Studies in Roughness: A Fractal Primer

Measurements disagree when teams change resolution without modeling how the object changes with scale. Fractal roughness turns that disagreement into evidence instead of dismissing it as instrument noise.

Enterprise problem: organizations need metrics that remain interpretable when observation moves from seconds to days, functions to systems, or individual tasks to portfolios.

Whole-course context: after previewing turbulent bursts, the course now develops the geometry and scaling language needed to describe irregular paths across resolutions.

Today’s slice: this chapter builds intuition from coastline measurement, self-similarity, self-affinity, fractal dimension, and the Hurst exponent.

End-of-day evidence: you will calculate a ruler-length scaling estimate, operate roughness labs, and state the finite range over which a pattern holds.

Still unsolved: one roughness exponent cannot capture every market feature; later chapters add tails, long memory, regimes, and multifractal spectra.

Fractal language is often used decoratively. Precise definitions keep “fractal” tied to measurements across scale.

Term	Plain meaning
Scale	Resolution or interval used to observe a system
Self-similarity	Parts resemble the whole after equal rescaling in every direction
Self-affinity	Statistical resemblance requires different scaling along different axes
Fractal dimension	Exponent describing how measured detail grows as resolution becomes finer
Scaling law	Relationship that follows a power of scale over a stated range
Hurst exponent	Exponent H relating typical increment size to time interval
Roughness	Irregularity that persists under magnification
Monofractal	Process summarized by one scaling exponent
Multifractal	Process requiring multiple scaling exponents for different fluctuation sizes

Exact self-similarity is rare in data. Statistical scaling means selected distributions or moments transform approximately over a finite range; it does not mean every zoomed segment is a copy.

An irregular boundary has no single observed length independent of ruler size. A long ruler skips bays and headlands; a short ruler follows more detail and reports a larger total.

Lewis Fry Richardson collected such scale-dependent measurements, and Mandelbrot connected them to fractal dimension. The lesson extends beyond geography: the measurement protocol is part of the result. Asking “How long?” without asking “At what resolution?” is incomplete.

Markets raise the analogous question: how does price variation change from minute to hour to day? Software metrics change between request, minute, and month. A scaling law is valuable when it predicts this transformation and exposes where it stops.

Cover a curve with rulers of length ε. If N(ε) rulers are needed, measured length is approximately L(ε) = N(ε)ε.

For a smooth curve, halving ε roughly doubles N, so total length stabilizes. For a fractal curve, count may scale as N(ε) ∝ ε^-D, where D is dimension. Then L(ε) ∝ ε^(1-D). When D > 1, measured length grows as the ruler shrinks.

Taking logarithms gives log N(ε) ≈ -D log ε + constant. Estimate D from the negative slope of a log-count versus log-ruler plot over a justified range. A straight-looking segment with three points is weak evidence; report uncertainty and endpoints.

Time series use different units on horizontal and vertical axes, so self-affinity matters more than geometric self-similarity. If typical increments obey |X(t+h)-X(t)| ∝ h^H, time scales by a while amplitude scales by a^H.

Brownian motion has H = 0.5. Values above 0.5 indicate persistence in the modeled process; values below 0.5 indicate anti-persistence. Interpretation depends on what series is measured and whether trends or breaks have been removed.

Scale estimates become clearer with exact counts. Suppose rulers of length 1, 1/2, 1/4, and 1/8 require 8, 20, 49, and 120 pieces.

Using the first and last points:

D ≈ log(120/8) / log(8/1) = log(15) / log(8) ≈ 1.30

Measured lengths are 8, 10, 12.25, and 15 units. The reported length grows as resolution becomes finer, consistent with a curve rougher than a smooth line but not filling a plane.

This estimate is illustrative. A real analysis fits all points, checks residuals, compares nearby scale ranges, and accounts for finite measurement. Dimension 1.30 does not reveal the boundary’s causal history.

Visual roughness can be confused with the smooth trend underneath it. The 2D lab changes roughness while preserving a labeled smooth reference.

Use the deterministic teaching model as follows:

1. Press Reset and note the Roughness percentage.
2. Identify the Smooth trend and Fractal trace from labels and distinct encodings.
3. Increase Roughness one step and compare the added fine-scale deviations across the time axis.
4. Record the Roughness gap at low, default, and high settings.
5. Reset and confirm that the default curves and metric return exactly.

