Sensitivity Fields: A Unified Perspective

• Sensitivity fields are mathematical constructs that capture how infinitesimal perturbations propagate into measurable variations, expressed as gradients or response functions.
• They are applied in quantitative phase imaging, neural radiance fields, sensor metrology, and inverse scattering to enhance parameter estimation and optimize system responses.
• Practical methods include scattering relationships, Sobol indices, and variational techniques that assess, control, or nullify sensitivities in complex multi-domain problems.

Sensitivity fields are not a single standardized mathematical object across the research literature. The term and closely related constructions are used for the gradient of a scattered field with respect to object parameters in quantitative phase imaging (Barbastathis, 2024), for a scalar field per simulation parameter over a spatial domain (Evers et al., 2024), and for a learnable field that quantifies the spatially varying impact of geometric perturbations on rendering quality in Neural Radiance Fields (Kim et al., 22 Oct 2025). In adjacent literatures, the same conceptual role is played by figures of merit and response functions that quantify how strongly a device, field observable, or physical system responds to external perturbations, including effective area in SQUID multiplexers (Stiehl et al., 2010), sensitivity curves in space-based interferometers (Yu et al., 2023), and minimum detectable electric fields in active biological membranes (Mathew et al., 2024). The unifying theme is resolved response: a sensitivity field, in the broadest technical sense, encodes how infinitesimal or weak perturbations propagate into measurable variation.

1. Terminological scope and principal definitions

The literature uses “sensitivity field” and closely related expressions in several non-equivalent but structurally related ways. Some are explicit spatial fields; others are frequency-domain or parameter-domain response objects.

Domain	Quantity	Definition or role
Quantitative phase imaging	Sensitivity Field	$$\chi(\mathbf{r};\boldsymbol{\theta}) \equiv \dfrac{\partial \psi(\mathbf{r};\boldsymbol{\theta})}{\partial \boldsymbol{\theta}}$$ (Barbastathis, 2024)
Spatial simulation analysis	Spatial sensitivity field	A scalar-valued function $S_i(\mathbf{x})$ per simulation parameter over the domain (Evers et al., 2024)
NeRF protection	Sensitivity Field	$G_s:\mathbf{x}\mapsto s\in[0,1]$ (Kim et al., 22 Oct 2025)
SQUID multiplexers	Effective area	$A_{eff}=\Phi/B$ (Stiehl et al., 2010)
Space-based interferometers	Sensitivity curve	$S_O(f)=N_O(f)/R_O(f)$ (Yu et al., 2023)
Biological membranes	Sensitivity	Minimum detectable electric field or threshold electrical noise (Mathew et al., 2024)

These definitions differ in codomain, dimensionality, and operational meaning. In imaging and simulation, the quantity is explicitly spatial. In detector physics, it is often a transfer-function-derived figure of merit. In learning systems, it can be a trainable latent field constrained by downstream objectives. This suggests that “sensitivity field” is best treated as a family resemblance term rather than a uniquely defined object.

2. Inverse scattering and parameter estimation

In quantitative phase imaging, the sensitivity field is part of a parameter-estimation formulation in which the measured intensity is modeled as

$$g_m(\boldsymbol{\theta}) = |\psi(\mathbf{r}_m; \boldsymbol{\theta})|^2,$$

and the optimization target is

$$E(\boldsymbol{\theta}) = \frac{1}{4M} \sum_{m=1}^{M} \left[ \tilde{g}_m - |\psi(\mathbf{r}_m; \boldsymbol{\theta})|^2 \right]^2.$$

The parameter estimate is obtained by minimizing this loss, and the loss gradient depends explicitly on the Sensitivity Field through

$$\frac{\partial E}{\partial \boldsymbol{\theta}} = \frac{1}{M} \sum_{m=1}^M \Re \left\{ \psi^*(\mathbf{r}_m;\boldsymbol{\theta})\,\chi(\mathbf{r}_m;\boldsymbol{\theta}) \right\} \, \Delta_m(\boldsymbol{\theta}),$$

with $$\Delta_m(\boldsymbol{\theta}) = \tilde{g}_m - |\psi(\mathbf{r}_m;\boldsymbol{\theta})|^2$$ (Barbastathis, 2024).

