How Science Works | Why Scientists Use Different Theories at Different Scales

Suppose the observable of interest changes slowly, with characteristic angular frequency ω. The system also contains a much faster internal mode with characteristic frequency Ω. If ω/Ω is small, the fast mode may follow the slow motion closely enough that its detailed independent motion need not be tracked for the intended prediction.

This is a separation of scales. Similar ratios can involve energy, momentum or length: a probe wavelength much longer than microscopic spacing, for example. The useful ratio depends on the physics and observable, not on the word “small” alone.

A low-energy description still contains information about the omitted physics. Its parameters may include measured elastic moduli, masses, interaction strengths or higher-order coefficients. Eliminating a variable does not mean declaring that variable nonexistent or setting its influence to zero.

The target precision is part of the question. A correction of one percent may be negligible for a rough estimate and decisive for a precision comparison. The same effective description can therefore be adequate for one task and inadequate for another at the same physical scale.

Not every approximation is a formal effective field theory. The broader effective-theory viewpoint includes many reduced descriptions; field theories add spatially varying fields and a systematic organisation of allowed interactions. Our mechanical example isolates the logic without requiring advanced quantum mathematics.

Contents · Next chapter

Let x and y be two displacement coordinates, measured in metres, with masses m and M. Use potential energy V = kx²/2 + Ky²/2 − gxy. The constants k, K and g have units of newtons per metre. The coupling term makes the coordinates influence each other.

Choose positive masses and kK > g², with k and K positive. These conditions make the quadratic potential stable. Apply an external force f(t) only to x. The equations are m ẍ + kx − gy = f and M ÿ + Ky − gx = 0.

For the original numerical example, use m = 1 kilogram, M = 0.01 kilogram, k = 100 newtons per metre, K = 100 newtons per metre and g = 50 newtons per metre. Since kK = 10,000 and g² = 2,500 in consistent squared-stiffness units, the stability condition is satisfied.

The fast coordinate’s uncoupled frequency is Ω = √(K/M) = 100 radians per second. The uncoupled x frequency is √(k/m) = 10 radians per second. Coupling shifts the actual normal-mode frequencies; these uncoupled values initially identify the intended hierarchy rather than the final exact spectrum.

The model is a mathematical exercise, not an apparatus-building instruction. It assumes linear motion and no damping. We will analyse a specified harmonic response away from resonances, without adding arbitrary free fast oscillations from independent initial conditions.

Contents · Previous chapter · Next chapter

For a harmonic component at angular frequency ω, a second time derivative becomes multiplication by −ω². The y equation becomes (K − Mω²)y = gx. Provided that denominator is nonzero, y = gx/(K − Mω²).

Substitute this into the x equation. The force and displacement satisfy f = [k − mω² − g²/(K − Mω²)]x. This is an exact frequency-domain reduction for the specified harmonic component. It retains the fast coordinate’s influence through a frequency-dependent term.

Define S(ω) = g²/(K − Mω²). The effective dynamic coefficient multiplying x is k − mω² − S(ω). Although y has disappeared from the equation, it has changed the response of x.

At zero frequency, y = (g/K)x. The coupling lowers the static stiffness from k to k_eff = k − g²/K. In our example this is 100 − 25 = 75 newtons per metre. Simply deleting y and its coupling would incorrectly leave a stiffness of 100.

This is the first central lesson: unresolved motion can renormalise a parameter already present in the simpler model. The low-frequency theory need not retain a separate y coordinate to describe its leading effect, but it must use the correctly matched coefficient.

Contents · Previous chapter · Next chapter

Write r = ω/Ω. Since Ω² = K/M, the eliminated contribution becomes S(ω) = (g²/K)/(1 − r²). For |r| < 1, the geometric series gives 1/(1 − r²) = 1 + r² + r⁴ + further even powers.

Keeping only the constant term gives S₀ = g²/K. Keeping the first frequency correction gives S₁ = (g²/K)(1 + r²). Adding the next correction gives S₂ = (g²/K)(1 + r² + r⁴).

Through order ω², the reduced x equation has dynamic coefficient k_eff − m_eff ω², where m_eff = m + g²M/K². For our parameters, m_eff = 1.0025 kilograms. The omitted fast coordinate contributes both a static-stiffness shift and a small effective-inertia correction.

