---
title: Relative Strength (RATIO) Analysis
url: https://www.emergentmind.com/topics/relative-strength-ratio
type: topic
---

# Relative Strength (RATIO) Analysis

Relative strength denotes a comparative quantity that measures one effect, interaction, transition, evidential update, or population share relative to another. In the cited literature, it is not a single standardized object but a family of domain-specific constructs. It may be a direct physical ratio such as \(B(E2;2_2^+\rightarrow 0_1^+)/B(E2;2_1^+\rightarrow 0_1^+)\), an interaction-balance condition such as \(k_L/k_\theta\), a spin-orbit coupling ratio such as \(\alpha_0/\beta^\ast\), a posterior-to-prior evidential update \(RB(A\mid x)=\Pi(A\mid x)/\Pi(A)\), a density-ratio \(p_{\rm nu}(x)/p_{\rm de}(x)\) or its bounded relative form, a share such as \(TS_U/RS_U\), or a received-signal-strength ratio \(P_{r,j}/P_{r,i}\) [1401.4837, 2104.03198, 1003.1198, 2301.08994, 1810.12667, 2004.06294, 2107.00801]. In some works, moreover, “strength” is not itself a ratio but an ordering of latent utility differences inferred from local comparisons [2512.25023].

## 1. Formal patterns of relative strength

Across the cited works, relative strength appears in several recurring mathematical forms: a ratio of two observables, a posterior-to-prior update, a bounded comparison against a mixture denominator, or a proportion of a subgroup within a larger population. The unifying feature is comparative normalization: the numerator has meaning only relative to the chosen denominator or reference set.

| Setting | Representative quantity | Role |
|---|---|---|
| Nuclear spectroscopy | \(\dfrac{B(E2;2_2^+)}{B(E2;2_1^+)}\) | Relative quadrupole transition strength |
| Spin-orbit-coupled quantum wells | \(\dfrac{\alpha_0}{\beta^\ast}=\dfrac{U_{s+}+U_{s-}}{U_{s-}-U_{s+}}\) | Relative Rashba/Dresselhaus strength |
| Bayesian evidence | \(RB(A\mid x)=\dfrac{\Pi(A\mid x)}{\Pi(A)}\) | Evidential update from prior to posterior |
| Relative density-ratio estimation | \(r_\alpha(x)=\dfrac{p_{\rm nu}(x)}{\alpha p_{\rm nu}(x)+(1-\alpha)p_{\rm de}(x)}\) | Bounded comparison of two densities |
| Bibliometrics | \(\dfrac{TS_U}{RS_U}\) | Share of top scientists in a university |
| Visible-light positioning | \(\dfrac{P_{r,j}}{P_{r,i}}\) | Relative received optical power |

The denominator is not merely a scaling constant; it determines interpretation. In Bayesian work, a prior makes the relative belief ratio an evidential measure rather than a descriptive one. In relative density-ratio estimation, replacing \(p_{\rm de}\) by \(\alpha p_{\rm nu}+(1-\alpha)p_{\rm de}\) produces a bounded extension of the ordinary density-ratio. In bibliometrics, replacing an absolute count of top scientists by the share \(TS_U/RS_U\) removes the direct size effect that would favor large institutions [2301.08994, 2107.00801, 1810.12667].

A recurrent interpretive caution is that the existence of a ratio does not imply that the raw magnitude alone is the correct measure of “strength.” The Bayesian literature treats the sign of \(RB-1\) as the direction of evidence and separates its calibration from the ratio itself, while ResponseRank learns ordered utility gaps without introducing an explicit ratio variable [2301.08994, 1401.4215, 2512.25023].