The output is a teaching simulation, not a forecast or proof that a real series is fractal. Read the labeled curves and Roughness gap rather than selecting the most convincing-looking segment.

Roughness can vary across position and scale, which a single line suppresses. The 3D surface shows how fine structure sits inside coarse terrain while keeping the axes and concentration readouts explicit.

Operate the surface methodically:

1. Read the legend: x is position, y is scale, and z is roughness.
2. Move the Roughness slider and compare the reported surface range.
3. Select a position with the x focus slider and a scale with the y focus slider; record selected x, y, and z.
4. Hold one focus value fixed while moving the other to compare coarse and fine structure.
5. Orbit by pointer or keyboard, then Reset the parameter, focus, and camera.

The mesh is a finite generated terrain. Interpolation and camera perspective can make patterns appear smoother or steeper, so numeric readouts and axis labels carry the interpretation.

Power laws are easy to see in short log–log plots and hard to establish. Trends, seasonality, structural breaks, and mixtures can imitate scaling.

The estimated exponent depends on horizon range, estimator, overlap, and finite sample. Very short financial horizons contain market microstructure; very long horizons contain few observations and regime change. A scaling range should be chosen for substantive and statistical reasons, then tested elsewhere.

One exponent also compresses heterogeneity. If small and large fluctuations scale differently, a monofractal H is incomplete and a multifractal spectrum may help. Extra flexibility increases estimation risk, so validation must use withheld moments or periods.

Software architecture can look simple at repository level and rough at dependency-call level. Metrics must state whether they observe functions, services, deployments, incidents, or months.

A service graph’s edge count grows when dynamic calls, retries, and shared infrastructure are measured. A coarse diagram can therefore understate operational surface area. Conversely, counting every trace edge can obscure stable module boundaries.

Use multiple resolutions deliberately: code ownership for change planning, service dependencies for blast radius, and trace spans for incident diagnosis. Look for scale transitions where one abstraction stops predicting effort or failure.

LLM quality changes with token, answer, conversation, task family, and deployment scale. A single average can mix these resolutions into an uninterpretable number.

Token loss may improve smoothly with compute while rare long-context tasks remain rough. An answer-level judge can miss a conversation-level contradiction. Aggregating users can hide a subgroup tail.

Report metrics at aligned scales and test how error moments change with context length or interaction count. Scaling curves describe measured ranges; they do not guarantee indefinite improvement beyond available data.

Agent tasks often split recursively into plans, subtasks, tool calls, retries, and validations. The repeated branching resembles a scale hierarchy, but resemblance alone does not establish a fractal law.

Measure node count, depth, cost, and failure concentration as task resolution changes. A coarse “one task” metric can hide a hundred tool actions; a fine metric can miss that several actions share one mistaken premise.

Use scale-aware budgets: total run cap, branch cap, per-tool limit, and checkpoint frequency. Determine where cost grows faster than useful evidence and stop recursion before it becomes an uncontrolled workload surface.

Startup growth can appear smooth monthly while daily acquisition is dominated by campaigns and launches. Decisions fail when a scale convenient for reporting is mistaken for the scale of operations.

Measure signups daily for capacity, cohorts monthly for retention, and runway across financing horizons. A power-law claim about customer size should specify threshold and range; one large customer is not enough.

Scale analysis can reveal concentration and transition points. It cannot predict which channel or customer creates the next burst, so commitments should remain robust to range uncertainty.

Personal progress depends on observation scale. Daily noise can hide a monthly trend, while monthly totals can hide exhausting sequences.

Exercise measured by one session, weekly consistency, and yearly health tells different stories. Calendar busyness measured by event count misses preparation and recovery. No one ruler is universally correct.

Choose the ruler that matches the decision, compare at least one finer and coarser view, and avoid moralizing noise. Scale awareness supports kinder, more accurate planning rather than constant self-surveillance.

A scaling statement is weak when it omits the object, ruler, exponent, and range. This exercise produces a claim another analyst can reproduce and reject.

Choose a curve or ordered series with enough resolution. Define at least five ruler sizes or horizons before measuring. For a boundary, count covering segments; for a series, compute a typical increment magnitude or an unrooted absolute moment. Plot both axes logarithmically and fit only the range justified by measurement and mechanism.

