🔢 mathematics

DisCo Algebra: Why Disintegration and Combination Are Not Inverses

This paper introduces DisCo algebra, a compositional framework for joint probabilities that demonstrates how combining block-wise CDF summaries with regular conditional measures creates a renormalization operator whose information loss is precisely quantified by probability curvature, thereby revealing a fundamental limit on capturing cross-block dependence through scalar summaries.

Original authors: Xuexing Lu

Published 2026-10-19

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Xuexing Lu

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

In the vast landscape of modern science, from predicting weather patterns to training artificial intelligence, researchers constantly grapple with the challenge of understanding how many different things happen together. When we look at a complex system, we are often dealing with a single, massive probability distribution that describes the likelihood of countless variables interacting at once. To make sense of this, scientists have long tried to break these massive distributions down into smaller, simpler pieces, hoping to study the parts and then put them back together to understand the whole. The hope has always been that if you take a complex whole, separate it into its components, and then reassemble those components, you should end up with the exact same thing you started with. It is a natural assumption that decomposition and reconstruction are perfect mirrors of one another, a process where nothing is lost in the translation.

A new study by Xuexing Lu challenges this fundamental assumption, revealing that for high-dimensional probability, the act of breaking something down and the act of building it back up are not perfect inverses. The research introduces a new mathematical framework called DisCo algebra, which treats the separation of a joint distribution into smaller blocks and the recombination of those blocks as two distinct operations that do not simply cancel each other out. The study demonstrates that when you compress a complex distribution into simpler summaries and then try to reconstruct it, you inevitably lose a specific type of information. This loss is not a mistake in the method or a flaw in the calculation; rather, it is an intrinsic property of high-dimensional probability itself. The researchers prove that the gap between the original distribution and the reconstructed one is a measurable quantity they call "probability curvature," which represents the hidden dependencies between different parts of the system that cannot be captured by simple summaries.

To understand how this works, imagine a large dataset representing a complex system, such as the behavior of thousands of interacting variables. The researchers begin by dividing these variables into distinct groups, or blocks. For each block, they create a single, intrinsic summary statistic based on the cumulative probability of the data within that group. This summary acts like a compressed map of the block's internal behavior, reducing a multi-dimensional cloud of data into a single, manageable value. This process is called disintegration. It is a lossy compression, meaning that while the summary captures the overall shape of the block, it discards the fine-grained details of how the blocks relate to one another in specific configurations.

Once these summaries are created, the researchers attempt to rebuild the original system. They use a process called combination, which takes the summaries from the different blocks and tries to stitch them back together into a full probability distribution. This reconstruction relies on a specific rule: it assumes that once the summaries are fixed, the different blocks behave independently of one another given those summaries. In other words, the reconstruction builds a new distribution where the blocks are linked only by their summaries, but the internal details of one block do not influence the internal details of another once the summary values are known. The study shows that this reconstruction is a perfect embedding only if the original system already satisfied this condition of conditional independence. If the original system had complex, hidden connections between the blocks that went beyond what the summaries could describe, the reconstruction fails to recover the original.

The crucial finding is that the order of operations matters immensely. If you start with a set of compatible summaries and blocks, and then combine them to form a distribution, you can perfectly recover the original summaries and blocks by breaking them down again. However, if you start with a complex, real-world distribution, break it down into summaries, and then try to rebuild it, the result is almost always different from the starting point. The reconstructed distribution is a "renormalized" version of the original, one that has been smoothed over to fit the limitations of the summary-based approach. The difference between the original and the reconstructed version is exactly what the researchers call probability curvature. This curvature measures the amount of information lost during the compression, specifically quantifying the dependence between blocks that the simple summaries failed to capture.

This result is not a failure of the DisCo framework but a revelation about the nature of complex systems. The study proves that there is a hard limit to how much cross-block dependence can be captured by using a single scalar summary for each block. When the probability curvature is zero, the system is simple enough that the summaries tell the whole story, and the decomposition and recombination are perfect inverses. But when the curvature is non-zero, it means there is a layer of coordination between the blocks that is invisible to the summaries. This hidden layer represents a form of complexity that cannot be bypassed by simply choosing a better summary or a different mathematical trick. The information loss is real and quantifiable, defined by the difference between the true conditional distribution of the system and the conditionally independent product used in the reconstruction.

The researchers emphasize that this asymmetry is a feature of the mathematics of probability, not a defect in their method. The combination step is designed to be a lossless embedding, but only for inputs that are already compatible with the summary format. The disintegration step, by contrast, is a lossy compression that forces any arbitrary distribution into that format. The fact that you cannot get back exactly what you started with after this compression is a fundamental truth about high-dimensional data. It suggests that in fields ranging from statistics to physics, there are limits to how much we can simplify a system without losing essential information about how its parts interact. The study provides a rigorous way to measure exactly how much is lost, offering a new tool for understanding the complexity of the world around us.

Ultimately, this work reframes how we think about breaking down and rebuilding complex systems. It shows that the path from the whole to the parts and back again is not a straight line but a journey with a cost. The cost is the probability curvature, a precise measure of the hidden connections that vanish when we try to describe a complex whole with simple, independent parts. By proving that disintegration and combination are not inverses, the study establishes a new boundary for what is possible in the analysis of joint probability distributions, reminding us that some aspects of complexity are irreducible and must be acknowledged rather than ignored.