## 2. Physical interactions and collective phenomena

In nuclear structure, relative strength arises as a ratio of reduced \(E2\) transition probabilities. For \(^{94}\)Zr, the quantity extracted by low-\(q\) inelastic electron scattering is
\[
\frac{B(E2;2_2^+\rightarrow 0_1^+)}{B(E2;2_1^+\rightarrow 0_1^+)}=0.789(43).
\]
This ratio matters because, in a vibrational picture,
\[
B(E2;2^{+}_{1} \rightarrow 0^{+}_{1}) > B(E2;2^{+}_{i} \rightarrow 0^{+}_{1}),\quad i>1.
\]
An earlier DSAM \((n,n'\gamma)\) result had implied the opposite ordering, with the \(2_2^+\) strength larger by a factor \(1.63\), whereas the electron-scattering measurement independently confirmed the revised sub-unity interpretation. The analysis used the relative population of the \(2_1^+\) and \(2_2^+\) states in \((e,e')\), a momentum-transfer range \(q=0.17\text{--}0.51\,\mathrm{fm}^{-1}\), and a PWBA-based low-\(q\) expansion in which Coulomb effects and other systematic errors largely cancel in the ratio. The fitted transition-radius difference, \(\Delta R=-0.28(42)\,\mathrm{fm}\), was consistent with zero, and use of \(B(E2;2_1^+)=4.9(11)\,\mathrm{W.u.}\) led to \(B(E2;2_2^+\rightarrow 0_1^+)=3.9(9)\,\mathrm{W.u.}\) [1401.4837].

In elastic spring networks, relative strength refers to the balance between bond compression/stretching and bond reorientation. The model assigns harmonic responses
\[
F_L=-k_L dL,\qquad F_\theta=-k_\theta L d\theta,
\]
with Poisson’s ratio defined as \(\nu\equiv -\epsilon_{\rm trans}/\epsilon_{\rm axial}\). The central analytic result is
\[
\nu=0 \quad \text{when} \quad k_L=k_\theta,
\]
and this holds independently of the network geometry in the two-dimensional model considered. The derivation imposes no net perpendicular displacement of each bond under axial loading, so the equality of stretching and reorientation stiffnesses becomes a geometry-independent sufficient condition for vanishing transverse strain. The same paper interprets \(k_\theta/k_L\ll 1\) as a regime of high tunability and notes that the closer \(k_L/k_\theta\) is to unity, the harder it is to obtain variation in \(\nu\) [2104.03198].

These two examples illustrate distinct physical uses of relative strength. In \(^{94}\)Zr it is a spectroscopic ratio between competing collective transitions; in spring networks it is the balance of two elastic channels. In both cases, however, the ratio is used to discriminate between structurally different regimes: ordinary vibrational ordering in one case, and zero-Poisson behavior in the other.

## 3. Spin-orbit, signal, and positioning ratios

In zincblende quantum wells, relative strength is the ratio between Rashba and Dresselhaus spin-orbit interactions. For a [001]-grown quantum well with total Rashba coefficient \(\alpha=\alpha_0+\alpha_{\rm ex}\), the linear SU(2)-symmetric points occur at \(\alpha=\pm\beta\). Sweeping a gate voltage produces two critical voltages \(U_{s+}\) and \(U_{s-}\), and their asymmetry directly encodes the intrinsic structure inversion asymmetry. In the linear model,
\[
\frac{\alpha_0}{\beta}=\frac{U_{s+}+U_{s-}}{U_{s-}-U_{s+}},
\]
while with cubic Dresselhaus corrections the extracted quantity becomes
\[
\frac{\alpha_0}{\beta^\ast},\qquad \beta^\ast=\beta-\frac{1}{4}k_F^2\gamma.
\]
The method does not rely on directional anisotropy measurements; instead it uses the voltages at which the in-plane spin relaxation times \(\tau_+\) and \(\tau_-\) are maximal. For an asymmetrically \(n\)-doped \(\mathrm{Al}_{0.3}\mathrm{Ga}_{0.7}\mathrm{As}/\mathrm{GaAs}/\mathrm{Al}_{0.3}\mathrm{Ga}_{0.7}\mathrm{As}\) quantum well with \(N_D=4.0\times 10^{11}\,\mathrm{cm}^{-2}\), the paper reports approximately \(U_{s+}=0.4\,\mathrm{V}\) and \(U_{s-}=-0.08\,\mathrm{V}\), corresponding to \(\alpha_0/\beta^\ast\approx -0.67\) [1003.1198].

