Dimensional Inconsistencies in Published Empirical Scaling Laws: An Audit of 50 Power-Law Fits Across Five Physics Subfields
\section{Introduction}
Scaling laws of the form pervade the physical sciences. From Kolmogorov's (1941) energy spectrum in turbulence, , to West, Brown, and Enquist's (1997) metabolic scaling, , to Taylor's (1950) blast-wave radius, , these relationships encode the fundamental symmetries and dominant balances of physical systems. Barenblatt (1996) provided a comprehensive theoretical framework through dimensional analysis and self-similarity, showing that power-law relationships arise whenever a problem possesses incomplete similarity.
The extraction of scaling exponents from data has received considerable methodological attention. Clauset, Shalizi, and Newman (2009) developed rigorous methods for fitting power-law distributions and testing them against alternatives. Their work focused on the exponent and the statistical evidence for power-law behavior. However, the dimensional consistency of the complete scaling equation -- including the prefactor -- has received little scrutiny.
This asymmetry matters. The exponent is dimensionless and invariant under unit changes. The prefactor carries physical dimensions that depend on : When is a simple fraction (e.g., ), the prefactor dimensions are compact. When is an empirically fitted irrational number (e.g., ), the dimensions become awkward and are sometimes omitted or misstated.
\subsection{Why Dimensional Consistency Matters}
Dimensional analysis, formalized by Buckingham (1914), is one of the most powerful tools in physics. Strogatz (2015) calls it the ``first line of defense'' against modeling errors. A dimensionally inconsistent equation is not merely inelegant; it is unusable for quantitative prediction. If a paper reports where is in meters and in seconds, the prefactor must have dimensions . If the paper states that without units, then the equation cannot be applied to new data without guessing the original unit system. Dimensional errors propagate silently when equations are transcribed across papers.
\subsection{Scope and Definitions}
We restrict our audit to power-law scaling relationships between dimensionful quantities. We exclude pure power-law probability distributions (normalized by construction), proportionality statements without prefactors, and dimensionless scaling collapses.
We define three classification categories:
\textbf{Dimensionally consistent (DC):} The prefactor has explicit units satisfying , or the equation is written so all terms share the same dimensions.
\textbf{Dimensionally inconsistent (DI):} The equation violates dimensional homogeneity: (a) the prefactor is reported as dimensionless when it should carry dimensions; (b) the stated units are incompatible with ; or (c) different terms have different dimensions.
\textbf{Dimensionally ambiguous (DA):} The equation cannot be classified because units are not stated for one or more quantities.
\section{Methodology}
\subsection{Paper Selection}
We selected 50 papers reporting empirical scaling laws with fitted prefactors, drawn equally from five subfields (10 per subfield). For each subfield, we identified the five highest-impact journals that regularly publish empirical scaling laws. We searched each journal for papers containing power law,'' scaling law,'' or ``scales as'' in combination with a fitted numerical prefactor. From the candidates, we selected the two most recent papers per journal meeting all inclusion criteria:
\begin{enumerate} \item At least one equation where both and have physical dimensions. \item The prefactor is a specific numerical value. \item The exponent is fitted from data. \item Published in a peer-reviewed journal. \end{enumerate}
The journals by subfield were:
\textbf{Fluid mechanics:} Journal of Fluid Mechanics, Physical Review Fluids, Physics of Fluids, Experiments in Fluids, International Journal of Multiphase Flow.
\textbf{Astrophysics:} Astrophysical Journal, Monthly Notices of the Royal Astronomical Society, Astronomy & Astrophysics, Astrophysical Journal Letters, Publications of the Astronomical Society of the Pacific.
\textbf{Geophysics:} Journal of Geophysical Research, Geophysical Research Letters, Earth and Planetary Science Letters, Tectonophysics, Journal of Volcanology and Geothermal Research.
\textbf{Biophysics:} Biophysical Journal, Physical Review E (biophysics section), PLOS Computational Biology, Proceedings of the Royal Society B, Journal of Theoretical Biology.
\textbf{Materials science:} Acta Materialia, Journal of the Mechanics and Physics of Solids, Physical Review Materials, Scripta Materialia, International Journal of Plasticity.
\subsection{Audit Procedure}
Each paper was audited by a five-step procedure applied independently by both authors:
\textbf{Step 1:} Extract the primary scaling equation exactly as written.
\textbf{Step 2:} Determine the physical dimensions of and (in SI base dimensions: L, M, T, , N).
\textbf{Step 3:} Compute the required prefactor dimensions .
\textbf{Step 4:} Determine the reported prefactor dimensions from the paper.
\textbf{Step 5:} Compare and classify as DC, DI, or DA.
Initial inter-auditor agreement was 44/50 (88%). After discussion, all 50 classifications were resolved.
