Gutman Test: Graph & Statistical Methods

• Gutman Test encompasses graph invariants (Gutman index) used to quantify molecular structures and statistical tests based on likelihood and divergence.
• In graph theory, the Gutman index employs vertex classification and shortest-path distances to derive closed-form formulas for molecular graph families.
• In statistics, the Gutman Test uses measures like JS divergence for distribution comparison, achieving optimal error rates in hypothesis and multi-class tests.

The term "Gutman Test" encompasses two distinct domains: (1) discrete graph invariants, notably the Gutman index in chemical graph theory, and (2) statistical hypothesis testing, including distributional comparison using likelihood-based criteria and divergence-based methods. This article systematically delineates both contexts and their principal mathematical properties, focusing on modern instantiations in graph indices and distributional testing.

1. Gutman Index in Graph Theory

The Gutman index is a degree-weighted graph invariant defined for a connected graph $G=(V,E)$ with vertex set $V=\{v_1,...,v_n\}$, vertex degrees $d_i$, and pairwise shortest-path distances $d_{ij}$:

$$\operatorname{Gut}(G) = \frac{1}{2} \sum_{i=1}^n\sum_{j=1}^n d_i\,d_j\,d_{ij} = \sum_{1\le i<j\le n} d_i\,d_j\,d_{ij}$$

This quantity generalizes the classical Wiener index by incorporating vertex degrees as weights. Recent applications concern explicit computation of the Gutman index for families of molecular graphs modeling condensed-ring hydrocarbons (Sahir et al., 2023).

Closed-form Results for Pentagonal Chains

For the pentagonal cylinder chain $P_n$ and its Möbius variant $P'_n$ (each with $5n$ vertices, modeling linear or topologically twisted phenylenes), the Gutman index admits closed forms:

$$\operatorname{Gut}(P_n) = \begin{cases} 49n^3 + 64n^2 + 5n, & n \text{ even} \ 49n^3 + 64n^2 + 4n, & n \text{ odd} \end{cases}$$

$$\operatorname{Gut}(P'_n) = \begin{cases} 49n^3 + 64n^2 - 13n, & n \text{ even} \ 49n^3 + 64n^2 - 14n, & n \text{ odd} \end{cases}$$

These formulas result from detailed vertex classification (three types per pentagon: degree-3 and degree-2 classes), use of chain and cyclic automorphisms to reduce distance sum computations, and aggregation of arithmetic progressions of shortest paths (Sahir et al., 2023).

Methodological Architecture

• Vertex-type classification partitions vertices according to degree and adjacency, structuring the summation of weighted distances.
• Automorphism and symmetry arguments (standard and Möbius twist) minimize analytic cases.
• Explicit summation uses progression counts; Laplacian spectrum is not invoked for the Gutman index, in contrast to Kirchhoff-type indices.

Applications in Chemical Graph Theory

The Gutman index robustly correlates with molecular thermodynamic and structural descriptors. Its rapid computability for large, regular chain graphs enables direct integration into QSPR (quantitative structure–property relationship) studies for phenylenes and related hydrocarbons.

A notable asymptotic relation is $\operatorname{Gut} \approx 3\,\operatorname{Kf}^*$ (degree-Kirchhoff index) for large $n$, revealing an intrinsic connection between distance-based and resistance-based topological invariants (Sahir et al., 2023).

2. Gutman Test in Statistical Hypothesis Testing

The term "Gutman Test" in statistics denotes distributional comparison via likelihood-based or divergence-based methods. These critical two-sample and multi-classification tests can be formulated in distinct but related frameworks:

Two-way Likelihood Ratio (G) Test

For contingency tables, the G-test is derived from the likelihood ratio statistic:

$$G = 2\sum_{i=1}^r\sum_{j=1}^c O_{ij}\,\ln\left(\frac{O_{ij}}{E_{ij}}\right)$$

where $O_{ij}$ is the observed count and $E_{ij}$ the expected count under independence. Asymptotically, $G$ is distributed as $\chi^2$ with degrees of freedom $(r-1)(c-1)$ (Hoey, 2012). The G-test is exact for multinomial fits and preferred over Pearson's $\chi^2$ when expected counts are sparse or small.