In indoor visible-light positioning, the relevant ratio is the received signal strength ratio (RSSR). The enhanced camera-assisted RSSR method first estimates incidence angles \(\psi_i\) from image geometry, then combines them with photodiode measurements according to
\[
\frac{P_{r,j}}{P_{r,i}}=\frac{\|\mathbf d_i^{\mathrm w}\|^{m+2}}{\|\mathbf d_j^{\mathrm w}\|^{m+2}}\frac{\cos(\psi_j)}{\cos(\psi_i)},
\]
so that
\[
A_{ij}=\left(\frac{P_{r,j}}{P_{r,i}}\frac{\cos(\psi_{i,\mathrm{est}})}{\cos(\psi_{j,\mathrm{est}})}\right)^{\frac{1}{m+2}}
=\frac{\|\mathbf d_i^{\mathrm w}\|}{\|\mathbf d_j^{\mathrm w}\|}.
\]
A camera-derived inter-ray angle \(\alpha_{ij}\), together with the known LED spacing, then converts relative distances into absolute distances by Euclidean geometry, after which linear least squares estimates the receiver position. The method requires only 3 LEDs for both orientation-free 2D and 3D positioning, and a compensation algorithm based on single-view geometry is introduced when the photodiode-camera offset is non-negligible. Simulation results reported centimeter-level accuracy over \(80\%\) indoor area for both small and large photodiode-camera distances [2004.06294].

Both cases use ratios to suppress nuisance structure. In the quantum-well setting, the asymmetry of the two SU(2)-restoring voltages isolates intrinsic SIA relative to Dresselhaus coupling. In visible-light positioning, ratios of received powers remove the need for direct absolute power calibration and convert signal amplitudes into geometric constraints.

## 4. Bayesian evidence and calibration

In Bayesian inference, relative strength is formalized as the relative belief ratio. The principle of evidence is stated as: if the posterior probability of an event is greater than, less than, or equal to its prior probability, then there is evidence in favor of, against, or neither for nor against that event. For an event \(A\),
\[
RB(A\mid x)=\frac{\Pi(A\mid x)}{\Pi(A)}.
\]
Hence \(RB>1\) indicates evidence in favor, \(RB<1\) evidence against, and \(RB=1\) no evidence either way. For a parameter value \(\psi\), the corresponding continuous-form expression is
\[
RB_\Psi(\psi\mid x)=\frac{\pi_\Psi(\psi\mid x)}{\pi_\Psi(\psi)},
\]
with the equivalent predictive form
\[
RB_\Psi(\psi\mid x)=\frac{m(x\mid \psi)}{m(x)}.
\]
A central claim of this literature is that the numerical size of \(RB\) or of a Bayes factor is not, by itself, a correct universal measure of evidential strength. Strength is instead calibrated by the posterior probability
\[
\mathrm{Str}_{\Psi,\psi_0}(x)=\Pi_\Psi\!\big(RB_\Psi(\psi\mid x)\le RB_\Psi(\psi_0\mid x)\mid x\big),
\]
which compares the hypothesized value’s evidential update against the posterior distribution of evidential updates for competing values [2301.08994].

This distinction is operationalized in two-arm trials. For equivalence, the paper on Bayesian hypothesis assessment discretizes the treatment difference \(\mu_E-\mu_R\) into bins of width \(2\delta\), identifies equivalence with \(H_0:\psi=0\), and computes
\[
RB_\Psi(0)=\frac{\Pi((-\delta,\delta]\mid x_E,x_R)}{\Pi((-\delta,\delta])}.
\]
For non-inferiority it uses
\[
RB_\Psi((-\delta,\infty))=\frac{\Pi((-\delta,\infty)\mid x_E,x_R)}{\Pi((-\delta,\infty))}.
\]
In the reported blood-pressure example, \(RB_\Psi(0)=0.515\) with strength \(0.19\), interpreted as moderately strong evidence against equivalence, whereas non-inferiority yields \(RB=1.53\), interpreted as evidence for non-inferiority. The same paper emphasizes that relative belief analysis must be accompanied by prior-bias assessment, model checking, and prior-data conflict checks [1401.4215].

A common misconception is therefore explicitly rejected in this literature: the ratio is the evidential direction, but the strength of that evidence is a separate posterior-calibration problem. This separation distinguishes relative belief methods from practices that read raw Bayes-factor magnitudes through fixed verbal scales [2301.08994].