Use the 2D lab to compare the Smooth trend and Fractal trace as the Roughness slider changes, recording the Roughness gap. Estimate slope and residual patterns from your chosen external series rather than from the widget. On the 3D surface, compare coarse and fine structure with the position and scale focus sliders and selected x, y, and z values rather than camera impression.

Now challenge the claim. Shift the fitting endpoints, use a later window, remove a trend, and compare overlapping with non-overlapping increments. Record how much the exponent moves. If small choices produce large movement, report instability rather than averaging it away.

Use a sensitivity table:

Choice	Baseline	Alternative	Estimated exponent change
Scale range	stated endpoints	narrower interior	measured
Detrending	none or stated method	alternate method	measured
Window	calibration period	later period	measured
Increment construction	non-overlapping	overlapping	measured
Moment order	low order	higher order	measured

Explain each change mechanistically. Overlapping increments add dependence; high moments give extremes more influence; long horizons reduce block count; short horizons may contain measurement artifacts. An exponent without these choices is not portable evidence.

Then compare the exponent’s usefulness against a simple baseline. Does it improve a held-out moment, capacity estimate, or roughness classification enough to change the decision? If not, keep the simpler description. If different moment orders require materially different exponents, record that as motivation for a multifractal analysis rather than forcing one averaged H.

Finish with a bounded sentence: “For this variable, estimator, and range, typical magnitude scales approximately with exponent H; outside that range the evidence is insufficient.” Add one decision consequence, such as which monitoring resolution or agent budget changes. This is more useful than labeling the whole system fractal.

Attach an uncertainty interval and a scale-coverage table to that sentence. For every ruler size, record the number of usable blocks and whether measurement resolution truncates the signal. The finest scales can be dominated by quantization or market microstructure, while the coarsest scales may contain too few blocks for stable estimates. A straight line through either region is visual convenience, not evidence.

Finally, compare the estimated relation with a shuffled or phase-randomized control that preserves some distributional features while destroying ordering. If the same slope survives, the claimed mechanism may come from marginal variation rather than temporal structure. If it disappears, repeat on a held-out window before interpreting persistence. The control cannot prove a fractal mechanism, but it can rule out an overly simple explanation and sharpen the next experiment.

Fractal roughness is not a visual style; it is a measured relationship between detail and scale. The exponent summarizes how observation changes when the ruler changes.

choose object + ruler
        |
measure across resolutions
        |
fit exponent over finite range
        |
test alternate ranges and windows
        |
state what the exponent does not explain

This discipline prepares the later market argument. Before claiming universal scaling, preserve units, distinguish self-similarity from self-affinity, and report the range where evidence supports the rule.

Fractal concepts are safest when traced to primary definitions and measurements. These sources establish coastline scaling, self-affine stochastic paths, and long-range dependence.

• Benoit Mandelbrot and Richard L. Hudson, The Misbehavior of Markets, Basic Books.
• Benoit Mandelbrot, “How Long Is the Coast of Britain? Statistical Self-Similarity and Fractional Dimension”, Science 156(3775), 1967.
• Benoit Mandelbrot and John W. Van Ness, “Fractional Brownian Motions, Fractional Noises and Applications”, SIAM Review 10(4), 1968.
• H. E. Hurst, “Long-Term Storage Capacity of Reservoirs”, Transactions of the American Society of Civil Engineers 116, 1951.

Roughness becomes scientific only when tied to scale, exponent, range, and uncertainty. Visual resemblance is an invitation to measure, not a conclusion.

• Shorter rulers reveal more detail on an irregular boundary.
• Fractal dimension describes how count grows as ruler size shrinks.
• Time series are usually self-affine because time and amplitude scale differently.
• Brownian motion provides the H = 0.5 reference.
• Finite samples, trends, and breaks can create spurious scaling.
• A single exponent may fail when fluctuation sizes scale differently.

The primer is complete when every roughness claim includes a ruler and a range. Apply these checks to one curve or time series.

• [ ] I can distinguish self-similarity from self-affinity.
• [ ] I can calculate a two-point fractal-dimension estimate.
• [ ] I can explain the Hurst exponent in plain language.
• [ ] I can operate both labs using numeric and non-color readouts.
• [ ] I can state the fitted scale range and sample limitations.
• [ ] I can name three mechanisms that imitate scaling.
• [ ] I can identify a scale transition in a non-financial system.
• [ ] I can explain why simulated fractal terrain is not a forecast.