Central Limit Theorem for tensor products of free variables

This paper establishes a central limit theorem for tensor products of free variables, demonstrating that the limiting distribution is the semi-circle law for centered variables and a free interpolation between the semi-circle law and the classical convolution of two semi-circle laws for non-centered variables.

The Gist

The Big Picture: A New Kind of "Average"

Imagine you are a statistician trying to predict the behavior of a crowd. In the classical world (like flipping coins), if you flip enough coins and add up the results, the pattern always settles into a familiar "bell curve" (the Gaussian distribution). This is the famous Central Limit Theorem.

In the world of Free Probability (a branch of math dealing with quantum mechanics and random matrices), there is a similar rule. If you take a bunch of "free" (quantum-independent) variables and add them up, they don't form a bell curve; they form a semi-circle. This is the "Free Central Limit Theorem."

The Problem:
This paper asks a tricky question: What happens if we don't just add these variables together, but we multiply them in a specific, twisted way called a "tensor product"?

Think of a variable $a_k$ as a single person.

• Adding them: Putting them in a line and counting the total height.
• Tensoring them ($a_k \otimes a_k$): Taking that person, making a perfect clone, and having them stand side-by-side holding hands. Now you have a "double-person" unit.

The authors wanted to know: If you take many of these "double-person" units, normalize them, and add them up, what shape does the final crowd look like?

The Discovery: It Depends on the "Mean"

The authors found that the answer depends entirely on whether the original people ($a_k$) have a "center" or not.

Scenario A: The Centered Case (The "Zero-Mean" Crowd)

Imagine the original variables are "centered," meaning their average value is zero. They are perfectly balanced around a middle point.

• The Result: When you combine their "double-person" clones, the final crowd still forms a perfect semi-circle.
• The Analogy: It's like taking a group of people who are all standing exactly at the 0-meter mark, making clones, and adding them up. The chaos of the "cloning" process somehow cancels out, and you get the same smooth, semi-circular hill you would have gotten if you just added the original people.

Scenario B: The Non-Centered Case (The "Biased" Crowd)

Now, imagine the original variables are not centered. They have a bias; their average value is some number $\lambda$ (not zero).

• The Result: The final crowd does not form a semi-circle. Instead, it forms a strange, hybrid shape.
• The Analogy: Imagine the "double-person" units are now slightly unbalanced because the original people were leaning to one side. When you add them up, the result is a mixture of two different worlds:
  1. The quantum world (the semi-circle).
  2. The classical world (a shape you get from adding two semi-circles together in a traditional way).

The final shape is a "free interpolation" between these two. The exact shape depends on how strong the bias ($\lambda$) is compared to the natural variation (variance) of the people. If the bias is strong, the shape looks more like the classical mixture; if the bias is weak, it looks more like the quantum semi-circle.

Why Is This Hard? (The "Entangled" Puzzle)

The paper explains that this is difficult because of a "double layer" of independence.

1. Freeness: The different people ($a_1, a_2, a_3$) are "free" from each other (quantum independence).
2. Classical Independence: Inside the "double-person" unit ($a_k \otimes a_k$), the two legs of the tensor are actually independent in a classical sense.

It's like trying to solve a puzzle where the pieces are glued together in two different ways at the same time. The authors had to invent a new way to count and organize these pieces (using something called "partitions" and "crossing diagrams") to see the pattern.

The "Gotcha": They Are Not Free

One of the most surprising findings (Corollary 1.2) is a negative result.
Usually, in Free Probability, if you start with "free" variables, their sums behave predictably. The authors proved that if you take free variables and turn them into these "double-person" tensor units ($a_k \otimes a_k$), they are no longer free from each other.

• The Metaphor: Imagine you have a group of strangers who don't know each other (free). If you force each stranger to hold hands with their own clone, and then you try to treat the whole group of "cloned pairs" as a new group of strangers, it doesn't work. The act of cloning and pairing them creates a hidden connection between the pairs. They are "entangled" in a way that breaks the rules of free probability.

Summary of the Main Theorem

The paper establishes a new rule (Theorem 1.1):

• If you take free variables, make "double-person" tensors out of them, and sum them up:
  • If they are centered (mean = 0): You get a Semi-Circle.
  • If they are biased (mean $\neq$ 0): You get a Hybrid Shape that blends a semi-circle with a classical convolution of two semi-circles.

This hybrid shape is the "limiting law" for these specific types of quantum random variables, filling a gap in our understanding of how complex quantum systems behave when scaled up.

Technical Summary

Technical Summary: Central Limit Theorem for Tensor Products of Free Variables

Problem Statement
The paper addresses the asymptotic behavior of normalized sums of tensor products of free random variables. Specifically, given a sequence of free, self-adjoint, identically distributed random variables $(a_k)_{k \in \mathbb{N}}$ in a noncommutative probability space $(A, \tau)$ with mean $\lambda$ and variance $\sigma^2$, the authors investigate the convergence in distribution of the sequence:
$$S_n := \frac{1}{\delta\sqrt{n}} \sum_{k=1}^n (a_k \otimes a_k - \lambda^2)$$
in the product space $(A \otimes A, \tau \otimes \tau)$, where $\delta^2 = \text{var}(a \otimes a) = \sigma^2(\sigma^2 + 2\lambda^2)$.