Technical Summary: DisCo Algebra

Problem Statement
High-dimensional joint probability distributions are central to statistics and machine learning, yet a fundamental question remains unresolved: can a joint distribution be decomposed into simpler parts and subsequently reassembled such that decomposition and reassembly are inverse operations? Classical copula theory (Sklar's theorem) provides a factorization into one-dimensional marginals and a copula, but this is an analytic statement lacking an algebraic structure that explains how decompositions compose or how reconstructions plug together. Furthermore, it is unclear whether scalar summaries of block-wise dependencies are sufficient to capture all cross-block dependence in high-dimensional settings.

Methodology
The paper introduces DisCo algebra (Disintegration + Combination), a compositional framework built on the family of bare probability measures $P_n = \text{Prob}([0, 1]^n)$ without external transformations, parameter families, or enriched objects. The framework relies on two distinct algebraic structures:

Symmetric Cooperad (Disintegration):
- Mechanism: A joint measure $\mu \in P_n$ is decomposed along a coordinate partition $I = (I_1, \dots, I_k)$ .
- Intrinsic Summary: Instead of external transforms, the framework uses block-wise joint Cumulative Distribution Functions (CDFs) as intrinsic, measure-determined 1D summaries. For a block $I_j$ , the summary $\Psi_j$ maps the block's coordinates to a scalar representing the cumulative mass.
- Output: The disintegration map $\Delta_I$ produces a tuple of block marginal measures $(\mu_1, \dots, \mu_k)$ and a "skeleton" measure $\Phi_k$ on the space of these scalar summaries.
- Structure: This forms a symmetric cooperad where the cocomposition is defined via the disintegration theorem for Borel probability measures on standard Borel spaces.
Non-Unital Partial Operad (Combination):
- Mechanism: The combination map $\gamma$ reconstructs a joint measure from a skeleton $\Phi_k$ and block measures $(\mu_1, \dots, \mu_k)$ .
- Reconstruction: It utilizes regular conditional measures to integrate the product of block-wise conditional distributions against the skeleton measure: $\hat{\mu} = \int_{I^k} (\bigotimes_j \mu_j(\cdot | t_j)) d\Phi_k(t)$ .
- Constraints: This is a partial operad because it is only defined on an "admissible" subclass where the marginals of $\Phi_k$ match the pushforwards of the block summaries ( $(\pi_j)_*\Phi_k = (\Psi_j)_*\mu_j$ ). It is non-unital as it lacks a distinguished identity element; composition is strictly designed to pair with disintegration outputs.

Key Results
The paper establishes that disintegration and combination are not inverses in general, leading to the following core findings:

Asymmetry of Operations:
- Combination $\to$ Disintegration: This is the identity on admissible inputs. If one starts with compatible components, combines them, and then disintegrates, the original components are recovered.
- Disintegration $\to$ Combination: This defines a renormalization operator $R_I = \gamma \circ \Delta_I$ . Applying this to an arbitrary measure $\mu$ yields a reconstructed measure $\hat{\mu}$ that generally differs from $\mu$ .
Probability Curvature ( $K_I(\mu)$ ):
- The discrepancy between the original measure and the reconstructed measure is quantified by the probability curvature, defined as the expected conditional Kullback-Leibler (KL) divergence:
  $K_I(\mu) := \mathbb{E}_{t \sim \Phi_k} \left[ D_{KL} \left( \mu(\cdot | t) \,\middle\|\, \bigotimes_{j=1}^k \mu_j(\cdot | t_j) \right) \right]$
- $K_I(\mu) = 0$ if and only if $\mu$ lies in the image of the combination map (i.e., the cross-block dependence is fully captured by the scalar block summaries).
Information Loss Decomposition:
- The total KL divergence between the true measure and the product of block marginals decomposes into two terms:
  $D_{KL}(\mu \,\|\, \bigotimes \mu_j) = D_{KL}(\Phi_k \,\|\, \Phi_k^{\text{ind}}) + K_I(\mu)$
- The first term measures dependence between the scalar summaries (skeleton level).
- The second term ( $K_I(\mu)$ ) measures residual cross-block dependence that cannot be captured by 1D CDF-based summaries (curvature layer).

Significance and Claims
The paper frames these results not as a defect of the DisCo framework, but as a hardness result regarding high-dimensional probability itself:

Expressive Ceiling: The framework demonstrates a fundamental limit: one cannot fully describe arbitrarily complex cross-block dependence using a single scalar summary per block. The "curvature" represents the irreducible information loss when compressing high-dimensional dependencies into scalar interfaces.
Algebraic Rigor: By formalizing probability decomposition as a non-unital partial operad and a symmetric cooperad, the paper provides a rigorous algebraic structure for joint probability that goes beyond the analytic statements of classical copula theory.
Lossy vs. Lossless: Disintegration is characterized as a lossy compression (discarding residual dependence), while combination is a lossless embedding (reconstructing exactly what the skeleton can encode). The renormalization operator $R_I$ quantifies exactly what is lost in the compression.
Distinction from Fiber Information: The paper clarifies that probability curvature is distinct from intra-block complexity (fiber information). A block can be highly structured internally yet independent of others (zero curvature), or blocks can be individually uniform yet exhibit strong cross-block coordination (nonzero curvature).

In conclusion, DisCo algebra reveals that joint probability is a derived concept subject to an intrinsic "curvature" obstruction. Compression via scalar block summaries is never the inverse of construction; the gap between them is a measurable, non-negative quantity that defines the limits of CDF-based dimension reduction.

Technical Summary: DisCo Algebra

More like this