1 Semi-log plots and exponential relationships
1.1 Definition of semi-log plotting
A semi-log plot is a graph in which one axis uses a logarithmic scale while the other remains linear. Most often, the dependent variable is placed on the logarithmic axis. When the data follow an exponential relationship with respect to the linear variable, this scale pairing tends to convert the curved exponential trend into a straight line, simplifying interpretation and parameter extraction.
1.2 Exponential models suitable for semi-log visualization
Semi-log visualization is well suited to models where the quantity changes proportionally to its current value, producing the general form \[ y(x)=A e^{kx} \] for constants \(A\) (scale factor), \(k\) (rate constant), and independent variable \(x\) (e.g., time). Equivalent expressions using other exponential bases can also be displayed. More complex combinations, such as piecewise exponentials or mixtures of processes, may appear as multiple straight segments or as departures from linearity.
1.3 How exponentials become linear under log transformation
The core idea is that taking a logarithm transforms multiplication into addition and powers into products. For the exponential model \(y=Ae^{kx}\) with \(y>0\), applying the natural logarithm yields \[ \ln y=\ln A+kx. \] This is the equation of a straight line in variables \(\ln y\) versus \(x\), with slope \(k\) and intercept \(\ln A\). As a result, semi-log plots provide a graphical route to estimate the parameters that define exponential behavior.
2 Mathematical foundation
2.1 Transforming \( y = Ae^{kx} \) (or equivalent forms)
Let \(y(x)=Ae^{kx}\) and assume \(A>0\) and \(y>0\). Using a logarithm base \(b\), \[ \log_b y=\log_b A + kx\log_b e. \] For \(b=e\) (natural logarithm), the factor \(\log_e e\) equals 1, giving the simplified linear form \(\ln y=\ln A+kx\). For base-10 logarithms, the relationship includes a constant scaling factor between \(k\) and the slope on the transformed plot.
2.2 Interpreting slope and intercept after transformation
On a semi-log plot where \(x\) is linear and \(\log(y)\) is linear, the fitted line can be written as \[ \log_b y = m x + c, \] where \(m\) is the slope and \(c\) is the intercept. For the exponential model \(y=Ae^{kx}\), these parameters relate to the original constants by \[ m = k\log_b e,\quad c=\log_b A. \] Therefore, once \(m\) and \(c\) are determined from the straight-line representation, the exponential parameters can be recovered through \(k=m/ \log_b e\) and \(A=b^c\).
2.3 Handling base-\(e\) vs base-10 logarithms
Logarithms in charts may use either base \(e\) or base 10, depending on the software and conventions. The two choices differ by a constant factor: \[ \log_{10} y = \frac{\ln y}{\ln 10}. \] Consequently, slopes measured under different bases differ by the same multiplicative constant. Interpretations that compare rate constants across figures must account for which logarithm base was used on the plotted axis.
2.4 Conditions for validity (e.g., positive values)
Logarithmic transformation requires that the plotted argument be positive. For \(\ln y\) or \(\log_b y\), data points with \(y\le 0\) cannot be plotted directly on a log axis. Additionally, the underlying exponential relationship assumes a consistent multiplicative form; additive components or saturation can break the strict linear appearance on a semi-log scale.
3 Practical construction of the plot
3.1 Choosing linear vs logarithmic axes
The typical setup places the independent variable \(x\) on the linear axis and the dependent variable \(y\) on the logarithmic axis. This choice is convenient when the model is expressed as an exponential in \(x\). If the model instead exponentiates the other variable, reversing the axes or using a different transformation may be more appropriate. In all cases, the plotted transformation should align with the mathematical form being tested.
3.2 Selecting plot ranges and avoiding invalid points
A practical range should cover the region where the exponential model is expected to hold. Points near the limits of detection or where the response approaches a measurement floor or ceiling may distort the visual linearity. Because log scales cannot represent \(y\le 0\), analysts typically exclude invalid values or apply a justified preprocessing step only when it does not bias the interpretation.
3.3 Labeling conventions for log-transformed axes
Axis labels should clearly indicate that the vertical axis is logarithmic and specify the base if relevant (e.g., \(\log_{10}(y)\) or \(\ln(y)\), or “log scale” with accompanying notes). Clear labeling also includes units for \(x\) and \(y\), since the parameters inferred from a slope depend on how \(x\) is measured. When reporting, it is common to describe the transformation used to generate linearity, not just the presence of a log axis.
