Fragility Index: Cross-Disciplinary Measure

• Fragility Index is a cross-disciplinary measure that quantifies how susceptible systems, conclusions, or properties are to specific perturbations.
• It is applied in glass physics, clinical trials, extreme-value theory, and network analysis, each using tailored metrics like Angell slopes, outcome counts, or graph centrality.
• Its computation and interpretation vary by context, emphasizing controlled perturbations as key to reversing conclusions or identifying kinetic and systemic instabilities.

“Fragility Index” does not denote a single invariant across the research literature. In current usage it names a family of quantities that measure how easily a system, conclusion, trajectory, or property changes under a specified perturbation. In glass physics and soft matter it usually denotes a kinetic fragility extracted from an Angell plot; in clinical statistics it denotes the minimum number of outcome changes needed to overturn a result; in extreme-value theory it denotes a conditional exceedance count; in several network and machine-learning settings it denotes graph- or risk-based instability; and in one biomedical aging usage represented here it is explicitly identified with the Frailty Index (Mondal et al., 2020, Maity et al., 2024, Falk et al., 2011, Yang et al., 18 Feb 2025, Farrell et al., 2016).

1. Terminological scope

The common feature across these usages is a focus on susceptibility to reversal, clustering, or collapse under controlled perturbation. What varies is the perturbation itself: temperature or density change, outcome relabeling, censoring reclassification, threshold exceedance, graph reorganization, or language expansion. This suggests that “Fragility Index” is best understood as a cross-disciplinary label for conditional sensitivity measures rather than as a single standardized statistic.

Domain	Quantity	Core definition
Glass-forming liquids	kinetic fragility $m$	slope of an Angell plot at $T_g$
Soft colloids	density-controlled kinetic fragility $m$	slope at $\zeta_g$ on $\log \tau_\alpha$ versus $\zeta/\zeta_g$
Clinical trials	FI	minimum number of outcome changes needed to reverse a study conclusion
Single-arm survival trials	FI	minimum censored observations reclassified as events to overturn a posterior survival criterion
Extreme-value theory	FI	$FI=\lim E(N_s\mid N_s>0)$
Aging	Frailty Index $f$	deficit proportion $f=\frac{1}{n}\sum_i d_i$

These usages are all explicitly documented in the literature represented here (Jaiswal et al., 2016, Mondal et al., 2020, Baer et al., 2022, Maity et al., 2024, Falk et al., 2011, Farrell et al., 2016).

2. Kinetic fragility in glass physics and soft matter

In glass science, the fragility index is a kinetic descriptor of how rapidly transport or relaxation slows as the glass transition is approached. The standard Angell definition used for molecular, metallic, and network liquids is

$$m=\left.\frac{\partial \log \eta(T)}{\partial (T_g/T)}\right|_{T=T_g},$$

or equivalently through $T_g$0 rather than $T_g$1. Small $T_g$2 corresponds to strong liquids with nearly Arrhenius behavior, while large $T_g$3 corresponds to fragile liquids with strongly super-Arrhenius behavior. A major comparative result is the inverse correlation between $T_g$4 and the reduced Arrhenius crossover temperature $T_g$5: metallic liquids cluster near $T_g$6, molecular liquids near $T_g$7, and network liquids above $T_g$8; the same study reports $T_g$9 for metallic and molecular liquids and finds it uncorrelated with $m$0 (Jaiswal et al., 2016). Related work connects fragility to thermodynamic anomalies: the Prigogine-Defay ratio $m$1 approaches values near $m$2 at high $m$3 and rises above $m$4 for low-fragility liquids, with the empirical fit $m$5 and $m$6 (Loidl et al., 2024). In simple-liquid theory, fragility has also been estimated from structural inputs by combining DFT-based configurational entropy, Adam-Gibbs, and FNH-derived relaxation times (Premkumar et al., 2015). In bulk metallic glass formers, the principal fragility parameter is often the VFT strength parameter $m$7, where larger $m$8 means stronger behavior; the review literature reports moderately strong behavior for many bulk metallic glasses and discusses fragile-to-strong transitions in systems such as Vitreloy 1 (Busch et al., 2014).

Soft-colloid work preserves the Angell logic but replaces temperature by concentration or density. For deformable microgel suspensions, the operational variable is $m$9, the glass point $\zeta_g$0 is fixed by $\zeta_g$1 with $\zeta_g$2, and the kinetic fragility is defined as the slope

$\zeta_g$3

Under a Vogel-Fulcher-type density law, this yields $\zeta_g$4 with $\zeta_g$5, and the three analyzed samples give $\zeta_g$6 for SC-I, SC-II, and SC-III (Mondal et al., 2020). In chalcogenides, fragility is strongly processing-dependent: homogenized $\zeta_g$7 melts show a super-strong minimum near $\zeta_g$8, while homogenized $\zeta_g$9 melts reach $\log \tau_\alpha$0 at $\log \tau_\alpha$1; both studies argue that incomplete homogenization inflates apparent fragility and that super-strong compositions create diffusion bottlenecks during melt synthesis (Bhageria et al., 2013, Ravindren et al., 2013). In data-driven materials modeling, "ViscNet" predicts $\log \tau_\alpha$2, $\log \tau_\alpha$3, and $\log \tau_\alpha$4 as the latent parameters of the MYEGA viscosity law, so fragility becomes a directly predicted descriptor rather than a post hoc fit parameter (Cassar, 2020).