A central result is that the Sensitivity Field itself obeys a scattering relationship analogous to that of the original scattered field. In the thin-film approximation, the detector-plane sensitivity is generated by a modified illumination term. In the first Born approximation, it takes the convolutional form

$$\chi(\mathbf{r}; \boldsymbol{\theta}) = \int \psi_0(\mathbf{r'})\,\chi_v(\mathbf{r'}; \boldsymbol{\theta})\,h_G(\mathbf{r}-\mathbf{r}')\,d\mathbf{r}',$$

with

$S_i(\mathbf{x})$0

Under multiple scattering, the sensitivity field satisfies a Lippmann–Schwinger-type integral equation,

$S_i(\mathbf{x})$1

The paper’s stated interpretation is that the Sensitivity Fields are either produced by the same scattering potential but with a modified illumination field based on the parameter gradient, or by the same illumination but with a potential corresponding to the gradient of the original potential (Barbastathis, 2024).

This formulation makes sensitivity a first-class forward quantity rather than only a derivative computed after the fact. A plausible implication is that the conditioning of an inverse problem can be read directly from the structure of $S_i(\mathbf{x})$2, particularly when transitioning from thin-film or Born regimes to multiple-scattering regimes.

3. Spatial ensemble analysis and verification

In 3D simulation ensembles, spatial sensitivity fields are defined as scalar fields over the simulation domain, one per simulation parameter, that express how sensitive the simulation output is to each parameter at every spatial location (Evers et al., 2024). For parameter $S_i(\mathbf{x})$3, the field is written as $S_i(\mathbf{x})$4, and the result of analyzing all parameters is multi-field data: $S_i(\mathbf{x})$5 scalar fields over the same domain.

The paper supports several sensitivity-analysis methods computed individually at each spatial location. For Sobol indices,

$S_i(\mathbf{x})$6

where $S_i(\mathbf{x})$7 is the variance due to parameter $S_i(\mathbf{x})$8 and $S_i(\mathbf{x})$9 is the total variance. The paper also treats the Borgonovo $G_s:\mathbf{x}\mapsto s\in[0,1]$0 measure and Distance-based Generalized Sensitivity Analysis (DGSA), both evaluated voxelwise to produce scalar sensitivity volumes (Evers et al., 2024).

A major contribution is methodological rather than purely statistical: the interactive analysis of the resulting multi-field data. To avoid 3D occlusion, the approach linearizes the spatial domain using a data-driven space-filling curve. The central visualization combines Horizon Graphs and a line chart, alongside a Parallel Coordinates Plot, 3D surface rendering of selections, and a linked parameter dependency visualization. The system is validated on synthetic and real-world ensemble data, including blood flow and radiofrequency ablation cases (Evers et al., 2024).

A related but distinct line of work studies sensitivity in spatial verification rather than constructing sensitivity fields per se. In the SAL method for cloud processes, the threshold level used by object identification algorithms induces high sensitivity and unstable behavior of the object-dependent SAL scores, especially the S and L2 components. Two sensitivity indicators, $G_s:\mathbf{x}\mapsto s\in[0,1]$1 and $G_s:\mathbf{x}\mapsto s\in[0,1]$2, are derived from univariate cumulative distribution functions to assess this sensitivity without computationally expensive iterative recalculation. The paper further shows that, for large-scale cloud data, changes in the parameters may have larger effects on the object dependent SAL scores than a complete loss of temporal collocation (Weniger et al., 2016). This is a cautionary result: a “sensitivity analysis of fields” and a “field of sensitivities” are not interchangeable notions.

