A new research uses probability theory — and sixty years of U.S. economic data — to test one of the most consequential (and most overlooked) assumptions in political economy.
The Assumption Hiding in Plain Sight
If you’ve ever read a Marxist analysis of how profits equalize across industries, you’ve probably encountered something called the average rate of profit. The idea is straightforward: competition between capitalists drives different rates of profit in different sectors toward a common, system-wide average. This is one of the pillars of Marx’s theory of value in Capital, Volume III.
But there’s a quieter assumption underneath this one — so quiet that most discussions never mention it explicitly. To arrive at a uniform profit rate, Marx first assumes a uniform rate of surplus value across all productive sectors. In plainer terms: he assumes that the degree to which workers are exploited — the ratio of unpaid labor to paid labor — is roughly the same everywhere, whether you work in steel manufacturing, food processing, or textiles.
Adam Smith proposed this idea before Marx. Smith argued that if one job were obviously more exploitative (in the sense of yielding far more unpaid surplus per dollar of wages paid), workers and capital would flow toward or away from it until the differences vanished. Marx adopted this observation and, as scholar Jonathan Cogliano notes, elevated it to “the status of a central economic law” within his framework.
Yet the assumption has been challenged from multiple directions — Marxist and non-Marxist alike. Is it actually justified? Or is it a convenient simplification that distorts our understanding of how capitalism works?
José Mauricio Gómez Julián, of the Universidad Latina de Costa Rica, decided to approach the question from an unexpected angle: probability theory. His paper, published in Ciencia Económica (2022), asks whether the mathematical law that would need to hold for this assumption to be valid actually does hold — and then checks the answer against six decades of real-world data from the United States.
The Mathematical Backbone: The Law of Large Numbers
If you’ve taken any statistics course, you’ve likely met the Law of Large Numbers (LLN). It tells us that as you observe more and more instances of something random — coin flips, dice rolls, stock returns — the average of those observations settles down toward the true expected value.
There are two versions:
- The Weak Law (WLLN): With enough observations, the sample average is probably close to the expected value.
- The Strong Law (SLLN): With enough observations, the sample average is almost certainly equal to the expected value — a much stronger guarantee.
Gómez Julián’s insight is this: if you think of each productive sector of the economy as a random variable representing that sector’s rate of surplus value, then the LLN tells you what happens to the average across sectors as the number of sectors grows large. In mathematical language:
- Strong Law: The probability that the average surplus-value rate across sectors equals the global expected value, in the limit, is exactly 1.
- Weak Law: The probability that the average deviates from the global expected value by more than any tiny amount shrinks to zero.
If either version holds, you get the result Marx needs: across a sufficiently large number of sectors, the rates of surplus value converge to a common value — uniformity, or at least a powerful tendency toward it.
The Catch: Independence and Identical Distributions
Here’s where things get interesting — and where the classical LLN hits a wall.
The textbook version of the LLN requires two conditions:
- Independence: The random variables (sectoral surplus-value rates) must be statistically independent of each other.
- Identical distribution: Each variable must follow the same probability distribution.
Neither condition holds for the real economy. And Gómez Julián is admirably upfront about this. Sectors are deeply intertwined — the steel industry depends on mining, manufacturing depends on steel, services depend on consumer spending powered by manufacturing wages. The idea that one sector’s surplus-value rate has no relationship to another’s is economically unrealistic. Furthermore, different industries have different cost structures, different labor intensities, and different technologies. There is no reason their surplus-value rates should follow the same statistical distribution.
So does this kill the argument? Not at all. In fact, it’s the most intellectually interesting part of the paper.
Non-Classical Varieties: When the Rules Relax
Over the past several decades, mathematicians and econometricians have developed non-classical versions of the LLN that weaken or entirely drop the independence and identical-distribution requirements. Gómez Julián surveys several of these:
- Li, Rao, and Wang (1995) showed the LLN holds for random variables arranged on a lattice structure under certain conditions — a structure that, as it happens, economic data naturally exhibits.
- Adler and Rosalsky (1987) proved the law for weighted sums of independent, identically distributed random variables belonging to a normalized sum, generalizing the classical case.
