B-Value: Multidisciplinary Insights

• B-value is a domain-dependent parameter defined differently across seismology, diffusion MRI, and sensitivity analysis, each with distinct applications and interpretations.
• In seismology, the b-value represents the slope of the Gutenberg–Richter relation, indicating the balance between small and large earthquakes and correlating with stress conditions.
• In diffusion MRI and bias analysis, b-values determine diffusion weighting and act as thresholds for statistical significance, underscoring their role in imaging contrast and result robustness.

B-value is a domain-dependent scientific term. In seismology, the $b$-value is the slope parameter of the Gutenberg–Richter frequency–magnitude relation and governs the relative abundance of small and large earthquakes. In diffusion MRI, b-values are diffusion-weighting strengths that determine signal attenuation and contrast. In sensitivity analysis for combined estimators, the b-value is a critical maximum relative bias at which a result ceases to be statistically significant. The cited literature uses the same label for these distinct quantities, so the term must be interpreted within its disciplinary context (Köhler et al., 12 Feb 2026, Merisaari et al., 2020, Lin et al., 18 Feb 2026).

1. Seismological definition in the Gutenberg–Richter framework

In statistical seismology, the standard definition is the Gutenberg–Richter relation

$\log_{10} N = a - bM,$

where $N$ is the cumulative number of earthquakes with magnitude at least $M$, $a$ controls overall productivity, and $b$ is the slope of the magnitude–frequency distribution. A larger $b$-value means the distribution is relatively richer in small events, while a smaller $b$-value means relatively more large events (Köhler et al., 12 Feb 2026, Tan et al., 2018).

This interpretation is also used in energy form. Using the empirical scaling $E \propto 10^{cM}$ with $c \approx 1.5$, the magnitude distribution becomes

$\log_{10} N = a - bM,$0

so the $\log_{10} N = a - bM,$1-value is directly tied to the power-law exponent of earthquake energy release. A typical $\log_{10} N = a - bM,$2 corresponds to $\log_{10} N = a - bM,$3–$\log_{10} N = a - bM,$4 (Varotsos et al., 2012).

The same parameter can also be written through $\log_{10} N = a - bM,$5, which is common in likelihood-based formulations of the magnitude density

$\log_{10} N = a - bM,$6

This parameterization is useful when deriving estimators, completeness diagnostics, or spatial fields of the frequency–magnitude slope (Lippiello et al., 7 Apr 2025).

2. Physical interpretation, stress dependence, and proposed mechanisms

The seismological literature summarized here consistently treats the $\log_{10} N = a - bM,$7-value as physically informative, but not as a uniquely interpretable stress meter. Low $\log_{10} N = a - bM,$8-values are commonly associated with relatively higher proportions of large events and are often interpreted as indicating higher differential stress, stronger asperity-like behavior, or more critically loaded crust. High $\log_{10} N = a - bM,$9-values indicate relatively more small events and are often linked to more fractured, heterogeneous, or less highly stressed regions (Kamer, 2014, Mousavi et al., 22 May 2026).

A direct field example is provided by Axial Seamount, where about 60,000 microearthquakes in a $N$0 block of crust were analyzed under tidal loading of about $N$1 kPa. Above a threshold stress amplitude of about $N$2 kPa, the $N$3-value decreases systematically as tidal stress increases. The reported sensitivity is about $N$4 b-value units per kPa change in Coulomb stress, and the low-stress and high-stress groups differ at better than the 1% level using both Utsu’s test and a z-test (Tan et al., 2018). That result supports, but does not universalize, the use of $N$5-values to estimate small stress variations.

Several papers frame $N$6-value change as a precursor-like quantity. In a two-dimensional continuous-time forest-fire model, a decrease in the effective $N$7-value before large events is explained by aggregation of small stressed clusters into larger clusters; seismic quiescence and $N$8-value decrease arise from the same cluster-dynamics process (Mitsudo et al., 2015). In natural time analysis, a fall in $N$9 before a mainshock is interpreted as an increase in fluctuations of the order parameter $M$0, with temporal correlations in earthquake magnitudes contributing alongside heavy-tailed size statistics (Varotsos et al., 2012).

A more formal physical derivation is given by a recent coupled geometrical–mechanical model of fault networks. There, the $M$1-value follows from the power-law distribution of fault lengths $M$2, slip scaling $M$3, and seismic moment scaling $M$4, yielding

$M$5

In this view, $M$6 is neither purely geometric nor purely mechanical; it is a compound descriptor of fault-size statistics, rupture-area scaling, slip scaling, fault criticality, and fracture-energy dissipation (Pan et al., 6 Mar 2026).

3. Estimation, completeness, uncertainty, and mapping

Because catalog incompleteness strongly affects small magnitudes, $M$7-value estimation is method-sensitive. Standard maximum-likelihood estimation above a completeness threshold takes the form

$M$8

or, in Aki’s notation,

$M$9

These estimators are widely used, but their reliability depends on how completeness is handled (Köhler et al., 12 Feb 2026, Mousavi et al., 22 May 2026).

