📊 statistics

From DPPs to $k$ -DPPs: identifiability analysis via spectral decomposition

This paper analyzes the geometry of determinantal point processes (DPPs) via spectral decomposition to demonstrate that while full DPPs are identifiable up to discrete sign similarity, conditioning on cardinality to form $k$ -DPPs introduces fundamental continuous non-identifiability due to scale, sign, and eigenspace rotation invariances, particularly when the number of possible subsets is smaller than the dimension of the parameter space.

Original authors: Hideitsu Hino, Keisuke Yano

Published 2026-05-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hideitsu Hino, Keisuke Yano

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are organizing a party. You have a list of $N$ potential guests, and you want to invite a group of people who will get along well but also bring some diversity to the conversation. You don't want a group of clones; you want a mix of personalities.

In the world of statistics and machine learning, this is modeled by something called a Determinantal Point Process (DPP). It's a mathematical tool that helps you pick diverse groups of items (like guests, photos, or news articles) by calculating probabilities based on a "kernel matrix" (a big grid of numbers representing how similar or different everything is).

This paper by Hideitsu Hino and Keisuke Yano takes a deep dive into the geometry of these models, specifically looking at what happens when you change the rules of the game.

Here is the breakdown of their findings using simple analogies:

1. The Two Knobs: Volume and Orientation

The authors break down the complex math of the DPP into two main parts using a technique called spectral decomposition. Think of the kernel matrix as a piece of clay that can be stretched and rotated.

The Eigenvalues ( $\Lambda$ ): The "Volume" Knob.
Imagine these are the settings that control how many people show up to the party. They determine the probability of getting a small group, a medium group, or a large group.
The Eigenvectors ( $U$ ): The "Orientation" Knob.
Imagine these control who is in the group, given that you've already decided on the size. If you want a group of 3, this knob decides if it's three musicians, three chefs, or a mix. It controls the specific "flavor" or correlation within that specific group size.

2. The Full Party vs. The Fixed-Size Party

The paper compares two scenarios:

The Full DPP: You let the party size vary. The math says you can figure out the "Volume" and "Orientation" knobs, with one tiny catch: you can flip the signs of the numbers (like turning a dial from +5 to -5) without changing the outcome. It's a small, discrete ambiguity.
The k-DPP (The Focus of the Paper): You decide beforehand, "I only want a party of exactly $k$ people." You condition the model on this fixed size.

The authors discovered that fixing the party size changes the rules of the game completely.

3. The New Problems: Why You Can't See the Whole Picture

When you force the party size to be exactly $k$ , the ability to uniquely identify the settings (identifiability) breaks down in three specific ways:

The Scale Problem (The Volume Knob is Broken):
In the full model, you know exactly how "loud" the volume is. In the fixed-size model, you only know the relative loudness. If you turn the volume up by 10% everywhere, the probability of getting a specific group of $k$ people doesn't change. You can't tell the difference between a "100-watt" party and a "200-watt" party if the size is fixed.
The Sign Problem:
Just like in the full model, you can still flip signs (positive to negative) without changing the result.
The Rotation Problem (The Orientation Knob is Blurry):
This is the big new discovery. In the full model, the orientation is mostly clear. In the fixed-size model, you can't see the orientation directly. You can only see the squared shadows of the orientation.
Analogy: Imagine looking at a 3D object through a foggy window. You can see the outline (the squared minors), but you can't tell if the object is rotated slightly to the left or right. There are many different rotations that look exactly the same through the fog.

4. The "Foggy Window" Theorem

The authors prove a mathematical rule about when this "fog" gets really thick.

They found that if the number of possible groups of size $k$ (calculated as "N choose k") is smaller than the number of settings you are trying to tune in the matrix, then there are infinite ways to rotate the settings that produce the exact same result.

The Analogy: Imagine you are trying to solve a puzzle with 100 pieces (the settings), but you only have 20 clues (the possible groups of size $k$ ). Because you have fewer clues than pieces, there are endless ways to arrange the remaining pieces that still fit the 20 clues.
The Result: Unlike the full model, where the ambiguity is just a few discrete flips, the fixed-size model has continuous, infinite ambiguity. You could be in a slightly different "universe" of settings, and you wouldn't know it just by looking at the data.

5. The Fisher Information (The Map)

The paper also looks at the "Fisher Information," which is essentially a map of how sensitive the model is to changes.

In the full model, the map is clear.
In the fixed-size model, the map has a "flat spot" (a direction where the map gives no information). This flat spot corresponds exactly to the "Scale Problem" mentioned earlier. If you try to walk in that direction (changing the scale), the map doesn't tell you anything new.

Summary

The paper argues that while DPPs are great for modeling diversity, forcing a specific group size (k-DPP) creates a fundamental blind spot.

You lose the ability to know the absolute "scale" of the diversity.
You lose the ability to know the exact "rotation" of the diversity, only seeing a blurred, squared version of it.
If the group size is small relative to the total pool, this blindness becomes a massive, continuous fog where many different underlying realities look identical.

The authors conclude that to understand these models better, we need to accept these geometric limitations and perhaps develop new ways to learn from data that account for these "foggy" directions.

