Technological change cleanly splits into an 'earnings' gradient and a 'reallocation' divergence, allowing precise mapping from tech shocks to wages and worker sorting; calibrated to U.S. data, cognitive skill–biased technological change meaningfully reshapes occupational sorting and earnings patterns.

Main Finding

Any infinitesimal technological change in a multidimensional worker–job assignment model uniquely decomposes into two orthogonal components (via the Helmholtz decomposition). The gradient component (symmetric part) fully passes through into marginal earnings without changing the assignment; the divergence‑free component (antisymmetric part) induces pure reallocation of workers across jobs without changing marginal earnings. Equilibrium comparative statics reduce to solving a Poisson equation (equivalently a quadratic “gradient regression”) that pins down the earnings response and, implicitly, the reallocation response. Complementarities and the skill distribution determine how much of a general technological change goes into each margin. Empirically, for U.S. data and a cognitive‑skill–biased shock, the authors find simultaneous sorting and heterogeneous earnings responses consistent with this decomposition.

Key Points

• Model set-up

  • Continuum of workers and jobs, each with d-dimensional types; output yt(x,z) depends on technology t.
  • Competitive equilibrium characterized by optimal assignment (primal) and payoff splitting (dual). Firm first‑order condition equates marginal product to marginal earnings: ∇1yt(x,z) = ∇wt(x) on the assignment support.
  • Assumptions: analytic densities and production, Ma–Trudinger–Wang (MTW) and twist conditions to ensure uniqueness and smoothness of earnings and assignment maps.

• Helmholtz decomposition of technological change

  • Any vector field of marginal product changes ( ˙y-gradient in worker space) can be uniquely written as gradient + divergence‑free parts.
  • Interpretation:
    • Gradient component: change in marginal earnings (∇˙w) — no reallocation.
    • Divergence‑free component: reallocation flow — leaves marginal earnings unchanged.
  • Uniqueness implies a complete characterization of comparative statics without recomputing the full reassigned equilibrium.

• Poisson equation and quadratic characterization

  • Solving for the earnings change reduces to a Poisson PDE on the worker space with coefficients determined by local complementarities and the initial skill density.
  • The PDE solution is equivalently the solution to a quadratic optimization: the best L2 fit of technological change by a gradient field (a “gradient regression”).
  • Complementarity effect: stronger worker–job complementarities bias adjustments toward earnings changes; weaker complementarities bias them toward reallocation. Abundant worker types are costlier to reallocate, increasing passthrough to earnings.

• Special cases and intuition

  • Unidimensional case: technological changes are always gradients → full passthrough to wages, no reallocation.
  • Bilinear / rotationally invariant example: decomposition reduces to symmetric/antisymmetric matrix split (Sylvester eqn); in 2D the antisymmetric part implies rotations (workers shift toward jobs emphasizing their relatively stronger skill).
  • Symmetric technological change → pure earnings adjustment. Antisymmetric → pure reallocation.

• Empirical application (U.S.)

  • Infer worker skill distribution from ACS earnings and occupation data; estimated skill distribution is non‑Gaussian (multimodal, skewed, heavy‑tailed).
  • Simulate cognitive skill‑biased technological change (increasing marginal product of cognitive skill only):
  • High‑cognitive workers tend to replace those with more manual skill; low‑cognitive workers shift opposite.
  • Marginal earnings for cognitive skills rise for all workers, with stronger increases for high‑ability workers (gradient component).
  • Marginal earnings for manual skills rise for workers with high cognitive skill and fall for those with low cognitive skill, despite no direct change in the marginal product of manual skills (interaction via reallocation).

• Additional theoretical results

  • Equivalent characterization of reallocation via maximizing second‑order aggregate output gain.
  • The decomposition and PDE approach can be applied along a trajectory to cover large technological changes by integrating local responses.