The next term is proportional to ω⁴. In a time-domain effective equation it corresponds to a higher-derivative correction, interpreted perturbatively within the low-frequency regime. Treating a truncated higher-derivative equation as an exact theory at arbitrary frequency can introduce spurious solutions outside its intended domain.

The expansion is useful because the powers have an ordering. At sufficiently small r, r⁴ is much smaller than r². In this special soluble model the coefficients and exact remainder are known. In more complicated effective theories, power counting supplies an organisation, while coefficient sizes and errors require additional analysis.

Contents · Previous chapter · Next chapter

Our example has g²/K = 25 newtons per metre. At r = 0.1, the exact eliminated contribution is 25/(1 − 0.01) ≈ 25.252525. The constant approximation gives 25, and including r² gives 25.25. The relative errors in S are respectively 1 percent and 0.01 percent.

At r = 0.3, the exact S is approximately 27.472527. The first three truncations give 25, 27.25 and 27.4525. Their relative errors are 9 percent, 0.81 percent and 0.0729 percent. Adding a term provides a predictable improvement in this regime.

At r = 0.6, exact S = 39.0625. The same approximations give 25, 34 and 37.24, with relative errors of 36 percent, 12.96 percent and 4.6656 percent. Three terms are no longer enough for a one-percent description of this contribution.

At r = 0.9, exact S is approximately 131.578947. The approximations 25, 45.25 and 61.6525 remain far away, with relative errors of 81 percent, 65.61 percent and 53.1441 percent. The expansion technically converges for this r, but it converges too slowly to justify a short low-order description.

All these percentages concern the eliminated contribution S, not the entire displacement response. If k − mω² − S is close to zero near a coupled-system resonance, even a small error in S can cause a large relative error in x/f. An error budget must be propagated to the actual observable.

Contents · Previous chapter · Next chapter

The geometric series lets us calculate the remainder exactly. If S_n retains terms through r^(2n), its difference from the exact S is (g²/K)r^(2n+2)/(1 − r²). Relative to exact S, the error is simply r^(2n+2) for 0 ≤ r < 1.

For a one-percent relative-error target on S, the constant approximation requires r² ≤ 0.01, so r ≤ 0.1. Including the r² term requires r⁴ ≤ 0.01, so r ≤ approximately 0.31623. Including through r⁴ permits r ≤ approximately 0.46416.

These are model-specific accuracy limits, not universal rules that every effective theory works up to the same fraction of its breakdown scale. Here the exact function provides unusually strong information. Other expansions may be asymptotic, contain logarithms or have enhanced coefficients.

The formal breakdown scale Ω marks the internal frequency where the eliminated response has a pole in this representation. Near it, the low-order expansion is unsuitable. Restoring y or using the full coupled description is more informative than forcing the truncated series beyond its purpose.

A useful graph would place exact S and successive approximations against r from zero towards one. A second panel would show their relative errors and a horizontal one-percent target. It would make both the improvement from additional terms and the shrinking usefulness near the high-frequency scale visible.

Contents · Previous chapter · Next chapter

If the detailed two-coordinate model were unknown, low-frequency measurements could determine k_eff and m_eff. This is matching: choosing the effective parameters so that the reduced description reproduces relevant observations or a more detailed calculation at the working scale.

Matching is not complete validation. If two coefficients are fitted using two pieces of information, agreement with those same inputs provides limited independent evidence. Predictions at additional frequencies, with uncertainty propagated, can test whether the truncated description captures the response shape.

Different microscopic parameter combinations can share the same leading effective coefficients. Measuring only static stiffness and low-frequency inertia would therefore not uniquely reveal m, M, k, K and g. Higher-order terms or observations nearer the internal mode could distinguish descriptions that look identical at low resolution.

For a concrete static degeneracy, keep k = 100 but replace K = 100 and g = 50 with K = 25 and g = 25, in the same stiffness units. Both choices give g²/K = 25 and therefore k_eff = 75. With M still 0.01 kilogram, however, the internal scale becomes 50 rather than 100 radians per second. Static agreement has not fixed the dynamic scale.

This underdetermination is not a defect if the goal is a low-frequency prediction. It becomes a problem when someone claims the effective fit has uniquely discovered every microscopic mechanism. A theory’s successful domain and its ability to identify deeper structure are different achievements.