Technical Summary: From DPPs to k-DPPs: Identifiability Analysis via Spectral Decomposition

Problem Statement
Determinantal Point Processes (DPPs) offer a probabilistic framework for modeling repulsive interactions and diversity. While algorithmically tractable, the parameter space of DPPs possesses a complex geometric structure constrained by positivity, spectral bounds, and nonlinear relations among principal minors. Existing geometric analyses have primarily relied on diagonal scaling, which clarifies embeddings into exponential families but leaves cardinality and correlation inherently coupled.

A critical gap in understanding arises when conditioning a full DPP on a fixed cardinality $k$ , resulting in a $k$ -DPP. The central problem addressed in this paper is how this conditioning fundamentally alters the identifiability of the model parameters. Specifically, the authors investigate how the spectral decomposition $L = U\Lambda U^\top$ separates the control of cardinality (governed by eigenvalues $\Lambda$ ) from the conditional law within fixed-cardinality strata (governed by eigenspace orientation $U$ ), and how this separation affects parameter recovery.

Methodology
The authors employ a geometric approach centered on the spectral decomposition of the kernel matrix $L$ . The methodology involves:

Spectral Parameterization: Expressing the $k$ -DPP probability measure using the spectral components $\Lambda$ (eigenvalues) and $U$ (orthogonal eigenvector matrix).
Exponential Family Analysis: Analyzing the $k$ -DPP as a curved exponential family, particularly when the eigenspace orientation is fixed to the identity matrix ( $U=I_N$ ). This involves examining the sufficient statistics and the degeneracy of the Fisher Information Matrix (FIM).
Invariance Characterization: Identifying explicit transformations of the spectral parameters that leave the $k$ -DPP probability distribution invariant.
Dimension-Counting Theorem: Utilizing the log-likelihood score function and linear algebra (rank-nullity theorem) to bound the dimension of the space of non-identifiable directions. This compares the dimension of the parameter space ( $N(N+1)/2$ for symmetric matrices) against the dimension of the sample space ( $\binom{N}{k}$ ).

Key Contributions and Results

Spectral Parameterization and Fisher Information Degeneracy:
The paper establishes that when $U=I_N$ , the $k$ -DPP forms a non-minimal exponential family. The sufficient statistics satisfy a nontrivial affine relation ( $\sum T_i(A) = k$ ). Consequently, the Fisher Information Matrix (FIM) with respect to the natural parameters $\theta$ is degenerate. Specifically, the FIM degenerates exactly along the "common-shift" direction $\mathbb{R}\mathbf{1}_N$ (the space of vectors where all components are equal). At symmetric points, the FIM reduces to the canonical metric on the additive quotient space $\mathbb{R}^N / \mathbb{R}\mathbf{1}_N$ .
Three Explicit Invariances (Identifiability Gap):
The authors characterize the identifiability gap between the full DPP and the $k$ -DPP by establishing three explicit invariances under which the $k$ -DPP distribution remains unchanged:
- Scale Invariance: The eigenvalues $\Lambda$ are identifiable only up to a common positive scale factor $c$ ( $\Lambda \to c\Lambda$ ).
- Sign Similarity: The eigenvector matrix $U$ is identifiable only up to sign flips ( $U \to DU$ , where $D$ is a diagonal matrix of $\pm 1$ ).
- Eigenspace Rotation: $U$ is identifiable only up to rotations $Q$ that commute with $\Lambda$ ( $U \to UQ$ where $Q\Lambda Q^\top = \Lambda$ ).
  Unlike the full DPP, where non-identifiability arises only from discrete sign similarity, the $k$ -DPP introduces continuous scale invariance and rotational invariance.
Continuous Non-Identifiability via Dimension Counting:
The paper proves a dimension-counting theorem (Theorem 2) demonstrating that additional, non-trivial continuous non-identifiability exists whenever the number of possible subsets of size $k$ is strictly less than the dimension of the symmetric matrix space:
$\binom{N}{k} < \frac{N(N+1)}{2}$
In such cases, the space of non-identifiable directions has a dimension strictly larger than the one-dimensional scale cone. The authors provide an example where $L=I_N$ and specific perturbations $H_\rho$ yield identical $k$ -DPP distributions despite distinct eigenvalue ratios, confirming the existence of a continuous non-identifiable fiber.

Significance
The paper claims to provide a mathematically transparent foundation for understanding the geometry of DPPs, specifically highlighting the structural shift that occurs when conditioning on cardinality.

Theoretical Insight: It clarifies that conditioning on cardinality decouples the spectral parameters, rendering the absolute scale unidentifiable and restricting eigenspace information to squared minors.
Geometric Foundation: By characterizing the null directions of the Fisher metric and the specific invariances, the work offers a basis for future analyses of constrained parametrization, curvature, and statistical inference in DPPs.
Algorithmic Implications: The authors suggest that these findings point toward the development of kernel learning algorithms that explicitly leverage spectral decomposition, addressing the central problem of learning DPP kernels from data.

The authors maintain a modest scope, focusing on the geometric and identifiability properties of the model rather than proposing specific experimental applications or future algorithmic implementations beyond the general direction of spectral-based learning.