Data & Methods

• Theoretical methods

  • Continuous optimal transport / assignment framework (primal and dual Monge–Kantorovich formulation).
  • Use of PDE tools: Helmholtz decomposition, Poisson equation for determining the gradient component of the shock.
  • Quadratic optimization characterization: projection of the technology perturbation onto the space of gradients weighted by densities and local curvature (complementarities).
  • Closed‑form algebraic simplification for bilinear technologies and rotationally invariant densities (Sylvester / symmetric–antisymmetric matrix split).

• Empirical methods (as implemented)

  • Data: American Community Survey (ACS) — earnings and occupation information used to infer a 2D skill distribution (cognitive and manual).
  • Inference approach: map occupation/earnings information into a continuous skill distribution (details of mapping and estimation are in the paper; the main reported features are multimodality, skewness, heavy tails).
  • Counterfactual: impose a cognitive‑skill–biased marginal product change and compute local comparative statics via the Poisson / quadratic method to obtain predicted earnings and assignment changes without recomputing a global nonlinear equilibrium.

• Assumptions & limitations of methods

  • Deterministic assignment (no idiosyncratic matching shocks), continuous smooth distributions, and MTW/twist regularity.
  • Comparative statics are local (infinitesimal), but the authors show how to integrate along a path for large changes.
  • No explicit search frictions, unemployment, or within‑occupation heterogeneity in the baseline deterministic framework.

Implications for AI Economics

• Conceptual framing for AI shocks

  • AI/automation often constitutes multidimensional technological change (affecting cognitive vs manual, routine vs nonroutine, or fine‑grained task complementarities). This paper gives a clean decomposition: AI improvements can be separated into (i) components that directly raise marginal earnings (wage passthrough) and (ii) components that primarily reallocate workers across jobs (employment / task reallocation).
  • Empirical policy and interpretation: observed wage changes and worker flows in response to AI can arise from different parts of the same underlying technical change; observing large employment shifts with little wage change suggests a large divergence‑free component, whereas widespread wage increases with little reallocation suggest a gradient (symmetric) component.

• Practical diagnostics and estimation

  • The quadratic “gradient regression” / Poisson PDE provides an implementable way to decompose an estimated technological shock (e.g., task‑level productivity changes inferred from automation exposure, or AI performance gains mapped to tasks) into wage‑passthrough vs reallocation parts.
  • Estimation requires (i) multidimensional skill measures, (ii) estimates of local complementarities (second derivatives of output wrt skills or surrogates), and (iii) the skill density. O*NET plus occupation–skill mappings or rich microdata like ACS can be used, but results are sensitive to the skill mapping and smoothness assumptions.

• Predictions for inequality and labor market adjustment

  • Inequality: gradient components that raise marginal returns more for high‑skill types will amplify wage dispersion; divergence‑free components can produce heterogeneous directional job flows that create compositional wage effects even where direct marginal products are unchanged.
  • Complementarity matters: when AI increases tasks that strongly complement a given skill, wages for those skills rise (passthrough); when complementarities are weak, AI mainly reallocates labor across tasks/occupations.
  • Policy relevance: if AI generates large divergence‑free shocks, policy should prioritize mobility, retraining, and matching rather than wage subsidies; if gradient shocks dominate, redistribution and compensation policies may be more relevant.

• Modeling advice for AI economists

  • Avoid relying solely on unidimensional frameworks: they cannot capture divergence‑free reallocation effects (they predict full passthrough).
  • Use this decomposition to interpret empirical mixtures of wage and employment responses, and to design counterfactuals that separately evaluate wage and reallocation channels.
  • Consider estimating the gradient projection of AI performance changes to quantify the expected direct wage impact versus the expected need for reallocation support.

Limitations to bear in mind - Results are local (infinitesimal) though a path integration extension is available; large discrete shocks may require careful application. - The baseline framework abstracts from frictions (search, mobility costs), stochastic matching noise, and firm dynamics; these can attenuate or complicate the pure decomposition in practice. - Empirical implementation hinges on credible multidimensional skill measures and estimates of complementarities.

Overall, this paper provides a sharp, general, and constructive framework to decompose how AI‑type technological changes translate into wage vs reallocation margins in multidimensional labor markets, and supplies practical tools (PDE / quadratic projection) for empirical application.