3.4 Spreadsheet and software workflow
Most data tools support semi-log plotting via an option such as “log y-axis.” The workflow typically includes: (1) choosing the axes (linear \(x\), log \(y\)), (2) selecting appropriate tick marks and limits, (3) plotting the raw data, and (4) optionally overlaying a fitted line derived from regression on the transformed values (e.g., fitting \(\ln y\) versus \(x\)). When software offers both a “log-axis” and “log-transformed data” mode, the analysis should specify which approach was used, since regression results can differ if fitting is done after transformation rather than by transforming axis display alone.
3.5 Common formatting and scaling pitfalls
Common issues include forgetting that the straight-line property depends on the log transformation of the dependent variable, not merely on the display. Another pitfall is inconsistent axis bases across figures, which can shift slopes and inferred rates. Analysts also sometimes choose tick spacing that visually exaggerates curvature or, conversely, hides systematic deviations. For interpretability, axis limits should avoid compressing the region of interest into a small portion of the plot.
4 Reading and extracting parameters
4.1 Estimating the exponential rate constant from slope
If the semi-log plot corresponds to \(\ln y\) versus \(x\), then the slope equals the rate constant \(k\) directly for the model \(y=Ae^{kx}\). Under base-10 plotting, the slope \(m\) corresponds to \(m=k\log_{10}e\), so the conversion \(k=m/ \log_{10}e\) is required. In practice, analysts often perform a regression on the log-transformed values to obtain \(m\) (and thus \(k\)) more reliably than reading it by eye.
4.2 Estimating the scale factor from intercept
With the natural-log form, the intercept of the fitted line in \(\ln y=\ln A+kx\) is \(\ln A\). Exponentiating that intercept returns the scale factor \(A\). Using base-10, if \(\log_{10} y=mx+c\), then \(c=\log_{10}A\), so \(A=10^c\). This step links the visual fit on the transformed plot to the original exponential amplitude in the original units of \(y\).
4.3 Determining doubling/halving times (if applicable)
For growth (\(k>0\)), the doubling time \(t_d\) satisfies \(y(x+t_d)=2y(x)\). Under \(y=Ae^{kx}\), \[ \frac{y(x+t_d)}{y(x)} = e^{k t_d}=2 \quad\Rightarrow\quad t_d=\frac{\ln 2}{k}. \]
| For decay (\(k<0\)), the halving time has the same form with absolute value, \(t_{1/2}=\ln 2/ | k | \). Semi-log plots provide \(k\), enabling these times to be computed directly. |
|---|
4.4 Uncertainty considerations in parameter estimates
Parameter uncertainty depends on measurement noise and how errors propagate through the log transformation. Because \(\log(y)\) changes more rapidly when \(y\) is small, the same relative error in \(y\) can yield different variances in \(\log(y)\) across the range. Good practice includes fitting methods that estimate standard errors for slope and intercept on the transformed scale, then converting them back to uncertainties in \(k\) and \(A\) using appropriate transformations (e.g., via error propagation or resampling).
5 Diagnosing model fit and deviations
5.1 Recognizing perfect-line behavior vs curvature
For ideal exponential data without additional effects, the semi-log plot displays a straight line over the region where the model holds. Curvature indicates that the process is not purely exponential across the full range. A concave or convex bend may suggest a changing effective rate, parameter drift, or the presence of additive or saturating components.
5.2 Systematic deviations (offsets, multi-phase processes)
A consistent offset from the straight line can occur when the data follow a related but different form, such as \(y=Ae^{kx}+B\). Mixed processes—such as an initial fast transient followed by slower long-term behavior—can produce piecewise straight segments with different slopes. In those cases, a single global exponential fit may yield misleading “average” parameters.
5.3 Effects of noise and measurement floor/ceiling
When the signal approaches the detection limit or saturates near a maximum response, the semi-log representation often deviates from linearity. Measurement floors can truncate small values, biasing curvature toward the high side on the log axis. Similarly, saturation can flatten the curve, producing bending that might be misread as a change in the underlying growth/decay mechanism rather than an instrument or process constraint.
5.4 Testing alternative models (e.g., power laws)
Not all non-linear behavior on a semi-log plot implies failure of modeling; it may indicate a different functional form. For example, power-law growth/decay can appear non-linear on a semi-log scale. Analysts may compare candidate models using both visual diagnostics and quantitative goodness-of-fit measures, choosing the transformation and model that best reflect the data-generating mechanism.
6 Statistical and methodological considerations
6.1 Regression methods for log-transformed data
When fitting an exponential model using semi-log plotting, a standard approach is regression on the transformed response: fit \(\log(y)\) as a linear function of \(x\). This corresponds to assuming a particular error structure in the original scale, often approximately multiplicative noise. The fitted line on the transformed scale should then be mapped back to the original model \(y=Ae^{kx}\). If the analysis objective involves unbiased prediction on the original scale, bias corrections may be needed depending on the noise distribution.
