Can an unshifted log-normal variable be negative?

No. Its values are strictly positive.

Is μ the average of the original observations?

No. In this parameterization it is the mean of their natural logarithms.

Does taking logs prove the model is appropriate?

No. The transformed observations still need to be assessed for compatibility with a normal model and the sampling assumptions.

AI Glossary / Computer vision and multimodal AI / Intermediate

Log-normal distribution

Definition

A log-normal distribution describes a positive random quantity whose logarithm follows a normal distribution. The original values are typically right-skewed: many observations are relatively small, with a longer tail toward large values. It is a probability model, not simply another name for any skewed dataset.

What it means

If Y = ln(X) is normally distributed with mean μ and standard deviation σ, then X is log-normal. Here ln is the natural logarithm. The parameters describe the logarithms, so μ is not the ordinary mean of X and σ is not its ordinary standard deviation.

This distinction matters when interpreting measurements such as processing durations, object sizes, or positive multiplicative factors. A model that looks symmetric after taking logarithms can look strongly asymmetric in the original units.

This entry uses the standard unshifted, two-parameter family. Some references add a location or shift parameter, changing the lower boundary. Always check the parameterization before comparing software outputs or copying a formula.

How it works

Consider a positive quantity affected by many multiplicative factors. Taking logarithms converts their product into a sum. Under suitable conditions, that sum may be approximately normal, making a log-normal model a plausible candidate for the original quantity. This is a motivation to test the model, not a guarantee that it fits.

For positive observations, inspect both the original values and their logarithms. Estimate parameters using an appropriate statistical method and examine fit, especially in the tail relevant to the application.

To simulate the unshifted distribution, generate a normal value Y with the chosen μ and σ, then calculate X = exp(Y). To obtain a percentile, transform the corresponding normal percentile back with the exponential. Keep units and any reference scaling explicit when interpreting logarithms of measurements.

Three practical examples

1. Image-processing latency: A team considers a log-normal model for positive request durations. It checks the fit separately for small and large images because mixing different workloads can create a misleading overall shape.

2. Object-size analysis: A researcher studies positive areas measured from segmented objects. A log-normal candidate may summarize the distribution, but segmentation failures and minimum detectable sizes must be examined first.

3. Synthetic variation: A test designer samples positive multiplicative brightness factors around a chosen median. The factors generate controlled variation; they do not claim that real illumination necessarily follows the same model.

Think of it this way

Think of repeated percentage changes applied to a starting amount. Multiplication by 1.1 and later by 0.9 acts differently from adding and subtracting fixed amounts. A logarithmic scale turns those multiplying factors into additions.

The analogy explains why log-normal models arise naturally in some multiplicative settings. A short sequence of arbitrary percentage changes does not automatically produce a log-normal distribution.

Worked example

Suppose a hypothetical processing-time model has median ten milliseconds and logarithmic standard deviation σ = 0.5. Express X numerically in milliseconds, so μ = ln(10), approximately 2.303.

The median is exp(μ) = 10. The mean is exp(μ + σ²/2), giving 10 × exp(0.125), approximately 11.33 milliseconds. The mean exceeds the median because the upper tail pulls the average upward.

For the 95th percentile, use the standard normal percentile approximately 1.645. The result is exp(μ + 1.645σ) = 10 × exp(0.8225), approximately 22.76 milliseconds. Under this assumed model, about five percent of durations exceed that value.

These figures are model calculations, not measured service guarantees. If actual data contain a separate slow-processing mode, the fitted tail may be inadequate even when the median looks correct. Observed performance must therefore be checked against the intended probability statement.

Illustrative case study

Fictional case: A document-imaging provider models page-processing times with one log-normal curve. The fit looks reasonable across most requests, but the slowest observations occur mainly on damaged scans requiring an additional processing stage.

The analyst separates those workflows and compares their distributions. The team retains the original observations, checks whether each group supports the proposed model, and reports empirical percentiles alongside fitted estimates.

The lesson is that a convenient distribution should not conceal a meaningful process difference. The case is illustrative and does not assert that this model is best for document-processing latency.

Common misconceptions and mistakes

Do not assume that right-skewed data must be log-normal. Gamma, Weibull, mixtures, and other models can produce positive skewed observations.

Do not take logarithms of zero or negative values without examining what those values mean. Adding an arbitrary constant changes the model and can materially affect interpretation.

Do not confuse a normal distribution of logarithms with a logarithmic encoding of an image. One describes probability across observations; the other changes the representation of signal values. Finally, the median and arithmetic mean answer different questions and should not be substituted for one another in reports.

Business use

Log-normal models can help summarize positive quantities, generate simulation inputs, and estimate probabilities when supported by data. They are particularly useful when multiplicative variation offers a plausible explanation of the process.

For business use, define the population being modeled and the consequence of tail errors. Retain evidence about excluded observations, censoring, and measurement limits. Check the fit on fresh data and compare model-based estimates with empirical summaries. A single fitted curve should not replace investigation of operational changes or unusual cases.

Check your understanding

A log-normal model has median ten and mean 11.33. Are these values inconsistent?

Answer: No. Its right-skewed shape can make the mean larger than the median.

Related terms

Learn next

Plot positive observations on both ordinary and logarithmic scales. Compare empirical quantiles with fitted log-normal quantiles, paying particular attention to extremes and any observations limited by the measurement process.

Sources

NIST/SEMATECH Engineering Statistics Handbook, Lognormal Distribution: https://itl.nist.gov/div898/handbook/eda/section3/eda3669.htm

The primary statistical reference was opened to check definitions and parameterization. The numerical calculations, application examples, and provider scenario are original illustrations.

Previous: Log Loss Next: Log-odds