- Chen and Sung (2016) extended those results further: the variables no longer need to be identically distributed. They only need to be “stochastically dominated” by a single random variable, with certain weighting conditions.
- Sung (2011) showed that the strong law can hold even when variables are dependent on each other, provided their probability moments (roughly, their averages and variability) satisfy certain finiteness conditions.
The crucial point: these results collectively tell us that the LLN’s convergence conclusion can survive even when the classical assumptions are substantially violated — which is exactly the situation with sectoral surplus-value rates.
Gómez Julián argues that the economic dynamics described by Smith — workers and capital moving between sectors in response to unequal advantages — are precisely the kind of compensatory dependence mechanism that these non-classical versions accommodate. The variables aren’t independent, but their dependence is structured in a way that still drives convergence.
What the Data Actually Shows
The theoretical argument is compelling, but Gómez Julián doesn’t stop there. He turns to sixty years of U.S. data (1960–2020), sourced from the Bureau of Economic Analysis (BEA), to see what the empirical evidence says.
He calculates sectoral surplus-value rates using macroeconomic data on gross operating surplus (representing surplus labor time) and employee compensation (representing necessary labor time), following a standard operationalization of Marx’s categories. After carefully determining which sectors qualify as “productive” in the Marxist sense — a nontrivial task, since the service sector includes activities with very different relationships to surplus-value production — he arrives at 36 productive sectors.
Here’s what the statistical analysis found:
Finding 1: No Identical Distributions
A probability distribution fitting exercise (using the Bayesian Information Criterion) revealed that the 36 sectors’ surplus-value rates follow a patchwork of different distributions — Log-Normal, Cauchy, Uniform, Weibull, and Logistic — with none following a normal distribution. The identical-distribution requirement of the classical LLN is not met.
Finding 2: No Statistical Independence
A Pearson correlation analysis across all 630 possible sector pairs yielded a mean correlation of about 0.08 and a median of about 0.14. While these may look small, a deeper cut reveals that roughly 40% of sector pairs have correlations of 0.3 or above — a level that’s practically meaningful. The sectors are not independent. This makes intuitive sense: industries are connected through supply chains, labor markets, and shared macroeconomic conditions.
Finding 3: Differences Tend Toward Zero
This is the key finding. When Gómez Julián computed the differences between each sector’s surplus-value rate and the global average (both the mean and the median), he found that these differences exhibit a strong tendency toward reciprocal nullification — positive differences roughly cancel out negative ones. The sum of all differences relative to the global mean was essentially zero (on the order of 10⁻¹⁴). The mean of differences relative to the global median was 0.0012 — vanishingly small.
Distributional fitting on these differences revealed they follow a Cauchy distribution (when measured against the global mean) or a uniform distribution (against the global median), with the medians of these distributions sitting very close to zero.
In plain language: sectors deviate from the average in different directions, and those deviations largely cancel each other out.
Why This Matters
Gómez Julián’s paper makes two types of contributions that are worth distinguishing:
For Marxist political economy: If the uniformity assumption holds — even approximately, even as a tendency rather than an iron law — then a large body of research on the long-run behavior of the average rate of profit, both within countries and across the global economy, is on sounder footing than critics have suggested. Researchers studying the tendency of the rate of profit to fall (or not) can continue to work without needing to explicitly model sector-by-sector differences in exploitation rates, at least for aggregate, long-run analyses.
For probability theory and economics: The paper demonstrates a productive intersection between a specific question in political economy and the deep mathematics of convergence theorems. It shows that the non-classical LLN theorems aren’t just abstract curiosities — they have direct relevance to understanding real economic phenomena. The structured dependence between economic sectors isn’t a bug that invalidates the mathematical framework; it’s a feature that the right version of the framework already accounts for.
A Few Honest Caveats — And Why They No Longer Apply
The original 2022 paper was unusually transparent about its limitations, and that transparency is one of its strengths. Rather than forcing the data into inappropriate statistical procedures, it openly acknowledged where the available inferential tools broke down.
At the time, three important caveats remained.
First, formal hypothesis testing had to be abandoned.