3. Clinical and biomedical meanings

In clinical-trial statistics, the classical fragility index is defined as the minimum number of cases whose outcomes must be modified to reverse statistical significance. The literature represented here emphasizes that this quantity depends on two operations: selecting the cases to be changed and specifying how their outcomes are changed. A central criticism is that the classical FI only guarantees the existence of one favorable subset, so a small FI may depend on an atypical set of cases. The proposed remedy is the stochastic generalized fragility index, $\log \tau_\alpha$5, defined as the minimum $\log \tau_\alpha$6 such that more than $\log \tau_\alpha$7 of case collections of size $\log \tau_\alpha$8 have a permitted outcome modification reversing significance; $\log \tau_\alpha$9 controls which outcome changes are sufficiently likely to be allowed (Baer et al., 2022).

A newer survival-analysis generalization defines FI for time-to-event endpoints in single-arm clinical trials. Under an exponential survival model with Gamma prior, it is the smallest number of censored observations with the shortest censoring times that must be reclassified as uncensored events so that the posterior probability that the median survival exceeds a prespecified threshold $\zeta/\zeta_g$0 falls below a prespecified confidence level $\zeta/\zeta_g$1. The formal decision quantity is

$\zeta/\zeta_g$2

with posterior $\zeta/\zeta_g$3. In the reported case studies, FI values of $\zeta/\zeta_g$4 or $\zeta/\zeta_g$5 were obtained for three single-arm datasets (Maity et al., 2024).

A distinct biomedical usage appears in aging research, where the paper explicitly treats the “Fragility Index” as the Frailty Index. There the quantity is the deficit proportion

$\zeta/\zeta_g$6

computed in the model from the $\zeta/\zeta_g$7 most connected non-mortality nodes. The theoretical maximum is $\zeta/\zeta_g$8, but observational studies often report an apparent ceiling near $\zeta/\zeta_g$9. The cited network model attributes this submaximal ceiling not to an intrinsic biological bound but to a deficit sensitivity parameter $FI=\lim E(N_s\mid N_s>0)$0, with $FI=\lim E(N_s\mid N_s>0)$1 used as a false-negative rate in deficit detection (Farrell et al., 2016).

4. Extreme-value, exceedance, and systemic-risk formulations

In multivariate extreme-value theory, the fragility index is the asymptotic expected number of exceedances conditional on at least one exceedance. For a random vector $FI=\lim E(N_s\mid N_s>0)$2 and threshold $FI=\lim E(N_s\mid N_s>0)$3,

$FI=\lim E(N_s\mid N_s>0)$4

if the limit exists. In this literature, $FI=\lim E(N_s\mid N_s>0)$5 is interpreted as asymptotic stability and $FI=\lim E(N_s\mid N_s>0)$6 as fragility, because exceedances then tend to occur jointly rather than singly (Falk et al., 2011). For different margins satisfying tail comparability $FI=\lim E(N_s\mid N_s>0)$7, the same paper gives the explicit formula

$FI=\lim E(N_s\mid N_s>0)$8

thereby decomposing fragility into marginal tail weights $FI=\lim E(N_s\mid N_s>0)$9 and the D-norm that encodes extremal dependence (Falk et al., 2011).

In the process setting, the same object becomes a temporal clustering measure. For a continuous process $f$0 with sojourn time

$f$1

the paper proves that, under a functional domain-of-attraction condition for the copula process, the asymptotic conditional mean sojourn time coincides with the asymptotic fragility index and equals the generator constant $f$2 of the limiting max-stable process (Falk et al., 2011). This identifies fragility with a conditional cluster size in time rather than just with a count on a fixed grid.

A block-level generalization treats a system partitioned into blocks $f$3. If $f$4 counts how many blocks contain at least one exceedance above $f$5, then

$f$6

and, under identical continuous margins in the domain of attraction of a multivariate extreme-value law $f$7,

$f$8

The paper interprets this as the expected number of affected subsystems conditional on at least one subsystem being affected, and uses it to distinguish within-block from between-block fragility in financial data (Ferreira et al., 2011).

