🔬 physics

From First Principles to Multi-scale Decomposition:Mutual Information as a Segregation Index

This paper derives mutual information as the unique segregation index satisfying mean-minimisation and invariance properties, enabling a novel multi-scale decomposition that reveals previously inaccessible insights into ethnic residential segregation in England and Wales.

Original authors: Rohit Sahasrabuddhe, Renaud Lambiotte

Published 2026-03-17

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Rohit Sahasrabuddhe, Renaud Lambiotte

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

The Big Idea: Measuring How "Sorted" Our Neighborhoods Are

Imagine you are looking at a giant jigsaw puzzle of a city. Some pieces are red, some are blue, some are green. If you look at the whole picture, you see a mix of colors. But if you zoom in on specific sections of the puzzle, you might see that the red pieces are all clumped together in one corner, and the blue pieces are in another.

This is segregation: when people of different backgrounds (ethnicities, incomes, etc.) live in separate neighborhoods rather than mixing together.

For decades, scientists have tried to measure this "sorting" with math. But the problem is that there are too many ways to measure it, and different ways give different answers. This paper asks: "Is there one 'perfect' way to measure segregation that makes the most sense logically?"

The authors, Rohit and Renaud, say yes. They prove that the best way to measure segregation is a concept from information theory called Mutual Information.

Part 1: The Two Rules of a Good Ruler

To find the "perfect" ruler, the authors didn't just guess. They started from scratch (first principles) and asked: "What properties must a good measurement have?" They came up with two golden rules:

Rule 1: The "Average" Must Make Sense (Mean-Minimisation)

Imagine you have three different fruit baskets:

Basket A: Mostly apples.
Basket B: Mostly oranges.
Basket C: A mix of both.

If you want to describe the "average" fruit basket for the whole group, you should just combine all the fruit into one giant pile and look at the mix. You shouldn't invent a new, imaginary fruit basket that doesn't exist.

The Metaphor: Think of this like blending smoothies. If you have three different smoothies and you pour them all into one big blender, the result is the "mean." A good segregation index must agree that the "average" neighborhood is just the sum of all its parts. If you try to measure segregation using a weird, imaginary average, your math breaks.

Rule 2: The "Zoom Out" Rule (Invariance)

Imagine you are looking at a map of the world.

Zoomed In: You see every country (France, Germany, Italy).
Zoomed Out: You see continents (Europe, Asia, Africa).

If you zoom out and group France and Germany together into "Europe," the world shouldn't suddenly look more segregated just because you changed the map. If two groups were mixed together before, they should still look mixed after you group them.

The Metaphor: Think of it like sorting a deck of cards.

Fine Scale: You sort by exact card (Ace of Spades, 2 of Spades...).
Coarse Scale: You sort just by color (Red vs. Black).

If you switch from sorting by exact cards to just sorting by color, you shouldn't suddenly claim the deck is "more sorted." The measurement needs to stay consistent no matter how much you zoom in or out.

The Result: The "Mutual Information" Magic

The authors proved that there is only one mathematical tool that obeys both of these rules perfectly. That tool is Mutual Information.

Think of Mutual Information as a "Surprise Meter."

If you pick a random person in a city and ask, "Where do you live?" and that answer tells you nothing about their ethnicity, there is zero segregation. (The city is a perfect mix).
If you pick a random person and their address tells you exactly who they are ethnically (e.g., "If they live on this street, they are definitely Group A"), there is maximum segregation. (The city is perfectly sorted).

Mutual Information measures exactly how much "surprise" is removed when you know someone's location.

Part 2: The Superpower – Breaking It Down

The real magic of this paper isn't just measuring segregation, but breaking it down like a LEGO set.

Most old ways of measuring segregation are like a black box: they give you one big number (e.g., "The city is 50% segregated"). But they can't tell you why.

Mutual Information allows you to decompose the number. You can ask:

How much segregation is happening between big regions? (e.g., North vs. South London).
How much is happening inside those regions? (e.g., within the North).
How much is happening between broad groups? (e.g., White vs. Non-White).
How much is happening between specific groups? (e.g., Indian vs. Pakistani).

The Analogy: Imagine you are analyzing the noise in a crowded room.