While the classical Free Central Limit Theorem (FCLT) establishes that normalized sums of free variables converge to a semi-circular distribution, the tensor product structure introduces a complex dependence: the variables $a_k \otimes a_k$ are not free from one another, even though the underlying $a_k$'s are free. This is because the tensor product couples the "legs" of the variables, mixing classical independence (between the two tensor legs) with freeness (between different $k$). The paper aims to identify the limiting law of $S_n$ and determine whether it remains semi-circular or deviates in the general (non-centered) case.

Methodology
The authors employ a combinatorial approach rooted in free probability theory, specifically utilizing the moment-cumulant formula and the analysis of partitions.

1. Moment Analysis: The proof proceeds by calculating the moments $\tau'(S_n^p)$ of the normalized sum. Using the standard technique for limit theorems in free probability, the authors expand the moments into sums over partitions $\pi \in P(p)$.
2. Partition Decomposition: A key technical innovation is the decomposition of pair partitions $\pi \in P_2(p)$ using a bijection $\Phi$. This mapping decomposes any partition into a noncrossing "skeleton" $\hat{\pi}$ (the noncrossing closure) and a collection of connected pair partitions $\pi_T$ associated with the blocks of $\hat{\pi}$.
3. Recursive Removal of Blocks: To evaluate the contribution of connected partitions, the authors define an iterative process of removing blocks one by one. They introduce a "coloring" scheme ($w \in \{0, 1\}$) and a binary vector $\theta \in \{0, 1\}^4$ to track how the removal of a block affects the joint distribution of the remaining variables. This involves distinguishing between cases where the tensor legs are replaced by free copies ($\tilde{a}$) or kept as original variables ($a$).
4. Bipartite Graph Analysis: The authors analyze the intersection graph of the partitions. They demonstrate that the limit depends critically on whether this graph is bipartite. If the intersection graph contains an odd cycle, the contribution vanishes. If the graph is bipartite and connected, the contribution is non-zero and depends on the parameter $q = \frac{2\lambda^2}{\sigma^2 + 2\lambda^2}$.

Key Results
The main result, Theorem 1.1, establishes that $S_n$ converges in distribution to a measure $\mu_q$ defined as a free interpolation between two specific laws:
$$\mu_q := \sqrt{q} \left( \frac{1}{\sqrt{2}}\mu_{sc} + \frac{1}{\sqrt{2}}\mu_{sc} \right) \boxplus \sqrt{1-q} \, \mu_{sc}$$
where:

• $\mu_{sc}$ is the standard semi-circular distribution.
• $\boxplus$ denotes free convolution.
• $+$ denotes classical convolution.
• $q = \frac{2\lambda^2}{\sigma^2 + 2\lambda^2} \in [0, 1]$.

The limiting law $\mu_q$ is characterized by its free cumulants:

• Odd free cumulants vanish.
• The second free cumulant is 1.
• For even $n \ge 4$, the $n$-th free cumulant is $\kappa_n^{free}(\mu_q) = 2 \left(\frac{q}{2}\right)^{n/2} |P_{bicon}^2(n)|$, where $|P_{bicon}^2(n)|$ is the number of bipartite connected pair partitions of $n$.

Specific Cases:

• Centered Case ($\lambda = 0$): Here $q=0$. The limit reduces to $\mu_{sc}$, the standard semi-circular distribution. This confirms the heuristic that if the variables are centered, the tensor product behaves like a standard free sum.
• Non-Centered Case ($\lambda \neq 0$): Here $q > 0$. The limit is a non-trivial mixture.
  • If $q=1$ (which implies $\sigma=0$, i.e., deterministic variables), the limit is the classical convolution of two semi-circles (scaled), representing the sum of two independent semi-circular variables.
  • For $0 < q < 1$, the limit represents a "free interpolation" between the semi-circle and the classical convolution of two semi-circles.

Corollary on Freeness
A direct consequence of the main theorem is Corollary 1.2: If the variables $a_k$ are non-centered ($\lambda \neq 0$), the tensor products $\{a_k \otimes a_k\}$ are not free. If they were free, their normalized sum would converge to a semi-circular distribution, contradicting the result that the limit is $\mu_q$ with $q > 0$.

Significance and Claims
The paper claims to resolve the convergence and explicit identification of the limit for tensor products of free variables, a problem motivated by models in Quantum Information Theory (specifically random quantum channels and the spectrum of realigned quantum states).

The authors highlight that the difficulty lies in the interplay between classical independence (within the tensor legs) and freeness (across the tensor copies). They note that while similar computations exist for centered semi-circular variables, the general case requires a new understanding of the dependence structure.

The paper modestly suggests that the results and techniques developed could be useful for understanding the typical asymptotic spectrum of separable quantum states and the realignment operation in quantum information, areas where such random matrix models appear naturally. Furthermore, the result implies that any general notion of independence corresponding to the tensor case cannot simply reduce to standard freeness, as the limiting behavior differs significantly from the free convolution of the individual variables.