5. Networked, game-theoretic, and machine-learning uses

A graph-theoretic usage appears in chess. There the fragility index is the fragility score

$f$9

where $f=\frac{1}{n}\sum_i d_i$0 is the normalized directed betweenness centrality of piece $f=\frac{1}{n}\sum_i d_i$1 in the attack/defense interaction graph and $f=\frac{1}{n}\sum_i d_i$2 if the piece is under attack, $f=\frac{1}{n}\sum_i d_i$3 otherwise. The score measures how tactically unstable a position is when attacked pieces are also structurally central. On a dataset of $f=\frac{1}{n}\sum_i d_i$4 games, maximum fragility typically occurs around ply $f=\frac{1}{n}\sum_i d_i$5, the attacked piece with largest contribution is a pawn in about $f=\frac{1}{n}\sum_i d_i$6 of cases and a knight in about $f=\frac{1}{n}\sum_i d_i$7, and aligned fragility curves are fit by

$f=\frac{1}{n}\sum_i d_i$8

with a gradual buildup before and a longer decay after the peak (Barthelemy, 2024).

Financial-network work uses the term less uniformly. One study of $f=\frac{1}{n}\sum_i d_i$9 global financial indices explicitly does not define a single scalar called a fragility index, but instead treats fragility as a family of network-centric indicators on MST-plus-threshold correlation networks; during stress periods, mean correlation, number of edges, edge density, clustering coefficient, communication efficiency, clique number, and OR/MR/HR curvatures rise, while diameter, average shortest path length, modularity, and FR curvature fall (Samal et al., 2021). A related stock-market study uses average Ollivier-Ricci curvature as the operative fragility indicator, with

$$m=\left.\frac{\partial \log \eta(T)}{\partial (T_g/T)}\right|_{T=T_g},$$0

and concludes that the indicator is useful mainly as an ex-post crisis hallmark rather than as a reliable early-warning signal (García et al., 2024). This suggests that some network literatures operationalize fragility through a metric panel or a geometry-derived proxy rather than through a unique canonical scalar.

In machine learning, the fragility index can be defined as a risk-averse measure of confident misjudgment. For ranking error $$m=\left.\frac{\partial \log \eta(T)}{\partial (T_g/T)}\right|_{T=T_g},$$1 and target $$m=\left.\frac{\partial \log \eta(T)}{\partial (T_g/T)}\right|_{T=T_g},$$2, the proposed classifier-level quantity is

$$m=\left.\frac{\partial \log \eta(T)}{\partial (T_g/T)}\right|_{T=T_g},$$3

Under KL divergence, it yields the tail bound

$$m=\left.\frac{\partial \log \eta(T)}{\partial (T_g/T)}\right|_{T=T_g},$$4

so lower FI means lighter right tails of confident ranking errors (Yang et al., 18 Feb 2025). A related but reversed-polarity construct is the "Credibility Index via Explanation Stability" (CIES), which the paper explicitly says is not literally a Fragility Index; there

$$m=\left.\frac{\partial \log \eta(T)}{\partial (T_g/T)}\right|_{T=T_g},$$5

and low CIES corresponds to high explanation fragility under realistic perturbations (Vaduva et al., 5 Mar 2026).

6. Formal abstractions, non-universality, and recurring caveats

A particularly abstract reformulation appears in model theory. The fragility spectrum defines

$$m=\left.\frac{\partial \log \eta(T)}{\partial (T_g/T)}\right|_{T=T_g},$$6

as the minimal language expansion needed to degrade a complete theory $$m=\left.\frac{\partial \log \eta(T)}{\partial (T_g/T)}\right|_{T=T_g},$$7 from property $$m=\left.\frac{\partial \log \eta(T)}{\partial (T_g/T)}\right|_{T=T_g},$$8 to $$m=\left.\frac{\partial \log \eta(T)}{\partial (T_g/T)}\right|_{T=T_g},$$9, with semantic and syntactic versions shown equivalent and with monotonicity, subadditivity, and a chain rule. The paper presents $T_g$00 as having infinite fragility for stability and $T_g$01 as having fragility $T_g$02 for $T_g$03-stability, thereby recasting fragility as resilience of tameness properties under expansion rather than as statistical or dynamical sensitivity (Adilkhan, 13 Aug 2025).

Across the literatures, several cautions recur. In soft colloids, the fragility–Poisson-ratio relation is explicitly said to be model-dependent, restricted to Angell-like/Vogel-Fulcher-like curves, and “still at the level of a hypothesis” (Mondal et al., 2020). In the Prigogine-Defay analysis, the $T_g$04 relation is statistical rather than exact, with substantial scatter and even some values below $T_g$05 (Loidl et al., 2024). In clinical statistics, the classical FI can be dominated by an atypical subset of cases, which is precisely the defect targeted by the stochastic generalized fragility indices (Baer et al., 2022). In single-arm survival studies, the resulting FI depends on the exponential survival model, the Gamma prior, and the rule that the shortest censoring times are perturbed first (Maity et al., 2024). In financial-network work, one paper explicitly declines to define a single scalar fragility index (Samal et al., 2021), while another concludes that curvature-based fragility is mainly ex-post rather than predictive (García et al., 2024). Taken together, these results indicate that any interpretation of a fragility index is inseparable from the perturbation class, the reversal criterion, and the structural model chosen in the underlying field.