Finite-sample effects are also non-negligible. Ogata and Yamashina’s result for a GR-distributed sample gives

$a$0

so the estimated $a$1-value has a systematic positive bias of order $a$2. A later generalization shows

$a$3

where $a$4. For a true GR distribution, $a$5, and this condition can be used to identify the incompleteness magnitude $a$6 (Godano et al., 2023).

A major current concern is downward bias from incomplete catalogs. Synthetic ETAS experiments and five instrumental regional catalogs show that traditional threshold-based estimators tend to underestimate the true $a$7-value, whereas the b-more-positive estimator, based on positive magnitude differences between carefully selected earthquake pairs, recovers the true slope more accurately. The b-positive estimator is a special case with consecutive events only (Lippiello et al., 7 Apr 2025). Relatedly, one methodological framework for anomaly detection uses the b-positive method of Elst et al. so that only positive differences between successive magnitudes enter the estimate (Köhler et al., 12 Feb 2026).

Several papers argue that the usual thresholding workflow discards too much information. A probabilistic model for observable magnitudes writes the observed density as

$a$8

jointly modeling the true magnitude law and the magnitude-dependent detection probability. Maximum-likelihood estimation on the full observed distribution, rather than only on events above $a$9, is reported to improve $b$0-value estimation compared with hard-threshold approaches (Martinsson et al., 2018).

Spatial mapping introduces a second layer of uncertainty. A Bayesian inversion framework models $b$1 as a nonstationary Gaussian field with covariance aligned to local fault geometry, producing posterior expectations and posterior uncertainties for spatially varying $b$2-value maps (Holschneider et al., 2020). The HIST-PPM framework similarly estimates a smooth spatial field $b$3 by penalized likelihood with smoothing chosen by ABIC (Mousavi et al., 22 May 2026).

By contrast, the comment on Tormann et al.’s distance exponential weighted mapping argues that the optimization procedure requires a priori knowledge of the spatial $b$4-value distribution it seeks to infer. Changing the dominant synthetic $b$5-value changes the parameter set judged “optimal,” and the resulting maps can differ dramatically. The comment warns that apparent high- and low-$b$6 anomalies may be under-sampling artifacts, and inferred recurrence times can differ by up to two orders of magnitude depending on parameter choices (Kamer, 2014).

Sample size requirements are correspondingly stringent. Using synthetic catalogs generated from the angular frequency-magnitude distribution, the minimum complete-event counts needed to distinguish adjacent $b$7-values are reported as approximately $b$8 for resolution $b$9, $b$0 for $b$1, $b$2 for $b$3, and $b$4 for $b$5. Even crude binary discrimination between low and high $b$6 requires about 100 complete events, and reporting $b$7-value variations with $b$8 requires about 500 complete events (Kamer, 2014).

4. Spatiotemporal anomalies and earthquake forecasting

One recent methodological line does not treat a $b$9-value anomaly as a separately thresholded geophysical quantity. Instead, it constructs daily gridded $b$0-value fields over Japan on a $b$1 grid, producing a $b$2 array after 2000-01-01, and defines anomalies as spatiotemporal patterns in those fields that a classifier learns as predictive of an impending $b$3 event in the next day (Köhler et al., 12 Feb 2026).

In that framework, each learning sample is a $b$4 spatiotemporal block extracted from moving-window $b$5-value estimates. Positive samples are those followed by an $b$6 earthquake in the central cell on day $b$7; negative samples require no $b$8 event within an $b$9 L1 radius, no such event within $E \propto 10^{cM}$0 days, and at least 10 earthquakes for stable $E \propto 10^{cM}$1-value calculation. The hybrid network combines spatial 3D convolutions with a temporal convolutional network, uses LeakyReLU with $E \propto 10^{cM}$2, has 8081 parameters, and is trained with a strictly time-forward progressive meta-epoch scheme. Its output is an uncalibrated anomaly score, explicitly not a direct earthquake rate or calibrated hazard probability (Köhler et al., 12 Feb 2026).

A separate retrospective study in the Alborz region evaluates the spatial forecasting skill of $E \propto 10^{cM}$3-value maps against the background seismicity rate $E \propto 10^{cM}$4. There, $E \propto 10^{cM}$5 is treated as a stress-state indicator and $E \propto 10^{cM}$6 as a tectonic-loading indicator. The results show that $E \propto 10^{cM}$7 is consistently the stronger predictor at lower target magnitudes, while the $E \propto 10^{cM}$8-value improves steadily with increasing magnitude; the modified area skill score for $E \propto 10^{cM}$9 becomes positive above approximately $c \approx 1.5$0, which is interpreted as the magnitude at which low-$c \approx 1.5$1 anomalies begin to capture meaningful stress concentrations (Mousavi et al., 22 May 2026).

