Chapter 8: Polarization as Dimensional Collapse — The Polarization Illusion Theorem
“A nation divided against itself cannot stand. But what if the division is an artifact of the ruler we used to measure it?”
RUNNING EXAMPLE — DISTRICT 7
District 7’s local newspaper runs a series titled “A District Divided.” The lead article reports that 48% of voters identify as Republican and 47% as Democrat, with only 5% truly independent — the lowest rate of independents in the district’s recorded history. Partisan antipathy has risen sharply: 62% of Democrats say they view Republicans unfavorably, up from 35% two decades ago. Among Republicans, 68% view Democrats unfavorably, up from 38%. The series concludes that District 7 is “deeply and perhaps irreconcilably polarized.”
But is it?
The polarization narrative is built on one-dimensional measurements: party identification, favorability of the other party, presidential vote share. These measurements are projections of the six-dimensional preference manifold onto the partisan axis. When we examine the full manifold, a different picture emerges: District 7’s voters are sorted (party identification predicts their position on multiple dimensions), and they exhibit affective polarization (they dislike the other party), but they are only modestly policy-polarized (the actual distance between party centroids on the manifold is moderate). The “deep divide” is, in significant part, a projection artifact amplified by metric distortion.
The Polarization Narrative
The dominant narrative of American politics in the 2020s is polarization. The country is divided. The parties have moved apart. Compromise is impossible. The center has collapsed. The extremes have captured the debate.
The evidence for this narrative is real — on the one-dimensional axis:
Partisan sorting: Since the 1960s, the correlation between party identification and issue positions has strengthened dramatically. Conservative Democrats and liberal Republicans, once common, have nearly vanished. Party identification now predicts positions on economics, social values, environment, foreign policy, trust, and identity. The covariance matrix \Sigma has become more structured: the off-diagonal terms \Sigma_{1j} linking economics to other dimensions have increased, meaning that a voter’s economic position increasingly predicts their position on every other dimension.
Affective polarization: Americans’ feelings toward the opposing party have deteriorated sharply. In 1980, the average partisan rated the other party at roughly 45 on a 100-point thermometer scale; by 2020, the average had dropped below 20. Partisans increasingly view the other party not merely as wrong but as threatening, unpatriotic, and immoral.
Elite polarization: Elected officials have moved apart on the DW-NOMINATE scale. The ideological overlap between the parties in Congress — once substantial (moderate Republicans overlapping with conservative Democrats) — has vanished. The two parties occupy non-overlapping regions of the 1D ideological space.
Geographic polarization: Landslide counties (where one party wins by 20+ points) have increased from 26% of counties in 1992 to over 50%. Voters are sorting geographically, clustering with like-minded neighbors.
This evidence is not fabricated. But it is one-dimensional. And the one-dimensional projection creates an illusion.
The Polarization Illusion Theorem
Theorem 4 (Polarization Illusion). A preference distribution that is unimodal on the d-dimensional manifold \mathcal{P} — possessing a single connected region of high density — can appear bimodal when projected onto a 1D subspace \pi(\mathcal{P}) if the projection axis is oblique to the distribution’s principal axis. Conversely, a bimodal 1D projection is consistent with either:
(a) Genuine bimodality: two separated clusters on the manifold (true polarization), or
(b) Oblique projection: a single elongated cluster on the manifold whose projection onto an oblique axis produces two apparent peaks (illusory polarization).
Observable 1D polarization is necessary but not sufficient evidence for manifold-level polarization.
Proof. Consider a unimodal Gaussian distribution on \mathbb{R}^d with mean \mu and covariance \Sigma. The projection onto a unit vector \mathbf{e} produces a 1D distribution with mean \langle \mu, \mathbf{e} \rangle and variance \mathbf{e}^T \Sigma \mathbf{e}. This is always unimodal.
Now consider a unimodal distribution that is not Gaussian but has elongated tails or a ridge-like structure. Specifically, consider a distribution with density f(x) \propto \exp(-\frac{1}{2}x^T \Sigma^{-1} x) \cdot (1 + \alpha \cdot P_2(\mathbf{e}_1^T x / \sigma_1)), where P_2 is the Legendre polynomial of degree 2 and \mathbf{e}_1 is the first principal axis. For suitable \alpha, this distribution has a single mode at the origin but produces a bimodal projection onto any axis oblique to \mathbf{e}_1.
More generally, for any mixed distribution f(x) = \sum_k w_k \cdot \mathcal{N}(\mu_k, \Sigma_k) whose component means \mu_k are separated along the principal axis of \Sigma but not along an oblique axis, the projection onto the oblique axis can produce bimodality where the full distribution is unimodal on each component’s manifold neighborhood. \square
What the Theorem Means for American Politics
The Polarization Illusion Theorem does not say that American polarization is entirely illusory. It says that one-dimensional evidence of polarization is ambiguous: it is consistent with genuine manifold-level polarization and with sorting-induced illusory polarization. Distinguishing the two requires multi-dimensional data.