6.2 Weighting and error propagation under log scaling
Because transformation alters the relationship between measurement error in \(y\) and variability in \(\log(y)\), ordinary least squares on log-transformed data may not correspond to minimizing error on the original scale. Weighted regression can be used when measurement uncertainty varies with \(y\). In many practical situations, relative errors are closer to constant, which translates into approximately constant variance in \(\log(y)\) and motivates weighting choices accordingly.
6.3 Comparing goodness-of-fit across transformations
Goodness-of-fit can depend on the scale on which residuals are computed. A model that looks linear on a semi-log plot may still perform poorly under a likelihood or error metric defined on the original scale. Comparisons across transformations should be done carefully, using consistent criteria (e.g., information criteria based on a specified noise model) or residual analysis that matches the modeling assumptions.
6.4 Residual analysis on the appropriate scale
Residuals should be examined where the model assumptions are most appropriate. Plotting residuals versus \(x\) on the transformed scale can reveal systematic trends indicating violations of the exponential form or non-constant variance. Alternatively, residuals on the original scale can show whether the model captures the magnitude of deviations in a way consistent with the noise structure. A coherent diagnostic strategy uses both views when feasible.
7 Special cases and edge conditions
7.1 Dealing with zero or negative observations
Log axes exclude non-positive values. In datasets that include zeros, analysts must decide whether zero is a meaningful value (not just “missing”) or a censoring outcome. Options include removing those points (with potential bias), using a justified substitution (often problematic if it distorts the functional form), or adopting a model that naturally handles zeros (e.g., processes with additive components). Negative observations similarly require reevaluation because they cannot be represented on a logarithmic scale without redefining variables.
7.2 Shifts and additive terms (e.g., \( y = A e^{kx} + c \))
Additive offsets break the direct linearization of \(\ln y\) versus \(x\). For \(y=Ae^{kx}+c\), taking a logarithm yields \(\ln(Ae^{kx}+c)\), which is not equal to \(\ln A+kx\) unless \(c=0\) or \(Ae^{kx}\gg c\) over the relevant range. A common diagnostic is that semi-log plots show curvature that can change with \(x\), consistent with an additive term becoming dominant at one end of the range.
7.3 Saturation effects and effective exponentials
Saturation mechanisms—where growth slows as a maximum is approached—often yield sigmoidal behavior on linear axes and pronounced deviation from linearity on semi-log plots. In some windows, the early or late regime may approximate an exponential, resulting in an approximately straight segment. Interpreting the straight portion as a constant-rate process is reasonable only if the selected range corresponds to the regime where the exponential approximation holds.
7.4 Units and dimensional consistency after logging
The argument of a logarithm must be dimensionless in strict mathematical terms. In practice, plotting \(y\) (with units) on a log axis implicitly uses a convention that divides by a reference value with the same units, such as \(y/y_0\). Reporting should clarify the quantity being logged (e.g., \(\ln(y)\) of a dimensionless ratio versus \(\ln\) of a raw variable interpreted via a reference). Dimensional care ensures that extracted parameters correspond to the intended physical or measurement units.
8 Reporting and reproducibility
8.1 Best practices for documenting transformations
A reproducible report specifies: which axis was logarithmic, the log base (if known), the mathematical form of the model being tested, and whether regression was performed on transformed data. It also helps to mention how non-positive values were treated and what range of data was included in the fit.
8.2 How to cite axis scaling and transformation details
Figures should include explicit axis labels such as “\(\log_{10}(y)\)” or “\(\ln(y)\)” when possible, or clearly state “log scale” together with a note about the base used by the plotting software. Captions and methods sections can supplement this by describing the transformation used to obtain the straight-line expectation.
8.3 Recreating plots from raw data
To recreate results, one should provide the raw dataset (or a reference to it), the preprocessing steps, and the exact plotting and fitting parameters. If the model fit is obtained via regression on \(\log(y)\), the regression specification (e.g., which points were included and what weighting method was used) is necessary to reproduce the slope and intercept.
8.4 Typical checklist for semi-log exponential figures
A semi-log exponential figure is typically accompanied by: (1) clear axis labeling with log base and units, (2) a statement of the model form being evaluated, (3) the fitted line equation on the transformed scale, (4) uncertainty reporting for parameters when available, (5) notes about excluded points or invalid values, and (6) identification of the data range where linearity is assessed. Together, these elements allow other analysts to verify both the graphical diagnosis and the extracted parameters.