Old Method: "It's very noisy." (One number).
New Method: "It's noisy because the band is loud (Region A), but also because the kids are screaming in the corner (Region B), and specifically because the drummer is hitting the cymbal too hard (Specific Group)."

This allows researchers to see where the sorting is happening and which groups are separating from each other.

Part 3: What They Found in England and Wales

The authors tested this new method on real data from the 2021 Census of England and Wales. Here are the cool things they discovered:

Diversity $\neq$ Integration:
Two cities can have the exact same mix of people (same diversity), but look completely different.
- Harrow: A mix of people living together like a salad (low segregation).
- Luton: A mix of people living in separate clusters like a layered cake (high segregation).
- The old math might have missed this difference, but the new "Surprise Meter" caught it immediately.
The "Hidden" Sorting:
In the borough of Croydon (a very diverse area), they looked at the "Asian" group.
- Broad View: "Asians" seem to live together.
- Fine View: When they zoomed in, they saw that Indians, Pakistanis, and Bangladeshis were actually living in different parts of the borough.
- The math showed that if you just lump everyone into "Asian," you miss 32% of the actual segregation happening inside that group.
The "White" Surprise:
They found that "White British" people and "Other White" people (like Irish or Polish) also live in different neighborhoods. Sometimes, "Other White" people live closer to non-White neighbors than "White British" people do. This nuance only shows up when you use this detailed, multi-scale approach.

Why Does This Matter?

This paper gives us a universal translator for inequality.

For City Planners: It helps them see exactly where to build schools or community centers to bridge gaps.
For Policymakers: It stops them from saying "Our city is diverse" when, in reality, the different groups are living in separate silos.
For Everyone: It helps us understand that "segregation" isn't just one big problem; it's a complex web of small problems happening at different scales (between neighborhoods, between cities, between specific ethnicities).

The Bottom Line

The authors built a new, mathematically perfect ruler called Mutual Information. It's the only ruler that works consistently whether you are looking at a whole country or a single street. By using it, they showed us that the story of segregation in England and Wales is much more complex, nuanced, and interesting than we thought.

In short: They didn't just count the people; they figured out the story of how the people are arranged.

1. Problem Statement

Measuring residential segregation is critical for understanding inequality, yet existing indices often lack a rigorous theoretical foundation or fail to account for the multi-scale nature of segregation (both geographic and demographic).

The Challenge: Segregation operates at various scales (e.g., neighborhoods vs. regions; specific ethnicities vs. broad racial groups). Choosing a scale is arbitrary, and different scales reveal different patterns.
The Gap: While information-theoretic indices (like Theil's index) have been used, there was no derivation from "first principles" that uniquely identifies Mutual Information (MI) as the optimal segregation index. Existing methods often struggle to decompose segregation consistently across nested spatial and demographic hierarchies without losing interpretability or mathematical rigor.

2. Methodology: Derivation from First Principles

The authors derive Mutual Information as the unique segregation index by defining segregation as the mean demographic dissimilarity of spatial units from the population. They formalize this using a divergence function $f$ and impose two axiomatic properties to narrow down the choice of $f$ .

A. The Framework

Setting: $n$ demographic identities ( $X$ ) and $m$ spatial units ( $S$ ).
Definition: Segregation $I$ is the weighted mean divergence of local demographic compositions ( $P_{X|S=s}$ ) from the global population composition ( $P_X$ ):
$I = \sum_s P_S(s) f(P_{X|S=s} || P_X)$
Goal: Identify the specific divergence $f$ that best fits the intuition of segregation.

B. The Two Defining Properties

The authors identify two properties that uniquely characterize the Kullback-Leibler (KL) divergence (and thus Mutual Information):

Mean-Minimiser Property:
- Concept: If we have a set of distributions, the "best representative" distribution should be their weighted mean.
- Implication: This ensures that the global population composition ( $P_X$ ) is the unique minimizer of the average divergence from local compositions. This aligns with how population counts are aggregated.
- Mathematical Result: This property restricts the divergence to the class of Bregman divergences.
Invariance Property (Information Monotonicity):
- Concept: Merging demographic groups (coarse-graining) should not increase the measured divergence. The divergence should only remain unchanged if the internal composition of the merged groups is identical in both distributions being compared.
- Implication: This captures the intuition that ignoring fine-grained differences should not artificially inflate segregation.
- Mathematical Result: Among Bregman divergences, only the KL Divergence satisfies this invariance property for $n \geq 3$ .