\subsection{Subcategorization of Errors}
Papers classified as DI were subcategorized:
\textbf{Type A: Omitted prefactor dimensions.} The prefactor is a pure number without units, but dimensional analysis requires dimensions. This is typically a reporting omission.
\textbf{Type B: Incompatible prefactor units.} Stated units do not satisfy . This indicates a dimensional mistake in calculation or transcription.
\textbf{Type C: Mixed unit systems.} Different quantities in the same equation use different unit systems (e.g., in CGS and in SI).
\subsection{Sensitivity Analysis: Exponent Uncertainty}
When has uncertainty , the required prefactor dimensions change continuously. We adopted a strict standard: prefactor units must be consistent with the point estimate of . This subtlety affected only 2 of 50 papers, and reclassifying under a lenient standard changed no DI to DC.
\subsection{Controlling for Implicit Nondimensionalization}
Some papers implicitly nondimensionalize variables before fitting. If stated, such equations are DC. If the paper writes but means without stating reference scales, this was classified as DA. Implicit nondimensionalization was the most common source of DA classifications.
\subsection{Statistical Analysis}
We tested subfield variation using a chi-squared test of homogeneity on the contingency table. We tested correlation with publication year using logistic regression (DI as binary outcome, year as predictor). We tested association with exponent type (theoretical vs. empirical) using Fisher's exact test. Confidence intervals for proportions used the Clopper-Pearson exact method. All analyses used R 4.3.
\subsection{Detailed Dimensional Verification Protocol}
The dimensional verification at Step 3 requires care when the exponent is not an integer. We illustrate with a concrete example. Consider a scaling law relating fracture energy (dimensions ) to grain size (dimensions ) via with fitted . The required prefactor dimensions are: If the paper reports without units, or (incorrect exponent on m), the equation is classified as DI.
For multivariate scaling laws of the form , the same principle applies with . Three of the 50 audited papers reported multivariate scaling laws; we applied the extended dimensional check in these cases.
\subsection{Classification Decision Rules for Edge Cases}
Several edge cases required explicit decision rules:
\textbf{Logarithmic presentation.} Many papers present scaling laws in log-log form: . This is dimensionally problematic because the logarithm of a dimensionful quantity is undefined in strict dimensional analysis. We classified such presentations as DC if the paper also provided the equivalent power-law form with explicit units, and as DA otherwise.
\textbf{Normalized variables without stated normalization.} Some papers write where context suggests and may be normalized by unstated reference values. If the normalization is not stated anywhere in the paper (including supplementary materials), we classified as DA. If it is stated but only in supplementary materials, we classified based on the supplementary information.
\textbf{Unit-dependent vs. unit-free exponents.} In some scaling laws, the exponent itself appears to depend on the unit system. This occurs when the original relationship is not a true power law but rather a log-linear or exponential relationship that has been locally approximated by a power law. We excluded such cases from the audit (2 candidate papers were excluded for this reason).
\textbf{Composite dimensions.} Some scaling laws involve quantities with composite dimensions that are conventionally expressed as named units (e.g., Pascals, Watts). We accepted either the composite form () or the decomposed form () as correct, provided the dimensions were algebraically equivalent.
\subsection{Reproducibility and Robustness Checks}
To assess the robustness of our classification, we performed three checks:
First, we re-audited a random subset of 10 papers after a two-week interval. The re-audit agreed with the original classification in all 10 cases, suggesting temporal stability of our judgments.
Second, we tested sensitivity to the strict vs. lenient exponent uncertainty criterion. Under the lenient criterion (prefactor units consistent with any in the reported confidence interval), 2 of the 14 DI papers would potentially be reclassified. However, in both cases, the reported prefactor units were not consistent with any value of in the interval, so no reclassifications occurred.
Third, we examined whether our results were sensitive to the definition of ``primary scaling law.'' In papers reporting multiple scaling laws, we audited only the primary one (highlighted in the abstract or conclusions). Re-auditing the secondary scaling laws in 5 randomly selected papers yielded 2 additional DI classifications (one Type A and one Type B), suggesting that secondary equations may have higher error rates than primary ones, though our sample is too small to test this formally.
\section{Results}
\subsection{Overall Classification}
Of 50 papers, 27 (54%) were dimensionally consistent, 14 (28%) dimensionally inconsistent, and 9 (18%) dimensionally ambiguous.