4. Learned and variational constructions

In Neural Radiance Fields, the Sensitivity Field is introduced as a learnable quantity that quantifies the spatially varying impact of geometric perturbations on rendering quality (Kim et al., 22 Oct 2025). It is parameterized as an MLP,

$G_s:\mathbf{x}\mapsto s\in[0,1]$3

where $G_s:\mathbf{x}\mapsto s\in[0,1]$4 is a sigmoid-normalized scalar sensitivity. The companion Perturbation Field predicts appearance and geometry perturbations,

$G_s:\mathbf{x}\mapsto s\in[0,1]$5

The learned sensitivity then adaptively constrains geometric perturbations through a differentiable soft clamping operation,

$G_s:\mathbf{x}\mapsto s\in[0,1]$6

so that high sensitivity drives perturbations toward zero, while low sensitivity allows larger perturbations. Training jointly minimizes a protection loss and a naturalness loss,

$G_s:\mathbf{x}\mapsto s\in[0,1]$7

Empirically, the sensitivity-guided method is reported to outperform fixed-bound baselines in the naturalness–protection trade-off, while random or complementary sensitivities degrade quality (Kim et al., 22 Oct 2025).

In experimental mechanics, the closely related notion is the sensitivity-based virtual field. The virtual fields method identifies material parameters from full-field deformation data by balancing internal and external virtual work, and its performance depends critically on virtual-field selection. The sensitivity-based virtual fields framework develops three variants: discrete SBVFs, variation-based SBVFs formulated using directional Gâteaux derivatives, and analytically differentiated SBVFs tailored to strain-invariant-based models (Nikolov et al., 3 Sep 2025). The variational construction enforces

$G_s:\mathbf{x}\mapsto s\in[0,1]$8

with the stress variation written as

$G_s:\mathbf{x}\mapsto s\in[0,1]$9

so that

$A_{eff}=\Phi/B$0

The paper reports sharper cost-function minima, improved robustness under synthetic noise, automated parameter-wise virtual-field construction, and a quantified role for data richness through a sharpness metric $A_{eff}=\Phi/B$1 (Nikolov et al., 3 Sep 2025).

These two literatures differ sharply in application, but both use sensitivity fields to regularize or optimize an inverse problem. One learns a field over 3D space; the other constructs parameter-specific virtual displacements from constitutive sensitivities. The shared strategy is to encode local informativeness rather than enforce a globally uniform constraint.

5. Electromagnetic sensing and metrology

Several sensor literatures use sensitivity fields, or equivalent field-response metrics, to quantify weak-signal detectability and spatial response. In a Cs vapor cell with built-in parallel electrodes, a Rydberg-atom sensor measures electric fields in the ULF, VLF, and LF bands at 1 kHz, 10 kHz, and 100 kHz. The built-in electrodes overcome the low-frequency electric field screening effect caused by alkali-metal atoms adsorbed on the inner surface of the container, and an auxiliary DC field amplifies the AC response through the cross-term $A_{eff}=\Phi/B$2. The reported minimum detectable field strengths are $A_{eff}=\Phi/B$3V/cm, $A_{eff}=\Phi/B$4V/cm, and $A_{eff}=\Phi/B$5V/cm, with corresponding sensitivities $A_{eff}=\Phi/B$6, $A_{eff}=\Phi/B$7, and $A_{eff}=\Phi/B$8; the linear dynamic range is over 50 dB, with exact values 51 dB, 63 dB, and 61 dB, and the sensitivity is better than a 1-cm dipole antenna (Lei et al., 2024).

A different route to field imaging is proposed with paramagnetic polar molecules. Ensembles of $A_{eff}=\Phi/B$9 and $S_O(f)=N_O(f)/R_O(f)$0 molecules can image both electric and magnetic field components from the dc limit to the THz regime, with a 100-fold higher sensitivity than the previously implemented $S_O(f)=N_O(f)/R_O(f)$1Rb-atom scheme. For ac electric fields, the sensitivity is written as

$S_O(f)=N_O(f)/R_O(f)$2

and the predicted sensitivities are on the order of $S_O(f)=N_O(f)/R_O(f)$3–$S_O(f)=N_O(f)/R_O(f)$4 V/cm Hz$S_O(f)=N_O(f)/R_O(f)$5 for electric fields and about $S_O(f)=N_O(f)/R_O(f)$6 fT Hz$S_O(f)=N_O(f)/R_O(f)$7 for magnetic fields (Alyabyshev et al., 2012).