The combined $c \approx 1.5$2–$c \approx 1.5$3 forecast uses the intersection of low-$c \approx 1.5$4 and high-$c \approx 1.5$5 cells. In retrospective testing it achieves detection rates of $c \approx 1.5$6–$c \approx 1.5$7 at spatial alarm rates of $c \approx 1.5$8 and $c \approx 1.5$9 for $\log_{10} N = a - bM,$00 and $\log_{10} N = a - bM,$01, respectively. The same study emphasizes that $\log_{10} N = a - bM,$02-value is not a universal spatial forecast tool; its operational utility is target-magnitude dependent and complementary to background-rate information (Mousavi et al., 22 May 2026).

5. b-Values in diffusion MRI

In diffusion MRI, b-values are diffusion-weighting strengths rather than frequency–magnitude exponents. A $\log_{10} N = a - bM,$03 image has no diffusion weighting, while higher b-values increase diffusion contrast and reduce signal amplitude. In intravoxel incoherent motion imaging, low b-values are sensitive to the pseudo-diffusive perfusion component, high b-values suppress perfusion effects and emphasize tissue diffusion, and mid-range b-values help separate the two compartments (Jurek et al., 2024, Merisaari et al., 2020).

Brain IVIM studies show that signal noise is strongly b-value dependent. Two distinct peaks in coefficient of variation are reported: a physiological-noise peak at low b-values below $\log_{10} N = a - bM,$04, and a thermal-noise peak at high b-values where the diffusion-weighted signal is strongly attenuated. The choice of b-value distribution affects map homogeneity more than the number of averages, and a 12-minute optimized acquisition scheme is proposed as

$\log_{10} N = a - bM,$05

In that setting, $\log_{10} N = a - bM,$06 is relatively robust, whereas $\log_{10} N = a - bM,$07 and especially $\log_{10} N = a - bM,$08 are more sensitive to b-value design and noise structure (Merisaari et al., 2020).

Multi-b-value denoising exploits these cross-shell relationships directly. The MBD method feeds a CNN with images acquired along the same diffusion encoding direction but at different b-values, for example $\log_{10} N = a - bM,$09, an intermediate b-value, and a high b-value. In clinical data, MBD denoising of $\log_{10} N = a - bM,$10 outperformed N2N, CNNe, MPPCA, and ALGe in quantitative MAE; in spherical b-tensor encoding data, MBD denoising of $\log_{10} N = a - bM,$11 similarly outperformed the alternatives. The method uses the b-value dimension as a source of redundancy when directional redundancy is sparse (Jurek et al., 2024).

DW-SSFP introduces a further extension: it has no single intrinsic b-value because the signal is a mixture of coherence pathways with different diffusion sensitivities. An equivalent or effective b-value is therefore defined as the DW-SE b-value that would produce the same apparent ADC estimate as the measured DW-SSFP signal at a given flip angle. In this framework, the effective b-value is flip-angle dependent and reflects pathway weighting, relaxation, and non-Gaussian diffusion rather than a single fixed sequence parameter (Tendler et al., 2019).

6. Sensitivity-analysis b-value and related notation

In a distinct inferential setting, the b-value is defined for combining an unbiased but noisy estimator with a biased but more precise estimator. The model is

$\log_{10} N = a - bM,$12

with unknown relative bias bounded by

$\log_{10} N = a - bM,$13

A sequence of confidence intervals is then indexed by $\log_{10} N = a - bM,$14, and the b-value is the critical threshold

$\log_{10} N = a - bM,$15

that is, the maximum relative bias the result can tolerate before losing significance in the two-sided test $\log_{10} N = a - bM,$16 versus $\log_{10} N = a - bM,$17 (Lin et al., 18 Feb 2026).

This definition is explicitly a robustness measure rather than a physical parameter. Larger $\log_{10} N = a - bM,$18 means the conclusion is more robust to unknown bias; smaller $\log_{10} N = a - bM,$19 means the conclusion is more fragile. The paper develops this sensitivity-analysis logic for the precision-weighted estimator, the pretest estimator, and the soft-thresholding estimator, and recommends reporting the b-value associated with the soft-thresholding estimator because its confidence intervals are robust to unknown bias and achieve the lowest worst-case risk among the alternatives (Lin et al., 18 Feb 2026).

The capitalization of the term can be misleading because nearby literatures use related symbols for different quantities. In heavy-ion collisions, $\log_{10} N = a - bM,$20 denotes the deuteron coalescence parameter and acts as a probe of source volume and space-momentum correlations, with the approximate scaling $\log_{10} N = a - bM,$21 in the simplest picture (Gaebel et al., 2020). In aerospace certification, the B-basis value is the lower bound of a 95% confidence interval for the 10th percentile of a material-property distribution, so that at least 90% of the population is expected to exceed it with 95% confidence (Donfack-Siewe et al., 11 Jun 2026). These are separate definitions, not variants of the Gutenberg–Richter or MRI b-value.