Resolution

You have regularized your inverse problem, run the solver, and out comes a reconstruction — a smooth, plausible-looking image of whatever you were trying to recover. It looks convincing. But an honest question remains, and it is the one that separates careful science from wishful thinking: how much of that fine detail can you actually trust? Is the little bump on the left a real feature your data revealed, or a smear that regularization painted in?

The model resolution matrix answers exactly this. It quantifies how much your recovered solution is a blurred version of the true one — it tells you the genuine spatial resolution your data can support, so you know which features to believe and which to treat as artefacts.

The idea: recovered = blurred truth

Regularization buys stability — but at a price. The recovered model is not the true one; it is a blurred version of it, and the resolution matrix measures exactly how blurred. Write the regularized solution as \hat m = G^{\sharp} d for some inverse operator G^{\sharp} (the pseudoinverse, or a Tikhonov-filtered version). Substitute the noise-free data d = G\, m_{\text{true}}:

\hat m = G^{\sharp} G\, m_{\text{true}} = R\, m_{\text{true}}, \qquad R = G^{\sharp} G.

So R maps the truth to what you actually recover. If R = I, resolution is perfect — every parameter comes back exactly. In practice R \neq I: it is a blurred, smeared version of the identity, with a bright band along the diagonal that has some width. That width is the whole story.

Read it row by row: the resolution kernel

Look at a single row i of R. It tells you how the recovered value \hat m_i is built from the true model:

\hat m_i = \sum_j R_{ij}\, (m_{\text{true}})_j.

Instead of returning the single true value (m_{\text{true}})_i, your estimate is a weighted average of nearby true values — the weights are that row of R. This row is the resolution kernel (or averaging kernel) for parameter i: it shows how one true parameter gets spread across your estimate. A sharp, spiky row (close to a row of the identity) means good local resolution — you really are seeing that one point. A broad, smeared row means you can only see a blurred regional average, and the width of that kernel is your resolution length: the smallest separation at which two features can still be told apart.

Worked example 1 — a near-identity R (good resolution)

Here is a heatmap of a resolution matrix that is close to the identity. The bright band hugs the diagonal tightly: each row is a sharp spike sitting almost entirely on its own parameter. Reading row 5, say, the weight on (m_{\text{true}})_5 is large and its neighbours barely contribute — so \hat m_5 \approx (m_{\text{true}})_5. Two features one grid-cell apart would still show up as two separate bumps. This is what "the data resolves the fine structure" looks like.

Worked example 2 — a heavily smeared R (poor resolution)

Now the same problem with much heavier regularization. The diagonal has bled into a wide, fuzzy band: every row is a broad hump spanning several parameters. Reading row 5 again, the weight is shared roughly equally across cells 3–7 — so \hat m_5 is really a blob average of five neighbours, not the value at 5. Two features three cells apart now merge into one smear; you literally cannot tell them apart. Nearby parameters have become indistinguishable.

Worked example 3 — the resolution-vs-stability trade-off

Compare the two heatmaps and the moral is stark. Going from example 1 to example 2, all we did was turn up the regularization. That helped stability — it tamed the noise amplification that makes raw inversion explode. But it hurt resolution — it broadened every kernel and smeared R further from the identity. This is the fundamental tension:

\text{more regularization} \;\Longrightarrow\; \text{more stable, but lower resolution.}

It is the bias–variance trade-off wearing a different hat. Heavy regularization broadens the kernels (more bias, more blurring) but suppresses noise (less variance); light regularization sharpens resolution but lets the noise flood back in. There is no free lunch — only a choice of where to sit on the curve, and the resolution matrix is how you read off what that choice cost you.

Where the blur comes from

The resolution matrix is not a mysterious extra ingredient — it falls straight out of how you built the inverse. Through the SVD G = U \Sigma V^{\mathsf T}, a filtered pseudoinverse gives

R = G^{\sharp} G = V\, \mathrm{diag}(f_i)\, V^{\mathsf T},

where the filter factors f_i \in [0,1] say how fully each singular direction is retained. Keep every direction (f_i = 1) and R = VV^{\mathsf T} = I — perfect resolution, but also full noise amplification through the tiny singular values. Regularization pushes the f_i for small singular values toward 0: those directions are dropped, the noise is tamed, and precisely those dropped directions are the fine spatial detail you can no longer resolve. The blur in R is the shadow of the singular directions you chose to sacrifice.

The word is borrowed straight from optics, and the analogy is exact. A telescope cannot separate two stars closer together than its diffraction limit — they blur into a single smudge. Its resolving power is the smallest angle it can split. An inverse problem is a lens too: the rows of R are its point-spread functions, and the kernel width is its resolving power. "Two stars you cannot split" and "two parameters you cannot separate" are the same sentence in two fields — which is why the same word travelled between them.

This is the classic way people fool themselves with inverse problems. A reconstruction can look crisp and detailed — smooth curves, little bumps, plausible structure — and yet have genuinely poor resolution. The apparent sharpness comes from the display and the regularizer, not from information in the data.

The resolution matrix is the referee. Any feature finer than the resolution length is an artefact or a smeared average, not real recovered structure. Claiming to "see" a small feature that your kernels are too broad to resolve is over-interpretation, plain and simple — the data physically cannot support it. Before you point at a bump and tell a story about it, check that the row of R there is actually spiky. If it is a broad hump, your "feature" is just blurring, and the honest statement is a regional average, not a point.

When a team publishes a tomographic image of the Earth's mantle — say a warm plume rising under Hawaii, or a cold slab sinking beneath the Andes — the very first question a reviewer asks is: can your data actually resolve that? Rays criss-cross the deep Earth unevenly, so some regions are beautifully sampled and others are barely touched, and regularization quietly fills the gaps with smooth guesswork.

That is why resolution analysis is non-negotiable in the field. Seismologists carefully report the resolution of their images and run resolution tests (the famous "checkerboard test" puts in a known pattern of blobs and sees how badly the inversion smears it back out). A claimed discovery of some small deep-Earth structure must pass such a test. The resolution matrix is exactly the tool that separates "the data really shows this" from "regularization smeared some blob into looking like this" — a crucial guard against fooling yourself, and everyone else.