Gurau's spectral density is not a probability measure for individual real symmetric tensors

This paper demonstrates that while the "spectral density" derived from Gurau's resolvent trace yields valid probability measures on average for certain random tensor ensembles, it fails to correspond to a probability measure for individual deterministic tensors, thereby proving it is not a pointwise-defined probability distribution.

The Gist

Imagine you have a giant, multi-dimensional puzzle piece called a tensor. In the world of mathematics, these are like super-charged versions of matrices (which are just grids of numbers). Just as a matrix can be squashed down to reveal its "soul" (its eigenvalues and a shape called a spectral density), mathematicians have been trying to do the same thing with these complex tensors.

Recently, a researcher named Gurau proposed a new "magic lens" (called a resolvent trace) to look at these tensors. The hope was that if you looked at a random tensor through this lens, you would see a smooth, predictable curve—a Spectral Density—that tells you everything about the tensor's hidden structure. This curve is supposed to act like a probability map, showing you how likely different values are to appear.

For a long time, everyone assumed this map always existed, especially when looking at average random tensors. It was like assuming that if you shake a bag of marbles, the average shape of the pile will always be a perfect hill.

The Big Surprise

This paper, written by Jerdee, Kunisky, and Moore, drops a bombshell: The map doesn't always exist.

They found specific, carefully crafted tensors (the "deterministic" ones) where the magic lens breaks. When you try to draw the probability map for these specific tensors, the math goes haywire. Instead of a nice, smooth hill where all the numbers are positive (like a real probability map should be), the math spits out negative numbers.

The "Negative Probability" Problem

To understand why this is a big deal, let's use an analogy:

Imagine you are baking a cake. The "Spectral Density" is the recipe.

• Ingredients (Moments): The recipe calls for cups of flour, sugar, and eggs. In math, these are called "moments."
• The Rule: You can't have negative cups of flour. If a recipe says "use -2 cups of sugar," it's nonsense. You can't bake a cake with negative sugar.

The authors show that for certain tensors, the "recipe" derived from Gurau's lens asks for negative sugar.

• If you look at a random batch of tensors (like shaking a bag of marbles), the average recipe works perfectly. It gives you a valid cake.
• But if you pick out one specific, weirdly shaped tensor (like a single, oddly molded marble), the recipe for that specific one demands negative ingredients.

Since you can't have a probability distribution with negative values, the "Spectral Density" for that specific tensor does not exist. It's not just a blurry picture; the picture itself is impossible to draw.

How They Did It (The "Bad Cake" Recipe)

The team didn't just guess this; they built a counter-example.

1. They started with a simple idea: Can we make a mathematical object that behaves like a "bad cake"?
2. They looked at a specific formula involving a random variable (like rolling dice) and found that for 3D tensors (cubes), you can arrange the numbers so that the "fourth moment" (a measure of how "spiky" the data is) becomes negative.
3. They constructed a specific 3D tensor with 27 dimensions (a very high-dimensional cube) where the math forces this "negative sugar" result.

Why This Matters

This is a crucial discovery for two reasons:

1. It fixes a misconception: For years, people thought this "Spectral Density" was a universal property of tensors, similar to how every matrix has eigenvalues. This paper proves that for individual tensors, this concept is broken. It only works "on average" for random groups, not for every single specific case.
2. It opens new doors: The authors suggest that maybe we need to allow "signed measures" (recipes with negative ingredients) to make sense of this. It's like saying, "Okay, we can't bake a normal cake, but maybe we can bake a 'ghost cake' that exists in a different mathematical realm."

The Takeaway

Think of the "Spectral Density" as a shadow cast by a 3D object.

• For a random cloud of objects, the average shadow is a perfect, smooth circle.
• But for one specific, jagged rock, the shadow might be a shape that doesn't make sense (like a circle with a hole in the middle that somehow has negative area).

This paper says: "Don't assume the shadow always makes sense for every single rock. Sometimes, the rock is just too weird for the shadow to exist."

This forces mathematicians to rethink how they describe the "spectrum" of complex data structures, reminding us that what works for the crowd (the average) doesn't always work for the individual.

Technical Summary

Here is a detailed technical summary of the paper "Gurau's spectral density is not a probability measure for individual real symmetric tensors" by Jerdee, Kunisky, and Moore.

1. Problem Statement

The paper addresses a fundamental question in the extension of Random Matrix Theory (RMT) to higher-order tensors.

• Context: In classical RMT, the trace of the resolvent of a real symmetric matrix $X$, defined as $R_X(z) = \frac{1}{N}\text{Tr}(zI - X)^{-1}$, admits a Laurent series expansion where the coefficients $\frac{1}{N}\text{Tr}(X^k)$ correspond to the moments of the empirical eigenvalue distribution (a probability measure).
• Generalization: Gurau (2020) proposed a generalization of this resolvent trace to real symmetric tensors $T \in \text{Sym}_p(\mathbb{R}^N)$. The expansion coefficients, denoted $\frac{1}{N}I_k(T)$, are homogeneous polynomials in the tensor entries.
• The Open Question: For random tensor ensembles (e.g., Gaussian), the expected values of these coefficients converge to the moments of a well-defined probability measure (a tensorial generalization of the Wigner semicircle law). However, it was an open problem whether, for individual deterministic tensors $T$, the sequence $\frac{1}{N}I_k(T)$ constitutes the moment sequence of some probability measure $\gamma_T$ (the "Gurau spectral density").
• Hypothesis: If such a measure exists for every $T$, then the coefficients must satisfy the positivity constraints required for a moment sequence (e.g., the Hankel matrix of moments must be positive semi-definite).

