One of the most persistent misunderstandings in PMF analysis is the tendency to treat a Q/Qexp ratio close to unity (or a low Q value) as evidence that the solution itself is reliable.
It is not.
Q is an important diagnostic. Q is a measure of fit. A low Q value indicates that the model reproduces the observed data reasonably well within the uncertainty structure supplied to the fit. It says very little about whether the resulting factors are physically meaningful, chemically interpretable, or uniquely resolved.
What Q actually measures — and what it doesn’t
Q is defined as the sum of squared residuals weighted by the uncertainty matrix. Minimizing Q is the objective of the PMF algorithm. Q/Qexp — the ratio of obtained Q to the theoretically expected Q under ideal conditions — provides a normalized benchmark: values close to 1 suggest the model fit is consistent with the stated uncertainties; values substantially above 1 suggest either underestimated uncertainties or model misfit; values substantially below 1 often indicate over-fitting or over-specified uncertainty.
What Q does not encode is any information about the physical interpretability of the solution. The PMF algorithm has no knowledge of atmospheric chemistry, emission source characteristics, or the expected seasonal behavior of traffic emissions in a mid-latitude city. It merely optimizes a mathematical objective. Multiple rotationally equivalent solutions, or solutions with different factor numbers, can achieve similar Q values while producing substantially different, and differently interpretable, factor profiles. The numerical criterion does not discriminate between them. As a consequence, improving Q alone does not necessarily improve the scientific validity of the factorization.
Other important diagnostics
Residual diagnostics, factor behavior, and uncertainty analysis often provide more useful information than Q itself once the fit reaches a reasonable level.
Residual structure. Systematic residual patterns indicate where the model assumptions begin to fail. Systematic patterns in residuals indicate that the model is missing something. Random, homoscedastic residuals suggest the model is capturing the variance structure in the data appropriately. Persistent underprediction or overprediction for particular species, time periods, or concentration regimes frequently signals unresolved sources, factor mixing, or processes not adequately represented within the bilinear framework.
Factor profile and contribution interpretability. PMF factors are ultimately intended to represent real source contributions. A factor profile should therefore correspond to chemically coherent emissions behavior rather than merely satisfying statistical separation criteria. Temporal behavior provides an additional constraint: source contributions should evolve in ways consistent with known atmospheric processes, emissions activity, and meteorological conditions.
None of this implies that Q is unimportant. Poor Q values usually indicate that the model fit itself is inadequate. But beyond a certain point, differences in Q become far less informative than the physical plausibility and stability of the solution.
A PMF analysis should therefore not be evaluated primarily by how low its Q value becomes, but by whether the resulting factorization remains chemically interpretable, temporally coherent, statistically stable, and robust under uncertainty exploration. Good PMF solutions are not defined by good statistics alone. They are defined by consistency between the mathematics of the fit and the physical reality the model is intended to represent.