Distributed optical-fiber sensing introduces an explicitly spatial sensitivity profile. In OFDR measurements of Faraday rotation, Random Optical Grating using UV Exposure (ROGUE) reduces the noise floor relative to unexposed spun fiber; the unexposed fiber has an S/P minimum ratio 20 dB higher than the ROGUE, and the ROGUE yields a 40–45 dB increase in backscattering signal. With spatial filtering of the derivative of $S_O(f)=N_O(f)/R_O(f)$8, the local magnetic field is estimated as

$S_O(f)=N_O(f)/R_O(f)$9

The reported performance includes fields down to 10 mT with 10 cm spatial resolution and a simulated current-sensing noise floor of around 1 A with 40 probing loops spatial resolution (Leymonerie et al., 2023).

In superconducting readout systems, field sensitivity is often condensed into a figure of merit rather than represented as a field over space. For time-division SQUID multiplexers, the NIST designs use the effective area

$$g_m(\boldsymbol{\theta}) = |\psi(\mathbf{r}_m; \boldsymbol{\theta})|^2,$$0

to quantify pickup of external magnetic fields. Moving from half-loop to whole-loop gradiometric input coils in MUX09a reduces the SQ1 effective area to $$g_m(\boldsymbol{\theta}) = |\psi(\mathbf{r}_m; \boldsymbol{\theta})|^2,$$1m$$g_m(\boldsymbol{\theta}) = |\psi(\mathbf{r}_m; \boldsymbol{\theta})|^2,$$2 and the SQ2 effective area to $$g_m(\boldsymbol{\theta}) = |\psi(\mathbf{r}_m; \boldsymbol{\theta})|^2,$$3m$$g_m(\boldsymbol{\theta}) = |\psi(\mathbf{r}_m; \boldsymbol{\theta})|^2,$$4, compared with much larger values in MUX06a and MUX07a (Stiehl et al., 2010).

Quantum-sensing NMR provides another strong example of field-detection sensitivity reframed around nanoscale fluctuations. NV-based deuterium NMR detects statistical spin fluctuations rather than Boltzmann polarization, with a reported sensitivity enhancement of six to eight orders of magnitude over inductive detection while operating at magnetic fields two orders of magnitude lower than conventional NMR. The spin sensitivity is quantified by the minimum number of spins detected at $$g_m(\boldsymbol{\theta}) = |\psi(\mathbf{r}_m; \boldsymbol{\theta})|^2,$$5 in 1 second, with reported values of approximately $$g_m(\boldsymbol{\theta}) = |\psi(\mathbf{r}_m; \boldsymbol{\theta})|^2,$$6 spins/$$g_m(\boldsymbol{\theta}) = |\psi(\mathbf{r}_m; \boldsymbol{\theta})|^2,$$7 for PMMA-d$$g_m(\boldsymbol{\theta}) = |\psi(\mathbf{r}_m; \boldsymbol{\theta})|^2,$$8 and $$g_m(\boldsymbol{\theta}) = |\psi(\mathbf{r}_m; \boldsymbol{\theta})|^2,$$9 spins/$$E(\boldsymbol{\theta}) = \frac{1}{4M} \sum_{m=1}^{M} \left[ \tilde{g}_m - |\psi(\mathbf{r}_m; \boldsymbol{\theta})|^2 \right]^2.$$0 for phenanthrene-d$$E(\boldsymbol{\theta}) = \frac{1}{4M} \sum_{m=1}^{M} \left[ \tilde{g}_m - |\psi(\mathbf{r}_m; \boldsymbol{\theta})|^2 \right]^2.$$1, versus bulk-NMR values of approximately $$E(\boldsymbol{\theta}) = \frac{1}{4M} \sum_{m=1}^{M} \left[ \tilde{g}_m - |\psi(\mathbf{r}_m; \boldsymbol{\theta})|^2 \right]^2.$$2 and $$E(\boldsymbol{\theta}) = \frac{1}{4M} \sum_{m=1}^{M} \left[ \tilde{g}_m - |\psi(\mathbf{r}_m; \boldsymbol{\theta})|^2 \right]^2.$$3 spins/$$E(\boldsymbol{\theta}) = \frac{1}{4M} \sum_{m=1}^{M} \left[ \tilde{g}_m - |\psi(\mathbf{r}_m; \boldsymbol{\theta})|^2 \right]^2.$$4 (Singh et al., 28 Apr 2026).

