675–700, June 2002.
[15] M. Cramer and J. Eisert. A quantum central limit theorem for non-equilibrium systems: exact local relaxation of correlated states. New J. Phys., 12(5):055020, May 2010.
[16] V. Jaksic, Y. Pautrat, and C.-A. Pille. Central limit theorem for locally interacting Fermi gas. Commun. Math. Phys., 285(1):175–217, May 2009.
[17] Gérard Ben Arous, Kay Kirkpatrick, and Benjamin Schlein. A central limit theorem in many-body quantum dynamics. Commun. Math. Phys., 321(2):371–417, July 2013.
[18] Tom Michoel and Bruno Nachtergaele. Central limit theorems for the large-spin asymptotics of quantum spins. Probab. Theory Relat. Fields, 130(4):493–517, Dec 2004.
[19] D. Goderis, A. Verbeure, and P. Vets. Non-commutative central limits. Probab. Theory Relat. Fields, 82(4):527–544, Aug 1989.
[20] V. Jakšić, Y. Pautrat, and C.-A. Pillet. A quantum central limit theorem for sums of independent identically distributed random variables. J. Math. Phys., 51(1):015208, 2010.
[21] L. Accardi and Y. G. Lu. Quantum central limit theorems for weakly dependent maps. ii. Acta Math. Hung., 63(3):249–282, Sep 1994.
[22] Zhengwei Liu. Exchange relation planar algebras of small rank. Transactions of the American Mathematical Society, 368(12):8303–8348, Mar 2016.
[23] Chunlan Jiang, Zhengwei Liu, and Jinsong Wu. Block maps and Fourier analysis. Science China Mathematics, 62(8):1585–1614, Aug 2019.
[24] M. Hayashi. Quantum estimation and the quantum central limit theorem. Am. Math. Soc. Trans. Ser., 2(227):95–123, Sep 2009.
[25] Earl T. Campbell, Marco G. Genoni, and Jens Eisert. Continuous-variable entanglement distillation and noncommutative central limit theorems. Phys. Rev. A, 87:042330, Apr 2013.
[26] Simon Becker, Nilanjana Datta, Ludovico Lami, and Cambys Rouzé. Convergence rates for the quantum central limit theorem. Commun. Math. Phys., 383(1):223–279, Apr 2021.
[27] Raffaella Carbone, Federico Girotti, and Anderson Melchor Hernandez. On a generalized central limit theorem and large deviations for homogeneous open quantum walks. Journal of Statistical Physics, 188(1):8, 2022.
[28] Salman Beigi and Hami Mehrabi. Towards optimal convergence rates for the quantum central limit theorem. arXiv preprint arXiv:2310.09812, 2023.
[29] Kaifeng Bu, Weichen Gu, and Arthur Jaffe. Stabilizer testing and magic entropy. arXiv:2306.09292, 2023c.
[30] D. Gottesman. Stabilizer codes and quantum error correction. arXiv:quant-ph/9705052, 1997.
[31] Peter W. Shor. Scheme for reducing decoherence in quantum computer memory. Phys. Rev. A, 52:R2493–R2496, Oct 1995.
[32] Alexei Kitaev. Fault-tolerant quantum computation by anyons. Annals of Physics, 303(1):2–30, Jun 2003.
[33] D Gottesman. The Heisenberg representation of quantum computers. In Proc. XXII International Colloquium on Group Theoretical Methods in Physics, 1998, pages 32–43, 1998.
[34] Sergey Bravyi and David Gosset. Improved classical simulation of quantum circuits dominated by Clifford gates. Phys. Rev. Lett., 116:250501, Jun 2016.
[35] Sergey Bravyi, Graeme Smith, and John A. Smolin. Trading classical and quantum computational resources. Phys. Rev. X, 6:021043, Jun 2016.
[36] Sergey Bravyi, Dan Browne, Padraic Calpin, Earl Campbell, David Gosset, and Mark Howard. Simulation of quantum circuits by low-rank stabilizer decompositions. Quantum, 3:181, September 2019.
[37] Michael Beverland, Earl Campbell, Mark Howard, and Vadym Kliuchnikov. Lower bounds on the non-Clifford resources for quantum computations. Quantum Sci. Technol., 5(3):035009, May 2020.
[38] James R. Seddon, Bartosz Regula, Hakop Pashayan, Yingkai Ouyang, and Earl T. Campbell. Quantifying quantum speedups: Improved classical simulation from tighter magic monotones. PRX Quantum, 2:010345, Mar 2021.
[39] Kaifeng Bu and Dax Enshan Koh. Classical simulation of quantum circuits by half Gauss sums. Commun. Math. Phys., 390:471–500, Mar 2022.
