Linearized sensitivities and regularized resolution
Research context
research method · conditional · As of 2026-10-10
Linearizing d=g(m) about m0 gives a sensitivity matrix J.
Key knowledge
Linearizing d=g(m) about m0 gives a sensitivity matrix J. Weighted regularization selects a model using data covariance and prior penalties; the resolution operator indicates which parameter combinations the observations recover, rather than a universal spatial resolution.
Values and equations
d-g(m0) ≈ J δm; δm_hat = (J^T C_d^-1 J + C_m^-1)^-1 J^T C_d^-1 (d-g(m0)); R = (J^T C_d^-1 J + C_m^-1)^-1 J^T C_d^-1 J
- m0
- Linearization point and Gaussian prior mean
- δm
- Parameter perturbation with zero-mean Gaussian prior
- J
- Jacobian of the forward model
- C_d
- Positive definite observation covariance
- C_m
- Positive definite Gaussian prior covariance
- R
- Local model-resolution matrix
Conditions and scope
- Local linearization can fail for multimodal or strongly nonlinear inversions.
- Compare residuals, withheld data and regularization sensitivity; seismic paths and gravity/magnetic kernels have different null spaces.
- Finite-dimensional zero-mean Gaussian observation noise and perturbation prior δm~N(0,C_m), with m0 the prior mean and invertible stated precision. A different prior mean requires an offset term; the displayed formula follows locally by completing the square.
Sources · 1
References
Sources retain their recorded scope. A source link alone does not establish verification of every statement.
Evidence and review
The Bayesian Approach to Inverse Problems: Sections 1.1–1.3 and 4: likelihood, regularization, posterior approximation (selected source body inspected).
2026-10-10 · Bounded source check of the recorded statement and conditions. This does not certify a unique physical model, reproduce the cited study or replace independent branch review.
Entry details
Permanent link to this entry ↗- Record identity
- research-earth-linearized-resolution
- Revision
- research.1
- Structured reference record
- Postgraduate & research reference