6. Sensitivity reduction, immunity, and selective observables

Not all work on sensitivity fields seeks maximal response. A substantial literature aims instead to suppress unwanted field sensitivity while preserving desired channels of detection. In low-angular-momentum Rydberg states of $$E(\boldsymbol{\theta}) = \frac{1}{4M} \sum_{m=1}^{M} \left[ \tilde{g}_m - |\psi(\mathbf{r}_m; \boldsymbol{\theta})|^2 \right]^2.$$5Rb, a non-resonant microwave dressing field at 38.465 GHz is used to eliminate the static electric dipole moment difference between the $$E(\boldsymbol{\theta}) = \frac{1}{4M} \sum_{m=1}^{M} \left[ \tilde{g}_m - |\psi(\mathbf{r}_m; \boldsymbol{\theta})|^2 \right]^2.$$6 and $$E(\boldsymbol{\theta}) = \frac{1}{4M} \sum_{m=1}^{M} \left[ \tilde{g}_m - |\psi(\mathbf{r}_m; \boldsymbol{\theta})|^2 \right]^2.$$7 states in dc fields of approximately 1 V/cm. The dressed-state energy is expanded as

$$E(\boldsymbol{\theta}) = \frac{1}{4M} \sum_{m=1}^{M} \left[ \tilde{g}_m - |\psi(\mathbf{r}_m; \boldsymbol{\theta})|^2 \right]^2.$$8

leading to an effective differential dipole moment

$$E(\boldsymbol{\theta}) = \frac{1}{4M} \sum_{m=1}^{M} \left[ \tilde{g}_m - |\psi(\mathbf{r}_m; \boldsymbol{\theta})|^2 \right]^2.$$9

and the nulling condition

$$\frac{\partial E}{\partial \boldsymbol{\theta}} = \frac{1}{M} \sum_{m=1}^M \Re \left\{ \psi^*(\mathbf{r}_m;\boldsymbol{\theta})\,\chi(\mathbf{r}_m;\boldsymbol{\theta}) \right\} \, \Delta_m(\boldsymbol{\theta}),$$0

The reported standard deviation of the measured transition frequency over 0 to 1.5 V/cm decreases from about 7 MHz without dressing to 0.06 MHz with dressing, while the anomalous spectral doublet is attributed to polarization ellipticity in the dressing field (Jones et al., 2013).

For circular Rydberg states, non-resonant dressing can null second-order as well as first-order dc sensitivity, but the directional alignment of dressing and dc fields is critical. The perturbative nulling condition for the quadratic response is

$$\frac{\partial E}{\partial \boldsymbol{\theta}} = \frac{1}{M} \sum_{m=1}^M \Re \left\{ \psi^*(\mathbf{r}_m;\boldsymbol{\theta})\,\chi(\mathbf{r}_m;\boldsymbol{\theta}) \right\} \, \Delta_m(\boldsymbol{\theta}),$$1

and the paper emphasizes that sensitivity to field misalignment is significantly larger for circular states than for low-angular-momentum Rydberg states of Rb (Ni et al., 2015). A separate bichromatic scheme for Cs $$\frac{\partial E}{\partial \boldsymbol{\theta}} = \frac{1}{M} \sum_{m=1}^M \Re \left\{ \psi^*(\mathbf{r}_m;\boldsymbol{\theta})\,\chi(\mathbf{r}_m;\boldsymbol{\theta}) \right\} \, \Delta_m(\boldsymbol{\theta}),$$2 and $$\frac{\partial E}{\partial \boldsymbol{\theta}} = \frac{1}{M} \sum_{m=1}^M \Re \left\{ \psi^*(\mathbf{r}_m;\boldsymbol{\theta})\,\chi(\mathbf{r}_m;\boldsymbol{\theta}) \right\} \, \Delta_m(\boldsymbol{\theta}),$$3 uses two microwave tones to reduce dc-field sensitivity by factors of 95 and 1600, respectively, and can cancel both second- and fourth-order terms in the polarizability of a single Rydberg state (Booth et al., 2017).