## 5. Machine learning: density ratios, preference strength, and privacy

In distributional learning, relative strength is often a density-ratio. Given numerator and denominator densities \(p_{\rm nu}\) and \(p_{\rm de}\), the ordinary density-ratio is \(r(x)=p_{\rm nu}(x)/p_{\rm de}(x)\). The meta-learning work on relative DRE instead uses the bounded \(\alpha\)-relative density-ratio
\[
r_\alpha({\bf x})=\frac{p_{\rm nu}({\bf x})}{\alpha p_{\rm nu}({\bf x})+(1-\alpha)p_{\rm de}({\bf x})},
\]
with
\[
r_\alpha({\bf x})\le \frac{1}{\alpha}\qquad (\alpha>0).
\]
The model extracts dataset-level information from two few-shot support sets, constructs an instance embedding conditioned on both datasets, and fits a linear relative density-ratio model in the embedded space. Its adaptation step has a closed-form solution,
\[
\tilde{\bf w}=({\bf K}+\lambda{\bf I})^{-1}{\bf k},
\]
which enables explicit meta-optimization of post-adaptation test error. The reported applications are relative DRE, dataset comparison, and outlier detection [2107.00801].

ResponseRank addresses a different problem: pairwise preferences ordinarily encode only direction, not strength. The paper models latent utility \(u:\mathcal X\to\mathbb R\) and treats preference strength as the utility gap \(u(x)-u(y)\) or \(|u(x)-u(y)|\). Rather than introducing an explicit ratio variable, it constructs local rankings of pairwise comparisons from noisy proxy signals such as response time, inter-annotator agreement, or returns. After converting each pair into winner-loser form \(q'_i=(w_i,l_i)\), it learns
\[
s_\theta(w_i,l_i)=u_\theta(w_i)-u_\theta(l_i)
\]
with a Plackett–Luce objective over rankings that append a virtual anchor \(\lambda_0\) of score \(0\). For singleton rankings the objective reduces to Bradley–Terry. The paper states that the learned utility is cardinal up to a positive affine transformation, not a literal ratio-scaled quantity, and introduces Pearson Distance Correlation to evaluate whether predicted absolute utility gaps correlate with true ones [2512.25023].

Differential privacy introduces yet another ratio problem. For a generic count ratio \(Z=X/Y\) and for relative risk \(p_x/p_y\), the paper on differentially private ratio statistics argues that directly perturbing the ratio is poor because the sensitivity can be large, while noising the counts and then taking the ratio as post-processing works well. Its default Laplace mechanism uses
\[
\widetilde X=X+\mathrm{Lap}\!\left(\frac{2}{\varepsilon}\right),\qquad
\widetilde Y=Y+\mathrm{Lap}\!\left(\frac{2}{\varepsilon}\right),\qquad
\widetilde Z=\widetilde X/\widetilde Y,
\]
and for bias control recommends the bounded denominator
\[
\widetilde Z=\widetilde X/\max(\widetilde Y,1).
\]
For relative risk, with \(X\sim Bin(n_x,p_x)\) and \(Y\sim Bin(n_y,p_y)\), the private estimator \(\widetilde p=(\widetilde X/n_x)/(\widetilde Y/n_y)\) is proved consistent under independent zero-mean noise with variance \(O(n_x)\) and \(O(n_y)\), and asymptotically valid confidence intervals are developed. The paper’s practical conclusion is that privacy noise is most problematic in the same small-denominator regimes in which the non-private ratio is already statistically unstable [2505.20351].

These machine-learning and inference settings show three distinct uses of relative strength: bounded comparison of distributions, ordered comparison of utility margins, and privacy-preserving estimation of count ratios. Only the first and third are literal ratios; the second is explicitly rank-based.

## 6. Optimization and strategic interaction

In cutting-plane theory, relative strength is defined through approximation quality. For the two-row mixed-integer relaxation with two free integer variables and nonnegative continuous variables, the nontrivial facets are split, triangle, and quadrilateral inequalities. The paper compares the closures \(S_f^k\), \(T_f^k\), and \(Q_f^k\) using Goemans-style scaling with respect to the integer hull \(R_f^k\). Its principal quantitative statements are
\[
T_f^k\subseteq S_f^k,\qquad Q_f^k\subseteq S_f^k,
\]
and
\[
R_f^k\subseteq T_f^k\subseteq 2R_f^k,\qquad
R_f^k\subseteq Q_f^k\subseteq 2R_f^k,
\]
whereas split closure has no bounded approximation ratio:
\[
\forall \alpha>1,\ \exists(f,r^1,\ldots,r^k)\ \text{such that}\ S_f^k\not\subseteq \alpha R_f^k.
\]
In this setting, “relative strength” is not a ratio of two numbers but the inclusion or scaling relation between closures, and the conclusion is that triangle and quadrilateral cuts are qualitatively stronger than split cuts in the two-row model [1701.06536].