The reason was purely statistical rather than economic. Classical inferential procedures—Student’s t tests, Wilcoxon tests, and most conventional non-parametric alternatives—are built on assumptions that the data simply did not satisfy. Sectoral surplus-value rates are neither independent nor identically distributed. They are linked through supply chains, technological change, capital mobility, and macroeconomic shocks. Even bootstrap procedures could not fully solve the problem because ordinary resampling may weaken dependence between resamples while leaving the internal dependence structure fundamentally unchanged. Consequently, the 2022 paper relied primarily on descriptive statistics together with probability-theoretic arguments instead of formal significance testing.
Second, the classification of productive sectors inevitably involved theoretical judgment.
Although the paper carefully justified the inclusion and exclusion of economic activities using Marxian categories and modern national accounting, reasonable scholars could still debate where certain services belong within the circuit of capital.
Third, the empirical evidence came exclusively from the United States.
The descriptive regularities were remarkably strong, but demonstrating that the same convergence mechanism operates under different institutional settings naturally remained an empirical question.
Those were genuine limitations in 2022.
Today, however, the first—and arguably the most important—of them has largely been overcome.
A much more comprehensive methodological paper (Gómez Julián, 2026; SSRN 5172185) develops an entirely new inferential framework specifically designed for exactly the type of data that made the original analysis difficult: dependent, heterogeneous, and unbalanced observations. Instead of trying to force classical statistical tests to work outside the assumptions under which they were derived, the newer paper constructs hypothesis testing from the ground up for this class of problems.
The key innovation is recognizing that the convergence of sectoral surplus-value rates is fundamentally a law-of-large-numbers problem under dependence, not an independent-samples problem. The framework therefore combines three complementary asymptotic structures—triangular arrays (TAC), correlation-weighted sums (WSC), and mixingale processes (MPC)—which respectively model hierarchical dependence, contemporaneous intersectoral dependence, and temporal dependence. Rather than treating these as competing approaches, the paper proves conditions under which they become metrically equivalent and therefore support the same inferential conclusions.
The inferential consequences are substantial.
Instead of abandoning significance testing because dependence invalidates classical procedures, the new framework explicitly extends the Neyman-Pearson paradigm to dependent observations, derives dependence-aware confidence regions, establishes rigorous Type I error control under strong-mixing assumptions, and integrates Bayesian and frequentist inference into a single coherent architecture. Robust procedures—including fixed-b heteroskedasticity-and-autocorrelation-robust inference, block bootstrap techniques that preserve dependence, adaptive conformal inference, composite and Whittle likelihoods, and hierarchical Bayesian estimation—serve as mutually reinforcing validation mechanisms rather than isolated alternatives.
In other words, what had been acknowledged as a methodological limitation in the 2022 paper became the central research question of the later work.
Rather than concluding that inference was impossible under dependence, the subsequent research asks a more fundamental question: what should hypothesis testing look like when dependence is the normal state of the data rather than an exception? The result is a unified inferential framework specifically intended for datasets that violate the assumptions of classical statistics—precisely the situation encountered with sectoral surplus-value rates.
The other caveats remain, although they are considerably less problematic than before. The classification of productive sectors continues to depend on theoretical interpretation, because that issue belongs to political economy rather than statistics. Likewise, expanding the empirical analysis to additional countries remains a desirable avenue for future research. Yet the principal statistical objection—that no valid inferential procedure existed for dependent sectoral data—has now been directly addressed through a purpose-built mathematical framework.
Looking back, the 2022 paper can therefore be read as identifying an important statistical obstacle, while the later work attempts to remove it. Together, the two papers form a coherent research program: first demonstrating that the convergence hypothesis is theoretically plausible and descriptively supported, and then developing the inferential machinery required to test that hypothesis rigorously without relying on unrealistic assumptions of independence or identical distributions.
Gómez Julián, J.M. (2022). Sobre la validez del supuesto de uniformidad en las tasas de plusvalía sectorial desde la teoría de las probabilidades. Ciencia Económica, 11(17). DOI: 10.22201/fe.24484962e.2022.11.17.2
Gómez Julián, J.M. (2026). Hypothesis Testing for Dependent Variables with Unbalanced Data: A Unified Framework: Theory, Robustness, and Software. SSRN Electronic Journal. DOI: 10.2139/ssrn.5172185.