At the opposite extreme, transition-edge sensors show no observable sensitivity to strong dc electric fields. A Mo/Au TES in a parallel-electrode configuration experienced fields up to 90 kV/m, with no significant or systematic changes in current-voltage characteristics, thermal conductance, or response time. The fitted time constants at 0 and 20 V were 0.370 ms and 0.368 ms, and upper bounds were reported as $$\frac{\partial E}{\partial \boldsymbol{\theta}} = \frac{1}{M} \sum_{m=1}^M \Re \left\{ \psi^*(\mathbf{r}_m;\boldsymbol{\theta})\,\chi(\mathbf{r}_m;\boldsymbol{\theta}) \right\} \, \Delta_m(\boldsymbol{\theta}),$$4 and $$\frac{\partial E}{\partial \boldsymbol{\theta}} = \frac{1}{M} \sum_{m=1}^M \Re \left\{ \psi^*(\mathbf{r}_m;\boldsymbol{\theta})\,\chi(\mathbf{r}_m;\boldsymbol{\theta}) \right\} \, \Delta_m(\boldsymbol{\theta}),$$5 (Patel et al., 2023). Here, “sensitivity” is operationally a null result, but the measurement is still cast in the language of field response.

Selective observables can also remove nuisance sensitivities. In magnetic-field and atomic-EDM measurements, initializing atoms in the $$\frac{\partial E}{\partial \boldsymbol{\theta}} = \frac{1}{M} \sum_{m=1}^M \Re \left\{ \psi^*(\mathbf{r}_m;\boldsymbol{\theta})\,\chi(\mathbf{r}_m;\boldsymbol{\theta}) \right\} \, \Delta_m(\boldsymbol{\theta}),$$6 state and measuring all individual magnetic sublevels yields a shot-noise-limited phase uncertainty

$$\frac{\partial E}{\partial \boldsymbol{\theta}} = \frac{1}{M} \sum_{m=1}^M \Re \left\{ \psi^*(\mathbf{r}_m;\boldsymbol{\theta})\,\chi(\mathbf{r}_m;\boldsymbol{\theta}) \right\} \, \Delta_m(\boldsymbol{\theta}),$$7

which is smaller than conventional Larmor precession by a factor $$\frac{\partial E}{\partial \boldsymbol{\theta}} = \frac{1}{M} \sum_{m=1}^M \Re \left\{ \psi^*(\mathbf{r}_m;\boldsymbol{\theta})\,\chi(\mathbf{r}_m;\boldsymbol{\theta}) \right\} \, \Delta_m(\boldsymbol{\theta}),$$8. When the populations in the even or odd magnetic sublevels are combined, the resulting signals are independent of the tensor Stark shift and the second order Zeeman shift (Tang et al., 2018). This is not a spatial sensitivity field, but it is a precise example of engineering sensitivity to preserve odd-in-$$\frac{\partial E}{\partial \boldsymbol{\theta}} = \frac{1}{M} \sum_{m=1}^M \Re \left\{ \psi^*(\mathbf{r}_m;\boldsymbol{\theta})\,\chi(\mathbf{r}_m;\boldsymbol{\theta}) \right\} \, \Delta_m(\boldsymbol{\theta}),$$9 physics while eliminating quadratic contaminants.

