Scales, Hierarchy

For a given real system, a single “one true” model rarely exists. Several plausible models sit on the table at once — some coarser, some finer — and the modeler’s job is to pick the level of detail that matches the question being asked. The candidates form a hierarchy, stacked from cheap-and-rough at the bottom to expensive-and-accurate at the top, and the choice of where to land in that stack is the scale decision.

Two questions govern the choice within the hierarchy.

• What is necessary? Which scale does the desired outcome actually require? Some questions can be answered from a rough overall picture; others need the fine detail that a coarse model would smooth over. The question fixes a floor below which the model is no longer useful.
• What is feasible? What can the available resources — data, compute, implementation effort, calendar time — actually support? They fix a ceiling above which the model cannot be built or solved in practice.

A model worth working with sits in the gap between the floor and the ceiling. When the floor exceeds the ceiling — when the question demands more than the resources allow — the only honest response is to revise one of them: relax the question or expand the budget.

Where scale shows up in practice

The “scale” axis takes different shapes across application areas. Four representative examples sketch the range.

• Flow through a cylinder. A fluid is pumped into one end of a pipe at known inflow conditions, and the velocity inside is wanted everywhere along the way. The pipe is 3D, but the model doesn’t have to be. A 1D model tracks the flow along the pipe’s length only, treating each cross-section as uniform — fine when the flow really is uniform across the pipe. A 2D model also resolves how the velocity changes from the center out to the wall — needed once friction at the wall starts slowing things noticeably. A 3D model resolves every direction — needed when the flow swirls, or the pipe has bends or obstructions inside. Each step up captures more detail at higher cost.

• Population dynamics, USA 1840–1860. During American westward expansion, the U.S. population was both growing and moving west. A purely time-dependent model $p(t)$ tracks the country’s total population as a single number — suitable for how big is the country? questions. A space-and-time model $p(x, t)$, with $x$ running east-to-west, tracks the population density as it shifts west and grows — suitable for where are people now, and where will they be next decade? The first is an ODE, the second a PDE; the underlying system is the same.

• Simulation of circuits. Classical circuit simulation treats a board as a network of ideal components — resistors, capacitors, inductors, joined by ideal wires — governed by Kirchhoff’s laws (currents balance at every junction; voltages around every loop sum to zero). That gives a purely time-dependent ODE for all the currents and voltages on the board. As components shrink to micrometer and nanometer sizes, the wires themselves stop behaving ideally: each picks up its own small resistance, capacitance to its neighbors, and inductance — the parasitic effects the original picture ignored. Capturing them pushes the model from time-only equations to ones that have to track where on the board things happen, not just when. Same physics, two scales — with the threshold between them shifting over the decades as the silicon shrank.

Each example carries the same shape: one system, several legitimate descriptions at different scales, with the question being asked deciding which scale is necessary and the budget deciding which is feasible.

Multiscale problems

Each of those examples lets the modeler commit to a single scale and work at that scale alone — a 1D pipe model, a time-only population model, a Kirchhoff circuit network — with everything finer than the chosen scale simply dropped from the picture. A second class of problems doesn’t allow this. The fine scales are coupled to the coarse ones — what the small structure does shows up in the large-scale answer — and dropping them produces an unacceptable loss of accuracy.

The textbook case is fluid turbulence.

The standard fix is to summarize the small swirls instead of tracking them one by one. A turbulence model does exactly that: it captures the average effect of the small swirls on the larger flow and bakes that summary into the coarse simulation as extra correction terms. The tiny swirls themselves are never computed — only their summary effect appears, calibrated either from theory or from measurements.

Two general techniques sit underneath this idea, and both apply far beyond turbulence.

• Averaging in space and time. Replace the wildly-fluctuating fine-scale picture with its local average — a smoother, blurrier version — and derive equations for that smoothed version. What’s left over after smoothing — the wiggles around the average — gets folded back in as a small correction term.
• Homogenization. When the fine-scale structure has a regular character (a repeating pattern, or a random porous material that’s uniform overall), solve a small problem on the fine scale once, extract a few summary numbers from it — an overall permeability (how readily fluid flows through), an effective conductivity, an average reaction rate — and plug those into the coarse model as fixed parameters.

Both trade exact fine-scale detail for a coarse model that can actually be computed. The trade works because the fine-scale picture has enough regular structure that capturing its bulk effect in a handful of numbers is a faithful enough summary.

In practice, this leads to a workflow that walks up the hierarchy rather than committing to one rung. A simulation usually starts coarse — cheap, fast, and good enough for a first answer. If that answer doesn’t really settle the question, the model is refined: a denser grid, a smaller time step, or swapping a summarized fine-scale piece for one that’s simulated directly. Then it runs again. Stepwise refinement through the hierarchy, rather than betting everything on one scale, is how most realistic simulations actually get built.

Hierarchy centered around humans

A concrete way to see a model hierarchy in action is to fix a domain — say everything that touches the human body — and walk down through the scales it covers. The same overall subject appears at every level, but the question being asked changes, and at each level the natural mathematical machinery is different.

The third column says possible model deliberately: none of these is the only valid choice for its row, just one plausible candidate. A different team with different questions or different budgets might land on a different model at the same level. The point of the table is the vertical structure — that one biological subject hosts a stack of qualitatively distinct mathematical descriptions — not the specific row-by-row picks.

Reading top to bottom, two patterns stand out.

• Each row sits at a smaller spatial scale than the row above, and the total range is enormous — from the geographic extent of a country (millions of meters) at the top to the size of an electron (less than a billionth of a meter) at the bottom. One subject area, biology, spans the whole gap.
• The class of model changes as you go down the table, sometimes only a little and sometimes completely. Population dynamics covers both the global and local rows; continuum mechanics covers both blood flow in the heart and cellular transport; the rest jump into entirely different kinds of model. The trajectory runs from time-only equations at the top (population dynamics, organ-level system simulators), through network models for circulation (heart as pump, vessels as canals), to equations defined across continuous space (continuum mechanics), to individual particles moving under Newton’s laws (molecular dynamics), and finally into quantum mechanics for atoms and electrons.

Which row a given study sits on is a modeling commitment, governed by exactly the necessary-vs-feasible trade-off introduced at the start of the section. The fact that there are eight rows on a single subject is the point of the table: real modeling lives across a hierarchy of scales, not at any one of them.