[40] Xun Gao and Luming Duan. Efficient classical simulation of noisy quantum computation. arXiv:1810.03176, 2018.
[41] Kaifeng Bu and Dax Enshan Koh. Efficient classical simulation of Clifford circuits with nonstabilizer input states. Phys. Rev. Lett., 123:170502, Oct 2019.
[42] Dorit Aharonov, Xun Gao, Zeph Landau, Yunchao Liu, and Umesh Vazirani. A polynomial-time classical algorithm for noisy random circuit sampling. STOC, page 945–957, Jun 2023.
[43] Dax Enshan Koh. Further extensions of Clifford circuits and their classical simulation complexities. Quantum Information & Computation, 17(3&4):0262–0282, 2017.
[44] Sergey Bravyi and Alexei Kitaev. Universal quantum computation with ideal clifford gates and noisy ancillas. Phys. Rev. A, 71:022316, Feb 2005.
[45] Kaifeng Bu and Arthur Jaffe. Magic can enhance the quantum capacity of channels. arXiv:2401.12105, 2024.
[46] Koenraad Audenaert, Nilanjana Datta, and Maris Ozols. Entropy power inequalities for qudits. J. Math. Phys., 57(5):052202, 2016.
[47] Eric A. Carlen, Elliott H. Lieb, and Michael Loss. On a quantum entropy power inequality of Audenaert, Datta, and Ozols. J. Math. Phys., 57(6):062203, 2016.
[48] Dan Voiculescu, Nicolai Stammeier, and Moritz Weber. Free probability and operator algebras. European Mathematical Society, 2016.
[49] Terence Tao and Van H Vu. Additive Combinatorics, volume 105. Cambridge University Press, 2006.
[50] Terence Tao. Sumset and inverse sumset theory for Shannon entropy. Combinatorics, Probability and Computing, 19(4):603–639, 2010.
[51] Ioannis Kontoyiannis and Mokshay Madiman. Sumset and inverse sumset inequalities for differential entropy and mutual information. IEEE Transactions on Information Theory, 60(8):4503–4514, 2014.
[52] Ben Green, Freddie Manners, and Terence Tao. Sumsets and entropy revisited. arXiv preprint arXiv:2306.13403, 2023.
[53] W. T. Gowers, Ben Green, Freddie Manners, and Terence Tao. On a conjecture of Marton. arXiv preprint arXiv:2311.05762, 2023.
[54] Mokshay Madiman and Ioannis Kontoyiannis. Entropy bounds on abelian groups and the Ruzsa divergence. IEEE Transactions on Information Theory, 64(1):77–92, 2018.
[55] Imre Z. Ruzsa. Sumsets and entropy. Random Structures & Algorithms, 34(1):1–10, 2009.
[56] Mokshay Madiman. On the entropy of sums. In 2008 IEEE Information Theory Workshop, pages 303–307, 2008.
[57] Mokshay Madiman, Adam W. Marcus, and Prasad Tetali. Entropy and set cardinality inequalities for partition-determined functions. Random Structures & Algorithms, 40(4):399–424, 2012.
[58] Thomas M Cover. Elements of information theory. John Wiley & Sons, 1999.
[59] Elliott H. Lieb and Mary Beth Ruskai. Proof of the strong subadditivity of quantum‐mechanical entropy. Journal of Mathematical Physics, 14(12):1938–1941, 11 1973.
[60] D. Gross. Hudson’s theorem for finite-dimensional quantum systems. J. Math. Phys., 47(12):122107, 2006.
[61] Ashley Montanaro and Tobias J. Osborne. Quantum Boolean functions. Chicago Journal of Theoretical Computer Science, 2010(1), January 2010.
[62] Daniel Gottesman. Class of quantum error-correcting codes saturating the quantum Hamming bound. Phys. Rev. A, 54:1862–1868, Sep 1996.
[63] Derek W. Robinson and David Ruelle. Mean entropy of states in classical statistical mechanics. Communications in Mathematical Physics, 5(4):288–300, 1967.
[64] Oscar E. Lanford and Derek W. Robinson. Mean entropy of states in quantum‐statistical mechanics. Journal of Mathematical Physics, 9(7):1120–1125, 10 1968.
[65] Robert König and Graeme Smith. Limits on classical communication from quantum entropy power inequalities. Nature Photon, 7(2):142–146, 2013.
[66] Robert König and Graeme Smith. The entropy power inequality for quantum systems. IEEE Trans. Inform. Theory, 60(3):1536–1548, 2014.
[67] G. De Palma, A. Mari, and V. Giovannetti. A generalization of the entropy power inequality to bosonic quantum systems. Nature Photon, 8(3):958–964, 2014.