C. The Resulting Index

Since KL divergence is the only function satisfying both properties, the resulting segregation index is Mutual Information:
$I(X, S) = \sum_s P_S(s) d_{KL}(P_{X|S=s} || P_X)$
This is mathematically equivalent to the difference between the entropy of the population and the mean entropy of spatial units, or the KL divergence between the joint distribution and the product of marginals.

D. Multi-scale Decomposition

A key methodological contribution is the ability to decompose Mutual Information additively across scales:

Spatial Decomposition: Total segregation = Segregation between regions + Mean segregation within regions.
Demographic Decomposition: Total segregation = Segregation between supergroups + Mean segregation within supergroups.
This allows researchers to quantify exactly how much information is lost when moving from fine scales (e.g., specific ethnicities, small neighborhoods) to coarse scales (e.g., broad racial groups, large districts).

3. Key Contributions

Theoretical Uniqueness: The paper provides a first-principles derivation proving that Mutual Information is the only segregation index that satisfies both the mean-minimiser and invariance properties.
Multi-scale Decomposability: Unlike other indices, Mutual Information admits a strict additive decomposition for both geographic and demographic hierarchies. This allows for the isolation of contributions from specific regions and demographic supergroups.
Open Source Implementation: The authors provide Python-based open-source software to perform these multi-scale analyses.
Empirical Application: Application of the framework to 2021 Census data from England and Wales, demonstrating how the index reveals structural insights invisible to standard metrics.

4. Results: Case Study (England and Wales)

Using 2021 Census data (19 fine ethnic identities across 188,000+ Output Areas), the authors analyzed residential segregation:

Scale Sensitivity:
- Spatial: More than 50% of segregation is explained by differences between districts. However, moving from Lower Super Output Areas (LSOAs) to Output Areas (OAs) reveals significant internal segregation within LSOAs.
- Demographic: At coarse scales, ~65% of segregation is between broad ethnic groups. At fine scales, this drops to ~58%, indicating that intra-group differences (e.g., between Indian and Pakistani subgroups) are significant.
Diversity vs. Segregation:
- There is a positive correlation between diversity and segregation, but it is not deterministic.
- Case Study (Harrow vs. Luton): Both districts have similar diversity and population sizes. However, Harrow is well-mixed (low segregation), while Luton exhibits high segregation with White neighborhoods surrounding an Asian core. Standard diversity metrics would miss this structural difference; MI decomposition captures it.
Intra-group Segregation (Croydon):
- In Croydon (a highly diverse London borough), coarse-graining demographics (grouping into broad categories) loses 48% of the segregation information.
- Decomposition: The "Asian" supergroup accounts for 32% of the within-group segregation (due to distinct spatial clustering of Indians vs. Pakistanis), while the "Black" supergroup accounts for only 12%.
- Aggregation Sensitivity: The choice of demographic grouping matters. A "White vs. Non-White" split explains less segregation than a "White British vs. Non-White British" split in Croydon, whereas the opposite is true in other boroughs. This highlights that the "optimal" demographic scale is region-dependent.

5. Significance and Future Directions

Theoretical Rigor: By grounding the index in information theory axioms, the paper resolves ambiguity in choosing segregation metrics. It establishes Mutual Information not just as a convenient tool, but as the mathematically necessary choice for specific desiderata.
Policy and Planning: The decomposability allows urban planners to identify where (which regions) and among whom (which subgroups) segregation is most acute. This moves beyond "how much" segregation exists to "how it is structured."
Beyond Segregation: The authors note that this framework applies broadly to clustering problems where one compares a partition of entities to metadata (e.g., network communities vs. industry codes).
Future Work: The paper suggests extending this to null models for statistical significance testing and developing methods to handle "partial" similarities between identities (moving beyond binary distinct/identical groupings).

In summary, the paper establishes Mutual Information as the gold standard for multi-scale segregation analysis, providing both a rigorous mathematical foundation and a powerful practical toolkit for dissecting complex patterns of inequality.