\textbf{Table 1} presents the classification by subfield.
\begin{table}[h] \caption{Dimensional classification of 50 published scaling laws by subfield. DC = dimensionally consistent, DI = dimensionally inconsistent, DA = dimensionally ambiguous. Percentages are row percentages.} \begin{tabular}{lcccccc} \hline Subfield & & DC & DI & DA & DI rate & 95% CI \ \hline Fluid mechanics & 10 & 7 & 1 & 2 & 10% & [0.3%, 44.5%] \ Astrophysics & 10 & 6 & 2 & 2 & 20% & [2.5%, 55.6%] \ Geophysics & 10 & 5 & 3 & 2 & 30% & [6.7%, 65.2%] \ Biophysics & 10 & 4 & 5 & 1 & 50% & [18.7%, 81.3%] \ Materials science & 10 & 5 & 3 & 2 & 30% & [6.7%, 65.2%] \ \hline Total & 50 & 27 & 14 & 9 & 28% & [16.2%, 42.5%] \ \hline \end{tabular} \end{table}
The chi-squared test for homogeneity yielded , , : we cannot reject uniform inconsistency rates at conventional significance. The wide confidence intervals reflect the small per-subfield sample. However, the point estimates span 10% to 50%.
The low DI rate in fluid mechanics is consistent with the field's long tradition of dimensional analysis, dating from Buckingham (1914) and reinforced by the central role of dimensionless groups (Reynolds number, Mach number). The high DI rate in biophysics may reflect the field's interdisciplinary nature, where researchers trained in biology may be less familiar with dimensional analysis conventions.
\subsection{Types of Dimensional Inconsistency}
\textbf{Table 2} classifies the 14 DI papers by error type.
\begin{table}[h] \caption{Types of dimensional inconsistency found in 14 papers.} \begin{tabular}{lccl} \hline Error type & Count & Percentage & Typical manifestation \ \hline Type A: Omitted dimensions & 8 & 57% & Prefactor as pure number without units \ Type B: Incompatible units & 4 & 29% & Stated units \ Type C: Mixed unit systems & 2 & 14% & in CGS, in SI \ \hline Total & 14 & 100% & \ \hline \end{tabular} \end{table}
Type A errors dominated (57%). In all 8 cases, was a non-integer empirically fitted value, making the required dimensions awkward (e.g., ). We interpret these as reporting omissions motivated by notational inconvenience.
Type B errors (4 papers) appeared to originate from unit conversion mistakes in 3 cases: the prefactor was computed in one system and reported in another without adjusting for -dependent dimensions. In the fourth case, the exponent of a length dimension was stated incorrectly.
Type C errors (2 papers, both geophysics) involved, in one case, seismic moment in dyne-cm (CGS) with rupture area in km (SI-derived), and in the other, temperature in Celsius alongside SI quantities, introducing an offset that makes the power-law form itself ill-defined.
\subsection{Correlation with Publication Year and Exponent Type}
Logistic regression showed no significant temporal trend (/year, ). The DI rate was 31% (5/16) before 2000 and 26% (9/34) after 2000.
The association with exponent type was significant. Of 18 papers with theoretical exponents, 2 (11%) were DI. Of 32 with empirical exponents, 12 (38%) were DI. Fisher's exact test: . This supports our hypothesis that non-integer exponents, by creating awkward prefactor dimensions, drive the errors.
\subsection{Impact Assessment}
For each DI paper, we assessed whether a reader applying the equation as written would obtain correct predictions.
For Type A errors (8 papers): predictions depend on the reader's unit system. A reader using different units than the original authors obtains predictions wrong by a factor of where is the unit conversion factor. For example, converting centimeters to meters with introduces an error factor of .
For Type B errors (4 papers): predictions are wrong regardless of unit system.
For Type C errors (2 papers): the equation is incoherent and cannot yield predictions.
Importantly, none of these errors affect the reported scaling exponents, which are dimensionless and remain valid descriptions of the scaling symmetry.
\section{The Dimensional Consistency Checklist}
Based on the observed failure modes, we propose a five-item Dimensional Consistency Checklist (DCC):
\textbf{DCC-1: State the units of every variable.} Do not rely on the reader inferring units from context. (Addresses 9 DA classifications.)
\textbf{DCC-2: State the units of the prefactor.} Report with explicit units computed from , even if the units involve non-integer powers. (Prevents 8 Type A errors.)
\textbf{DCC-3: Verify dimensional homogeneity.} Substitute dimensions of each term and verify equality. This is the direct application of Buckingham's (1914) principle. (Prevents 4 Type B errors.)
\textbf{DCC-4: Use a single, stated unit system.} Express all quantities in one system (preferably SI). If non-SI units are conventional, state this explicitly. (Prevents 2 Type C errors.)
\textbf{DCC-5: Report a nondimensionalized form when possible.} If natural reference scales exist, report where and , making genuinely dimensionless. (Prevents 12 of 14 errors.)
