Adaptive finite element methods refine meshes where the error is largest. The error estimator tells the algorithm where to refine. Standard residual estimators compare the computed solution to the equation it should satisfy and measure the discrepancy. These estimators work, but they carry a constant that grows with polynomial degree p — higher-order elements give more accuracy per element but make the error harder to estimate.
The paper proves that equilibrated-flux estimators are immune to this scaling. The contraction factor of the adaptive algorithm stays bounded independent of p. Whether you use piecewise linear, cubic, or degree-twenty polynomials, the estimator's reliability doesn't degrade.
The immunity comes from the estimator's construction. Instead of comparing the solution to the residual (which depends on the approximation space), the equilibrated-flux estimator constructs a locally equilibrated vector field — one that satisfies the conservation law element by element — and measures the distance between this field and the computed flux. The construction works with local Neumann problems on element patches, and the equilibrium constraint is physical, not numerical. It doesn't reference the polynomial degree at all.
This is not the same as saying the error is small. It says the estimate of the error is reliable, regardless of how the approximation was built. A good estimator that scales with p gives you less and less useful information as p increases — eventually, the guaranteed bound is too loose to guide refinement. An estimator that doesn't scale with p gives equally useful guidance at any polynomial order.
The through-claim: robustness follows from what the estimator measures. Residual estimators measure the distance to the approximation space, so they inherit the space's complexity. Equilibrated-flux estimators measure the distance to physical law, which doesn't get more complex as the numerics do. The choice of what to compare against determines whether complexity propagates into the estimate. Physics doesn't scale with p. Neither does the estimator built from it.