When fitting data, keep parameter uncertainty separate from truncation uncertainty. The first concerns imperfect knowledge of coefficients retained in the model. The second concerns terms deliberately omitted. A precise coefficient fit can coexist with a poor prediction if the omitted contribution is too large.

Contents · Previous chapter · Next chapter

A systematic effective theory retains the low-scale degrees of freedom and writes interactions allowed by the relevant symmetries, ordered by an expansion scheme. It does not arbitrarily keep only the terms that make an attractive fit. Symmetry can forbid some terms and relate others.

In our simple undamped time-reversal-symmetric harmonic example, the response expansion contains even powers of frequency. A dissipative model can introduce other structures, including frequency-dependent imaginary parts. Copying our even real series into a system with friction would omit relevant physics.

Appelquist and Carazzone’s 1975 original study examined low-momentum behaviour and formulated a decoupling theorem under its field-theoretic assumptions. It helped establish how heavy degrees of freedom can influence low-energy physics through adjusted parameters and suppressed contributions. The result is conditional, not a claim that all heavy physics always disappears in every theory.

Weinberg’s 1979 original paper on phenomenological Lagrangians developed the role of symmetry-consistent interactions in organising low-energy descriptions. Burgess’s author-written introduction provides a broader account of scale hierarchies and power counting. These field-theory ideas extend the logic illustrated by the oscillator; the mechanical calculation is not a derivation of every quantum-field result.

The point is disciplined compression. Many possible microscopic details can be represented by a manageable number of coefficients at a chosen accuracy. As the required precision increases, more terms may become relevant and the experiment must supply enough information to constrain them.

Contents · Previous chapter · Next chapter

The original 2015 study on quantifying EFT truncation errors adapted Bayesian procedures to uncertainty from omitted orders. Such methods make assumptions about coefficient behaviour and use order-by-order information. Their uncertainty statements are therefore conditional model-based assessments, not automatic rigorous bounds like the exact geometric remainder in our example.

A 2025 original analysis by Bub and colleagues calibrated nucleon–nucleon interactions in pionless effective field theory while including correlated truncation uncertainty. Its results exposed inconsistencies across orders for the particular power-counting construction studied. That is a useful example of uncertainty analysis challenging an approximation scheme rather than merely decorating a prediction with error bars.

The accessible author manuscript describes how low-energy constants and truncation-error parameters were inferred together. This does not mean every nuclear effective theory fails, or that Bayesian fitting can repair any missing mechanism. The conclusion concerns the studied formulations, assumptions and data range.

For any effective calculation, inspect whether successive orders improve agreement as expected and whether their differences fit the proposed error model. Check correlations between predictions: omitted terms can shift several nearby energies together, so treating every discrepancy as independent noise may exaggerate information.

Experimental uncertainty, numerical error and effective-theory truncation also need separate accounting. Reducing simulation roundoff cannot fix an omitted physical mode. Collecting more data can sharpen a disagreement without extending the approximation’s valid domain. A growing discrepancy may indicate the need for another term, different power counting or different degrees of freedom.

Contents · Previous chapter · Next chapter

For the worksheet, calculate S at r = 0.1 and 0.6 using the exact expression and the first three approximations. State the error relative to S itself. Then explain why that number alone cannot guarantee an equally accurate displacement prediction near a resonance.

Next choose a one-percent target and derive the permitted r range for each truncation. This makes the choice of theory depend on both physical scale and required accuracy. A leading approximation can be entirely adequate in one range and predictably inadequate in another.

Finally, imagine that only k_eff = 75 newtons per metre has been measured. Explain why this does not uniquely determine the five microscopic parameters. Identify what new type of frequency-dependent observation could constrain the omitted mode, rather than pretending more decimal places in the static measurement reveal everything.

A useful report states the observable, relevant scale ratio, retained variables, symmetry assumptions, matching information, truncation order and estimated error. It also identifies what observation would signal breakdown. Calling a model “effective” without those details provides little protection against overextension.

An effective theory can be reliable without being the final microscopic description. Conversely, a more microscopic model can be harder to calculate and less useful for the immediate task. Scientific judgement lies in choosing the simplest description that preserves the required physics and honestly controls the neglected effects.

Return to Physics and the How Science Works roadmap. Different theories at different scales are not a collection of excuses. When their matching and error limits are explicit, they form a connected system of testable approximations.

Contents · Previous chapter · Continue to Physics