Entropic Quantum Central Limit Theorem and Quantum Inverse Sumset Theorem

Abstract.

1. Introduction

1.1. Background

1.2. Outline and summary of our main results

2. Background

Definition 1 (Characteristic function).

Definition 2 (Stabilizer states [62, 30]).

Definition 3 (Minimal stabilizer-projection state).

Definition 4 (Clifford unitary).

Definition 5 (Mean state (MS) [1, 2]).

Definition 6 (Zero-mean state [1, 2]).

Definition 7 (Magic gap [1, 2]).

Definition 8 (Quantum convoltuion [1, 2]).

2.1. Basic properties of entropy

Definition 9 (Quantum entropy).

Definition 10 (Quantum relative entropy).

Lemma 11 ([1, 2]).

3. The entropic, quantum, central-limit theorem

Theorem 12 (Entropic q-CLT).

Proof.

Theorem 13 (Rate of convergence).

Proof.

Corollary 14.

Proof.

Theorem 15 (Rényi entropic q-CLT via magic gap).

Proof.

3.1. Entropic q-CLT for qubit systems

Definition 16 (Key Unitary).

Definition 17 (Convolution of multiple states).

Theorem 18 (Entropic q-CLT for qubits).

Proof.

Theorem 19 (Rényi entropic q-CLT via magic gap).

Proof.

4. Entropic quantum inverse sumset theorem

Definition 20 (Quantum-doubling constant).

Remark 21.

Theorem 22 (Quantum inverse sumset theorem using magic gap).

Proof.

Proposition 23.

Proof.

Definition 24 (Quantum tripling constant).

Proposition 25 (Quantum inverse sumset theorem for qubits).

Proof.

5. Quantum Ruzsa divergence

5.1. Definition and properties of the quantum Ruzsa divergence

Definition 26 (Quantum Ruzsa Divergence).

Definition 27 (Symmertized Quantum Ruzsa Divergence).

Theorem 28.

Proof.

Conjecture 29 (Triangle inequality for the quantum Ruzsa divergence).

Conjecture 30 (Convolutional strong subadditivity).

Lemma 31.

Proof.

Proposition 32.

Proof.

Theorem 33 (No subaddivity or supaddivity for the quantum convolution).

Proof.

Remark 34.

Proposition 35.

Proof.

Remark 36.

5.2. Magic measure via quantum Ruzsa divergence

Definition 37.

Definition 38 (Stabilizer channel).

Proposition 39.

Proof.

Proposition 40.

Proof.

6. Conclusion

7. Acknowledgement

8. Appendix

8.1. Properties of quantum convolution

Lemma 41 ([1, 2]).

References

Entropic Quantum Central Limit Theorem
and Quantum Inverse Sumset Theorem