The available multi-dimensional data — from ANES, the Cooperative Election Study, Pew Research surveys, and other sources — tells a more nuanced story than the 1D narrative:
Dimension by dimension:
d_1 (economic policy): Mild bimodality. Party centroids have separated modestly on economic issues — Democrats have moved somewhat left, Republicans have moved somewhat right. But the separation is moderate (roughly 1.5 standard deviations between centroids), and substantial overlap remains. Many voters in both parties hold centrist economic positions.
d_2 (social values): Substantial bimodality. This is the dimension where genuine polarization is most evident. Party centroids have separated sharply on issues like abortion, LGBTQ+ rights, and the role of religion in public life. The separation exceeds 2 standard deviations, and overlap has diminished.
d_3 (environmental priority): Unimodal with high variance. The distribution of environmental attitudes is single-peaked (most voters favor some level of environmental protection), but the variance is large (the range from “climate emergency” to “climate hoax” spans the full scale). The bimodality appears only when environmental attitudes are coded by party — evidence of sorting, not polarization. The underlying distribution of environmental concern is unimodal.
d_4 (foreign policy): Unimodal, poorly sorted by party. Foreign policy attitudes are only weakly correlated with party identification. Both parties contain hawks and doves, interventionists and isolationists. The 1D narrative of polarization has almost no basis on the foreign policy dimension.
d_5 (institutional trust): Strongly bimodal. This is the dimension of deepest genuine polarization — and the one least captured by the traditional left-right axis. Trust in institutions has collapsed among a large segment of the electorate, cutting across partisan lines (both left-populists and right-populists distrust institutions). The bimodality on d_5 is real and significant.
d_6 (identity): Moderately bimodal, heavily confounded with d_5. Identity salience has increased, and it correlates with institutional distrust. The d_5-d_6 axis is the axis of genuine polarization in the 2020s electorate — but it is not the traditional left-right axis.
The manifold picture: Americans are genuinely polarized on d_2 (social values) and d_5 (institutional trust), moderately sorted on d_1 (economics) and d_6 (identity), and barely sorted on d_3 (environment) and d_4 (foreign policy). The “polarized nation” narrative conflates the d_2-d_5 bimodality with a full-manifold bimodality that does not exist.
The Sorting-Polarization Distinction
The geometric framework draws a sharp distinction between two phenomena that the 1D projection conflates:
Sorting is the alignment of the covariance matrix \Sigma. When positions on different dimensions become correlated with party identity, the off-diagonal terms \Sigma_{ij} grow, and the electorate’s distribution becomes more structured — more predictable, more party-aligned. Sorting does not require the party centroids to move apart. It requires only that the within-party variance decrease (voters become more homogeneous within each party) and the cross-dimensional correlations increase (positions on different dimensions become linked through party identity).
Polarization is the separation of centroids. When the mean positions of the two parties’ supporters move apart on the manifold, the distance between centroids increases. Polarization is a statement about the first moment of each party’s distribution (the mean), not about the second moment (the covariance).
These are geometrically independent phenomena:
High sorting, low polarization: The covariance matrix is highly structured (party ID predicts positions on all dimensions), but the centroids are close (the parties’ average positions differ only modestly on each dimension). This describes an electorate that is neatly divided into two camps whose actual positions are not far apart. The 1D projection looks polarized (bimodal) because the sorted distribution has two concentrated clusters — but the clusters are close together on the manifold.
Low sorting, high polarization: The covariance matrix is unstructured (party ID does not predict positions on most dimensions), but the centroids are far apart on one or two dimensions. This describes an electorate with genuine policy disagreement on specific issues but no overall alignment of the preference structure with party identity. The 1D projection may not look polarized (the distribution may be unimodal) because the polarization is on specific dimensions, not on the 1D composite.
High sorting, high polarization: Both the covariance structure and the centroid separation are large. This is genuine full-manifold polarization — the most severe scenario. The 1D projection accurately reflects the manifold structure.
The 2020s American electorate exhibits high sorting with moderate polarization. America has experienced dramatic sorting since the 1960s: party identification now predicts positions on all six dimensions, the covariance matrix has become highly structured, and the two parties have become internally homogeneous. But the centroid separation is moderate on most dimensions — the party means are approximately 1.5 standard deviations apart, not the 3+ standard deviations that the “deeply divided nation” narrative implies.