2. Methodology

The authors employ a combination of invariant theory, statistical physics (Gaussian integration), and cumulant analysis to construct a counterexample.

A. Mathematical Framework

1. Invariant Polynomials: The authors utilize the theory of $O(N)$-invariant polynomials. They define "matching tensors" $w(\mu)$ based on matchings of indices. The coefficients $I_k(T)$ are shown to be proportional to connected cumulants of the random variable $Y = \langle T, g^{\otimes p} \rangle$, where $g \sim \mathcal{N}(0, I_N)$.
2. Resolvent Trace Relation: Using Gaussian integration by parts, they derive the relationship:
   $$\frac{1}{N}I_k(T) = \frac{1}{N} \frac{1}{p^{k-1}(k-1)!} \kappa_k(\langle T, g^{\otimes p} \rangle)$$
   where $\kappa_k$ denotes the $k$-th cumulant.
3. Moment-Cumulant Connection: For a sequence to be moments of a probability measure, the underlying random variable must have non-negative moments. Specifically, if the sequence corresponds to a measure, the fourth moment must satisfy $m_4 \geq 3m_2^2$ (for a centered variable), which implies the fourth cumulant $\kappa_4 = m_4 - 3m_2^2$ must be non-negative.

B. Construction Strategy

The authors aim to construct a deterministic tensor $T$ such that the associated random variable $Y = \langle T, g^{\otimes p} \rangle$ has a negative fourth cumulant ($\kappa_4 < 0$).

1. Polynomial Approximation: They seek a univariate polynomial $q(x)$ such that $\kappa_4(q(g_1)) < 0$ for a standard normal $g_1$.
   • They note that for degree $p=2$ (matrices), $\kappa_4$ is always non-negative due to the spectral theorem structure ($g^T T g = \sum \lambda_i g_i^2$).
   • For degree $p=3$, they identify a polynomial $q(x) = x - \frac{1}{10}x^3$ which yields a negative fourth cumulant.
2. Homogenization: To map this univariate polynomial to a tensor contraction $\langle T, g^{\otimes 3} \rangle$, they approximate the term $x$ using the law of large numbers. They replace $g_1$ with a term involving the average of squares of other Gaussian variables:
   $$Y_N = g_1 \left( \frac{1}{N} \sum_{i=2}^{N+1} g_i^2 \right) - \frac{1}{10}g_1^3$$
   As $N \to \infty$, the term in parentheses converges to 1 almost surely, recovering the behavior of $q(g_1)$.
3. Tensor Representation: Since $Y_N$ is a homogeneous cubic polynomial in $N+1$ variables, it can be represented exactly as $\langle T^{(N)}, g^{\otimes 3} \rangle$ for a specific symmetric tensor $T^{(N)}$.

3. Key Results

Theorem 1.1 (Main Result)

There exists a deterministic real symmetric tensor $T \in \text{Sym}_3(\mathbb{R}^{27})$ such that the coefficient $I_4(T)$ is negative. Consequently, the sequence $\frac{1}{N}I_k(T)$ cannot be the moment sequence of any probability measure.

Explicit Construction

The authors provide an explicit tensor $T$ for $N=26$ (dimension 27):

• $T_{1,1,1} = -1/10$
• $T_{1,i,i} = T_{i,1,i} = T_{i,i,1} = \frac{1}{3 \times 26} = \frac{1}{78}$ for $i \in \{2, \dots, 27\}$.
• All other entries are zero.

Verification

Calculating the fourth cumulant for this tensor yields:
$$\kappa_4(\langle T, g^{\otimes 3} \rangle) = -\frac{87}{250} + \frac{6}{26} + \frac{72}{26^2} + \frac{144}{26^3} < 0$$
Since the fourth moment of any probability measure must satisfy $m_4 \geq 3m_2^2$ (implying $\kappa_4 \geq 0$ for centered variables), the existence of a negative $I_4(T)$ proves that no such measure $\gamma_T$ exists for this specific $T$.

4. Significance and Implications

1. Limitation of Pointwise Spectral Density: The paper demonstrates that while the "spectral density" (Gurau's measure) is a robust concept for ensembles of random tensors (where it converges to a generalized semicircle law), it does not exist pointwise for individual deterministic tensors. The coefficients of the resolvent trace expansion do not necessarily form a valid moment sequence.
2. Distinction from Matrices: This highlights a fundamental structural difference between matrices ($p=2$) and higher-order tensors ($p \geq 3$). For matrices, the spectral theorem ensures that quadratic forms $g^T T g$ are always mixtures of $\chi^2$ distributions, preserving positivity. For tensors, the lack of a similar spectral theorem allows for the construction of "pathological" random variables via tensor contractions that violate moment constraints.
3. Signed Measures: The authors suggest that while a probability measure may not exist, the coefficients might correspond to a signed measure (a measure with negative mass). This aligns with previous work in tensor eigenvalue theory but shifts the interpretation of the "spectral density" from a physical probability distribution to a more abstract mathematical object.
4. Future Directions: The result clarifies the scope of free probability theory in the context of tensors. It suggests that future work must either restrict the definition of spectral density to specific classes of tensors or accept that the "density" is an ensemble-averaged property rather than an intrinsic property of a single tensor realization.

In summary, the paper resolves an open problem by proving that the analogy between matrix resolvents and tensor resolvents breaks down at the level of individual realizations, as the tensorial "moments" can violate the necessary positivity conditions for probability measures.