7. Broader physical and astrophysical usages

In space-based gravitational-wave interferometers, “sensitivity” is generalized from strain detection to ultralight bosonic fields. Using time-delay interferometry to suppress laser frequency noise, the response of TDI combinations to scalar and vector fields is characterized through averaged transfer functions, and the sensitivity curve of channel $$\Delta_m(\boldsymbol{\theta}) = \tilde{g}_m - |\psi(\mathbf{r}_m;\boldsymbol{\theta})|^2$$0 is defined by

$$\Delta_m(\boldsymbol{\theta}) = \tilde{g}_m - |\psi(\mathbf{r}_m;\boldsymbol{\theta})|^2$$1

The framework is applied to LISA, Taiji, and TianQin, with low-velocity dark-matter-like fields showing transfer-function scalings of $$\Delta_m(\boldsymbol{\theta}) = \tilde{g}_m - |\psi(\mathbf{r}_m;\boldsymbol{\theta})|^2$$2 below $$\Delta_m(\boldsymbol{\theta}) = \tilde{g}_m - |\psi(\mathbf{r}_m;\boldsymbol{\theta})|^2$$3 and $$\Delta_m(\boldsymbol{\theta}) = \tilde{g}_m - |\psi(\mathbf{r}_m;\boldsymbol{\theta})|^2$$4 for $$\Delta_m(\boldsymbol{\theta}) = \tilde{g}_m - |\psi(\mathbf{r}_m;\boldsymbol{\theta})|^2$$5 Hz (Yu et al., 2023). This usage extends the sensitivity-field idea into frequency space and coupling-constraint space rather than physical space.

Cosmology supplies another related extension: sensitivity of field statistics to model parameters. In simulations of massive neutrinos, the cosmic velocity field is decomposed into magnitude, divergence, vorticity, and dispersion. The vorticity spectrum is reported as the most sensitive observable to the neutrino mass sum $$\Delta_m(\boldsymbol{\theta}) = \tilde{g}_m - |\psi(\mathbf{r}_m;\boldsymbol{\theta})|^2$$6, with the null hypothesis of massless neutrinos incompatible with vorticity and divergence spectra from $$\Delta_m(\boldsymbol{\theta}) = \tilde{g}_m - |\psi(\mathbf{r}_m;\boldsymbol{\theta})|^2$$7 eV at $$\Delta_m(\boldsymbol{\theta}) = \tilde{g}_m - |\psi(\mathbf{r}_m;\boldsymbol{\theta})|^2$$8-values 0.03 and 0.07, respectively (Zhou et al., 2021). This is not a “sensitivity field” in the explicit spatial-map sense, but it is a field-based sensitivity analysis in which the response of different field components carries distinct inferential power.

Biophysical theory introduces yet another formulation, where sensitivity is the minimum detectable electric field of an active membrane. The membrane polarization $$\Delta_m(\boldsymbol{\theta}) = \tilde{g}_m - |\psi(\mathbf{r}_m;\boldsymbol{\theta})|^2$$9 is modeled by an overdamped Langevin equation with passive noise, active noise, and spatial correlations. In the active case, the polarization fluctuation spectrum is

$$\chi(\mathbf{r}; \boldsymbol{\theta}) = \int \psi_0(\mathbf{r'})\,\chi_v(\mathbf{r'}; \boldsymbol{\theta})\,h_G(\mathbf{r}-\mathbf{r}')\,d\mathbf{r}',$$0

and the rms voltage fluctuation setting the noise floor is

$$\chi(\mathbf{r}; \boldsymbol{\theta}) = \int \psi_0(\mathbf{r'})\,\chi_v(\mathbf{r'}; \boldsymbol{\theta})\,h_G(\mathbf{r}-\mathbf{r}')\,d\mathbf{r}',$$1

The paper argues that activity, particularly spatial correlation effects from protein crowding, can lower the minimum detectable field below the thermal noise limit predicted by equilibrium statistical mechanics (Mathew et al., 2024).

Taken together, these works show that sensitivity fields are best understood as resolved response objects whose mathematical realization depends on the host problem. They may be gradients, scalar volumes, transfer functions, virtual fields, or figures of merit; they may be designed to maximize detectability, to regularize estimation, or to null unwanted coupling. The concept remains unified not by a single formula, but by a shared objective: making the structure of response explicit enough to analyze, optimize, or suppress.