In finite strategic games, relative strength is likewise defined by inclusion between iterated outcomes of elimination procedures. The paper studies four operators: iterated elimination of strategies weakly dominated by a pure strategy (\(LW\)), weakly dominated by a mixed strategy (\(MLW\)), strictly dominated by a pure strategy (\(LS\)), and strictly dominated by a mixed strategy (\(MLS\)). Because the local operators are non-monotonic, the paper introduces global versions and proves equality of local and global \(\omega\)-outcomes. The resulting hierarchy is
\[
LW^\omega \subseteq LS^\omega,\qquad
MLW^\omega \subseteq MLS^\omega \subseteq LS^\omega.
\]
An important negative result is that the apparently natural inclusion \(MLW^\omega\subseteq LW^\omega\) fails. Here relative strength means that one procedure leaves a game no larger than another after full iteration, not that one-step elimination is more aggressive [0706.1617].

These optimization and game-theoretic uses differ sharply from ratio-based physics or statistics. “Strength” is comparative, but the comparison is between closures, operators, or limit outcomes rather than between scalar observables.

## 7. Institutional and astrophysical diagnostics

In bibliometrics, relative strength is a size-independent institutional indicator. The ratio proposed for universities is
\[
\mathrm{RATIO}_U=\frac{TS_U}{RS_U},
\]
where \(TS_U\) is the number of top scientists employed by university \(U\) and \(RS_U\) is the total research staff or faculty. Top scientists are those at or above the 90th percentile of national productivity within each scientific disciplinary sector, with productivity measured by the paper’s fractional scientific strength \(FSS_p\). The same logic is applied at field and discipline levels via \(\mathrm{RATIO}_{U,s}\) and \(\mathrm{RATIO}_{U,d}\). The paper interprets this ratio as an indicator of competitive strength, reports a Spearman correlation of \(0.924\) between the overall ranking by TS ratio and the ranking by average faculty productivity \(FSS_U\), and finds a negligible Pearson correlation of \(-0.129\) between TS ratio and university size [1810.12667].

In active galactic nuclei, relative strength is the luminosity of narrow-line region emission relative to broad-line region emission. Stern and Laor study the narrow H\(\alpha\)-to-broad H\(\alpha\) ratio and the \([\mathrm{O\,III}]\lambda5007\)-to-broad H\(\alpha\) ratio. Their main empirical relations are
\[
L_{\mathrm{nH\alpha}}\propto L_{\mathrm{bH\alpha}}^{0.7},
\qquad
L_{\mathrm{[OIII]}}\propto L_{\mathrm{bH\alpha}}^{0.7},
\]
so the narrow-to-broad ratios decline with AGN luminosity. They derive
\[
\frac{L_{\mathrm{bol}}}{10^{43}\,\mathrm{erg\,s^{-1}}}
=
4000\left(\frac{L_{\mathrm{[OIII]}}}{10^{43}\,\mathrm{erg\,s^{-1}}}\right)^{1.39},
\]
which implies that \(L_{\mathrm{bol}}/L_{\mathrm{[OIII]}}\) decreases from \(\sim 3000\) at \(L_{\mathrm{bol}}=10^{46.1}\,\mathrm{erg\,s^{-1}}\) to \(\sim 300\) at \(L_{\mathrm{bol}}=10^{42.5}\,\mathrm{erg\,s^{-1}}\). The favored interpretation is a decrease in the narrow-line region covering factor with increasing luminosity, from \(\sim 0.4\) at \(10^{42.5}\,\mathrm{erg\,s^{-1}}\) to \(\sim 0.04\) at \(10^{45.5}\,\mathrm{erg\,s^{-1}}\), with low-luminosity type 1 AGN therefore often appearing as intermediate types because the narrow component dominates the observed H\(\alpha\) profile [1207.5543].

These examples show that relative-strength measures can function as diagnostics of composition rather than of absolute scale. In the university case, the ratio distinguishes concentration of elite researchers from raw institutional size. In the AGN case, declining narrow-to-broad line strength reveals a changing balance between emitting regions and motivates a covering-factor interpretation.

Source: https://www.emergentmind.com/topics/relative-strength-ratio