3. Gutman Test for Distributional (Two-sample) Testing

In modern information-theoretic statistics, the Gutman test is deployed for comparing two empirical distributions, typically using the Jensen–Shannon (JS) divergence (Harsha et al., 14 Jan 2026):

$$D_{\rm JS}(T \| R) = \frac{1}{2} D_{\rm KL}\left(T \Big\| \frac{T+R}{2}\right) + \frac{1}{2} D_{\rm KL}\left(R \Big\| \frac{T+R}{2}\right)$$

For samples $X^n \sim P_1$, $Y^m \sim P_2$:

$$\text{Accept } H_0 \Longleftrightarrow D_{\rm JS}(\hat P_{X^n}, \hat P_{Y^m}) \le \gamma$$

The threshold $\gamma$ is set so that the type I error does not exceed a prescribed level $\alpha$. This approach is shown to be the generalized likelihood ratio test (GLRT) for the composite null hypothesis $P_1 = P_2$.

First-order and Second-order Asymptotics

In the regime where sample sizes grow proportionally, the optimal error exponent (Stein's exponent) is governed by the Bhattacharyya distance:

$$\lim_{n\to\infty} -\frac{1}{n}\ln\beta_n = 2 D_B(P_1, P_2), \quad D_B(P_1,P_2) = -\ln \sum_z \sqrt{P_1(z)P_2(z)}$$

Second-order refinements yield explicit $O(\sqrt n)$ corrections involving KL-variances at the Bhattacharyya-optimal mixture and quantiles of the $\chi^2_{k-1}$ distribution. This guarantees precise finite-sample error control (Harsha et al., 14 Jan 2026). Using invariant divergences (those whose Hessian is proportional to the Fisher metric) in place of JS maintains identical second-order performance.

Robust Goodness-of-fit Formulation

Two-sample Gutman testing can be recast as robust goodness-of-fit on product spaces, leveraging known connections between likelihood ratios and JS divergence minimization (Harsha et al., 14 Jan 2026).

4. Gutman Test in Multiple Classification and Sequential Testing

The classical Gutman M-ary test decides, given training sequences from $M$ unknown distributions $P_1,...,P_M$, whether a test sequence $Y^n$ matches any training distribution or should be rejected.

The test uses the generalized Jensen-Shannon (GJS) divergence with weight $\alpha$:

$$\operatorname{GJS}(P,Q;\alpha) = \alpha\,D(P\|\mu) + D(Q\|\mu), \quad \mu = \frac{\alpha P + Q}{1+\alpha}$$

The rule assigns $Y^n$ to the unique distribution $i$ where $\operatorname{GJS}(\hat P_i, \hat Q;\alpha)$ is minimal and below a threshold $\lambda$, otherwise rejects (Zhou et al., 2022).

Error exponent region

The achievable error exponent for fixed-length tests with reject option is

$$\lambda^* = \min_{i<j} \operatorname{GJS}(P_i, P_j; \alpha)$$

and rejection is minimized at this value. Sequential and two-phase extensions interpolate between fixed-length and optimal sequential exponent regions, using a two-phase protocol to achieve the full error exponent region without requiring a final reject option.

5. Comparative Summary and Principal Connections

Table: Gutman Tests — Summary of Domains

Context	Statistic/Index	Main Use
Graph theory	$\operatorname{Gut}(G) = \sum d_i d_j d_{ij}$	Molecular descriptor/QSPR
Contingency tables	$G = 2\sum O_{ij} \ln(O_{ij}/E_{ij})$	Likelihood-ratio test
Two-sample testing	JS divergence thresholding	Distribution equality testing
Multiple classification	GJS-divergence selection with reject option	M-ary sequence classification

The Gutman index and the various Gutman tests share the unifying theme of weighting pairwise comparisons (distances in graphs, divergences between distributions), but their implementations and applications are domain-specific. In both cases, symmetry, minimality, and error exponent optimality feature prominently. In distributional testing, recent advances reveal the Gutman test as the optimal GLRT in two-sample problems and establish its second-order asymptotic efficiency for practical, finite samples (Harsha et al., 14 Jan 2026). The two-phase protocol generalizes Gutman’s original test and achieves full sequential optimality under universal constraints (Zhou et al., 2022). In graph theory, the Gutman index enables explicit quantification of structure-property relationships in large molecular systems (Sahir et al., 2023).