The 1D projection conflates sorting and polarization because both produce bimodal 1D distributions. The geometric framework distinguishes them by examining the full covariance structure: sorting increases inter-dimensional correlation without increasing centroid distance; polarization increases centroid distance without necessarily affecting inter-dimensional correlation.
Affective Polarization as Metric Distortion
Affective polarization — the increasing dislike of the opposing party — is the most dramatic trend in American politics over the past two decades. It has grown faster than policy polarization, is more widely distributed across the electorate, and has more dramatic behavioral consequences (unwillingness to socialize across party lines, disapproval of interparty marriage, dehumanization of the other party’s supporters).
The geometric framework interprets affective polarization as metric distortion: a systematic inflation of perceived manifold distances between in-group and out-group members.
Formally, let d_{\mathcal{P}}(v_1, v_2) be the actual manifold distance between two voters (computed from their six-dimensional positions), and let \hat{d}(v_1, v_2) be the perceived distance (the distance that each voter estimates, based on their knowledge of the other’s party affiliation and media-mediated impressions). Affective polarization is the condition:
\hat{d}(v_{\text{in}}, v_{\text{out}}) \gg d_{\mathcal{P}}(v_{\text{in}}, v_{\text{out}})
where v_{\text{in}} is an in-group voter and v_{\text{out}} is an out-group voter. The perceived distance exceeds the actual distance, often dramatically. A Democrat and a Republican who are 1.5 standard deviations apart on the manifold may perceive themselves as 4 or 5 standard deviations apart — because each perceives the other through a distorted metric that inflates cross-party distances on all dimensions simultaneously.
The metric distortion is self-reinforcing: inflated perceived distance justifies stronger dislike, which provides more reason to maintain the inflated perception. The feedback loop produces a curvature singularity on the political manifold — a region of infinite perceived distance that may correspond to finite actual distance.
Where Affective Polarization Lives
Affective polarization is not uniform across the manifold. It is concentrated on specific dimensions and in specific regions:
Strongest on d_2 (social values) and d_6 (identity): The dimensions with the most emotional salience — where positions are tied to moral identity and group belonging — show the largest gap between actual and perceived distance.
Moderate on d_1 (economics): Economic disagreements, while real, produce less affective polarization than social and identity disagreements. Voters who disagree on tax policy do not hate each other. Voters who disagree on abortion or immigration often do.
Weak on d_3 (environment) and d_4 (foreign policy): These dimensions produce relatively little affective charge. Cross-party agreement on specific environmental and foreign policy questions is substantial and largely uncontroversial.
Concentrated near sacred-value boundaries: The curvature singularity is located near the sacred-value boundaries on d_2 (the abortion divide, the gun rights divide) and d_6 (the immigration divide, the racial justice divide). Voters on opposite sides of these boundaries experience near-infinite perceived distance.
The Empirical Test: ANES Data, 1972-2024
The Polarization Illusion Theorem provides the theoretical framework. The empirical test requires dimension-by-dimension analysis of voter preferences over time — examining whether the observed 1D polarization corresponds to genuine manifold-level separation or to sorting-induced projection artifacts.
The American National Election Studies provide the longest continuous time series of multi-dimensional political preference data, spanning from 1948 to the present. While the specific items have changed over the decades, sufficient overlap exists to track the six dimensions approximately:
d_1 (Economic policy), 1972-2024: The standard government services/spending scale shows a modest increase in bimodality. The gap between the party means on this scale has grown from approximately 0.8 standard deviations in 1972 to approximately 1.4 standard deviations in 2024. This is genuine policy polarization on the economic dimension — but it is moderate, not extreme. The distributions still overlap substantially: the interquartile ranges of the two parties’ economic positions intersect.
d_2 (Social values), 1972-2024: This is the dimension of greatest genuine polarization. Attitudes on abortion, gay rights, and gender roles show increasingly bimodal distributions since the 1990s, with the gap between party means growing from approximately 1.0 standard deviations in 1980 to approximately 2.2 standard deviations in 2024. The social values dimension has experienced genuine manifold-level separation — two clusters moving apart — not merely sorting.
d_3 (Environmental priority), 1992-2024: Environmental attitudes were approximately unimodal and non-partisan through the 1990s. Since 2000, partisan sorting has introduced apparent bimodality, but the underlying distribution of environmental concern remains single-peaked. The party means have separated by approximately 1.5 standard deviations, but much of this separation reflects sorting (environmentally concerned voters selecting into the Democratic party) rather than polarization (voters becoming more or less environmentally concerned).
d_4 (Foreign policy), 1972-2024: Foreign policy attitudes remain poorly sorted by party and approximately unimodal. Both parties contain hawks and doves, interventionists and isolationists. The partisan realignment on foreign policy (the populist-nationalist turn in the Republican party, the interventionist turn in the Democratic party’s foreign policy establishment) has produced some sorting since 2016, but the within-party variance remains large.