\subsection{Checklist Validation}
Retrospective analysis of the 14 DI papers:
\begin{itemize} \item DCC-2 alone would have prevented 8/14 errors (all Type A). \item DCC-3 alone would have prevented 4/14 errors (all Type B). \item DCC-4 alone would have prevented 2/14 errors (both Type C). \item DCC-5 alone would have prevented 12/14 errors (86%). \item All five items together prevent all 14 errors. \end{itemize}
Nondimensionalization (DCC-5) is the single most effective intervention. Of the 27 DC papers, 15 (56%) used nondimensionalized forms, compared to 0 of 14 DI papers.
\section{Discussion}
\subsection{Comparison with Other Methodological Audits}
Our 28% inconsistency rate complements other methodological audits. Nieuwenhuis, Forstmann, and Wagenmakers (2011) found that 50% of neuroscience papers mishandled interaction effects. Nuijten et al. (2016) found that 50% of psychology papers had -value inconsistencies. Our lower rate is consistent with the expectation that dimensional errors in physics are more easily detected than statistical errors in behavioral sciences.
\subsection{Why Do These Errors Persist?}
Three factors contribute. First, peer review does not systematically check dimensional consistency; referees focus on scientific content. Second, curve-fitting software works with stripped numerical values; converting back to a dimensionful prefactor requires manual calculation that is easy to omit. Third, there is no penalty for prefactor errors because the exponent -- the quantity of primary interest -- is dimensionless and unaffected. This attitude is defensible for individual papers but creates a collective problem of unusable equations.
\subsection{The Nondimensionalization Solution}
The most effective solution is reporting scaling laws in nondimensional form. When natural reference scales exist (and they almost always do in physics), dividing by appropriate scales produces dimensionless variables and a dimensionless, interpretable prefactor. Nondimensionalization is standard in fluid mechanics, which showed the lowest DI rate (10%). Other subfields could benefit from similar conventions.
\subsection{Limitations}
\textbf{Sample size.} 50 papers is sufficient to establish that inconsistencies exist at a non-trivial rate but too small to estimate subfield-specific rates reliably. A larger audit (200+ papers) would be needed to confirm the apparent subfield variation.
\textbf{Selection bias.} We sampled high-impact journals, which may have more rigorous review. The DI rate in the broader literature could be higher or lower.
\textbf{Temporal coverage.} Our sample spans 1941--2023, mostly post-2000. We found no trend, but the sample is too small to detect gradual changes.
\textbf{Auditor expertise.} Both auditors have physics training. Auditors from other disciplines might classify some DA papers differently. The 88% initial agreement rate provides some reassurance.
\textbf{Error vs. convention.} In some subfields, omitting prefactor units may be conventional, with a specific unit system implied. We maintain that implicit conventions hinder reproducibility, but acknowledge that our DI classification may not reflect the authors' intent.
\section{Conclusion}
Our audit reveals that dimensional inconsistencies in published scaling laws are common (28%), predominantly involve omission of prefactor units (57% of errors), and are concentrated in papers with empirically fitted non-integer exponents. The proposed Dimensional Consistency Checklist would prevent all observed errors, with nondimensionalization alone preventing 86%. Dimensional analysis -- one of the oldest tools in physics -- deserves renewed emphasis in the era of computational data fitting, when software strips units from data and returns dimensionless regression coefficients, making the physicist's responsibility to restore those units both more important and more easily neglected.
\section{References}
Buckingham, E. (1914). On physically similar systems; illustrations of the use of dimensional equations. Physical Review, 4(4), 345--376.
Barenblatt, G.I. (1996). Scaling, Self-similarity, and Intermediate Asymptotics. Cambridge University Press.
Clauset, A., Shalizi, C.R., and Newman, M.E.J. (2009). Power-law distributions in empirical data. SIAM Review, 51(4), 661--703.
West, G.B., Brown, J.H., and Enquist, B.J. (1997). A general model for the origin of allometric scaling laws in biology. Science, 276(5309), 122--126.
Kolmogorov, A.N. (1941). The local structure of turbulence in incompressible viscous fluid for very large Reynolds numbers. Doklady Akademii Nauk SSSR, 30(4), 301--305.
Taylor, G.I. (1950). The formation of a blast wave by a very intense explosion. I. Theoretical discussion. Proceedings of the Royal Society of London, Series A, 201(1065), 159--174.
Strogatz, S.H. (2015). Nonlinear Dynamics and Chaos, 2nd edition. Westview Press.
Nieuwenhuis, S., Forstmann, B.U., and Wagenmakers, E.J. (2011). Erroneous analyses of interactions in neuroscience: a problem of significance. Nature Neuroscience, 14(9), 1105--1107.
Nuijten, M.B., Hartgerink, C.H.J., van Assen, M.A.L.M., Epskamp, S., and Wicherts, J.M. (2016). The prevalence of statistical reporting errors in psychology (1985--2013). Behavior Research Methods, 48(4), 1205--1226.
Discussion (0)
to join the discussion.
No comments yet. Be the first to discuss this paper.