d_5 (Institutional trust), 1958-2024: This is the dimension of deepest divergence — but the divergence is not simply partisan. Trust in government has declined dramatically across the entire electorate since the 1960s (from approximately 75% expressing trust in 1964 to approximately 20% in 2024). The decline has been more pronounced among some groups than others, creating within-party variation that is as large as between-party variation. The d_5 distribution is now bimodal, with one cluster of residual high-trust voters (predominantly educated professionals and institutional insiders) and a large cluster of low-trust voters spanning both parties.
d_6 (Identity), 2004-2024: Identity salience has increased across the electorate. The partisan sorting on d_6 is substantial: strong identity-oriented voters have sorted into the Republican party (national identity) and the Democratic party (racial/ethnic identity), creating apparent bimodality on the 1D projection. But the underlying phenomenon is sorting, not polarization: voters have not become more identity-oriented in absolute terms so much as they have sorted into parties that emphasize different identity frameworks.
The manifold picture, summarized: Genuine policy polarization on d_2 (social values) and d_5 (institutional trust). Moderate sorting-driven bimodality on d_1 (economics), d_3 (environment), and d_6 (identity). Minimal partisan structure on d_4 (foreign policy). The “polarized nation” narrative is half right and half misleading: there is real separation on two dimensions, sorting on three dimensions, and minimal structure on one dimension. The full-manifold picture is far more nuanced than the 1D narrative suggests.
District 7: Decomposing the Divide
We apply the geometric decomposition to District 7’s “deep divide.”
The sorting analysis: District 7 is highly sorted. Party identification predicts positions on all six dimensions with correlation coefficients ranging from r = 0.3 (d_4, foreign policy) to r = 0.7 (d_2, social values). The covariance matrix \Sigma has strong inter-dimensional correlations within each party cluster.
The centroid analysis: The Democratic centroid in District 7 is approximately (-1.2, -0.8, 1.0, -0.3, -0.2, 0.5). The Republican centroid is approximately (1.0, 0.8, -0.5, 0.5, -1.2, 0.8). The Mahalanobis distance between centroids is 1.5 — moderate, not extreme. The centroid separation is largest on d_2 (1.6 units), d_5 (1.0 units), and d_3 (1.5 units), and smallest on d_4 (0.8 units) and d_6 (0.3 units).
The affective polarization analysis: Despite the moderate centroid separation, District 7’s voters perceive the gap as much larger. Democrats estimate Republicans’ average position as approximately (2.5, 2.0, -2.0, 1.5, -2.5, 2.0) — significantly more extreme than the actual Republican centroid. Republicans estimate Democrats’ average position as approximately (-2.5, -2.0, 2.5, -1.5, 1.0, -1.5) — again, more extreme than reality. The perceived inter-party distance is approximately 4.2 Mahalanobis units, nearly three times the actual distance of 1.5.
The manifold picture: District 7 is highly sorted (party ID strongly predicts manifold position), moderately polarized on the manifold (centroid distance 1.5), and severely afflicted by affective polarization (perceived distance 4.2, nearly 3x the actual distance). The newspaper’s “District Divided” narrative captures the affective distortion and the sorting, but it conflates both with policy polarization. On the manifold, the district is more united than divided — the centroids are 1.5 standard deviations apart, meaning the distributions overlap substantially. A randomly selected Democrat and a randomly selected Republican in District 7 have a roughly 30% chance of being closer to each other on the manifold than to their own party’s centroid.
The polarization is largely an illusion produced by three compounding distortions: the 1D projection (which makes sorting look like polarization), the media heuristic (which amplifies the conflict dimension), and affective metric distortion (which inflates perceived distances). On the full manifold, District 7 is a district where most voters share more than they realize — united by shared concerns about healthcare, education, and local infrastructure (d_1 and d_3), divided by genuine disagreement on social values (d_2) and institutional trust (d_5), and separated far more by perception than by position.
DISTRICT 7 — CHAPTER SUMMARY
We have proved the Polarization Illusion Theorem: a unimodal distribution on the manifold can appear bimodal on a 1D projection. Applied to District 7, the theorem reveals that the district’s apparent “deep divide” is a composite of three distinct phenomena: genuine sorting (party ID predicts manifold position), moderate policy polarization (centroid distance 1.5 sigma), and severe affective distortion (perceived distance 3x actual distance). The “polarized nation” narrative is not false, but it is geometrically misleading — it conflates a 1D projection artifact with a manifold-level reality.
In Chapter 9, we turn to gerrymandering: the deliberate manipulation of the political manifold’s topology through redistricting.