Optimal Kinematic & Angular Cuts

• Optimal kinematic and angular cuts are systematic criteria that use inequalities on observables like pT and angular measures to maximize signal-to-noise ratios.
• They are mathematically formulated via thresholds and optimized using methods such as Random Grid Search and Neural Gradient Optimization.
• These selection cuts balance physical realism with statistical precision, playing a key role in collider analyses, pedestrian motion studies, and cosmological measurements.

Optimal kinematic and angular selection cuts are systematic criteria imposed on measured physical observables—such as particle momenta, energy, rapidity, and angular variables—to maximize signal discrimination and background suppression, often subject to strict statistical and algorithmic optimization. These cuts are pivotal across experimental and phenomenological analyses that require statistically robust event selection in high-dimensional phase space, spanning searches at colliders, precision measurements, and pattern recognition in time-series trajectory data.

1. Mathematical Formulation of Kinematic and Angular Selection Cuts

Selection cuts are formalized as inequalities or intervals in observable space. A simple example is a lower threshold on transverse momentum, $p_T > p_{T,\min}$, or an upper bound on an angular observable, $|\Delta\phi| < \phi_{\max}$. For multi-dimensional selections, the feasible event region is defined as the intersection of accepted intervals:

$$C(x_1, \ldots, x_n) = \bigwedge_{j} \big[x_j \in [a_j, b_j]\big]$$

For processes where the physical constraint itself is non-linear or data-driven, cuts may be constructed following empirically determined boundaries such as power-law envelopes or resonance-dependent limits.

A prominent example is the universal power-law kinematic constraint on pedestrian walking, where the critical angular velocity $\omega_{c}(v)$ for speed $v$ is empirically determined as

$\omega_{c}(v) = C\,v^{-0.8}$

with $C$ calibrated to the temporal sampling interval and $\Delta t$, and the allowed sample is defined via $|\omega| \leq C v^{-0.8}$ (Wang et al., 18 Jun 2025). This approach generalizes to reject outlier or physically unfeasible events in trajectory or behavioral datasets.

In longitudinal phase-space analyses, such as photoproduction with resonance discrimination, mass-dependent sector boundaries are defined through resonance-specific kinematic extrema. For a two-body decay into particles $Y$ and $Z$ through intermediate resonance $R$, the optimal cut takes the form

$$\rho_{\text{low}}(M_R) < \arctan\left(\frac{p_Y^{z,\text{CM}}}{p_Z^{z,\text{CM}}}\right) < \rho_{\text{up}}(M_R)$$

where the endpoints $\rho_{\text{low,up}}(M_R)$ are explicitly evaluated from the resonance mass and kinematics (Pauli et al., 2018).

2. Algorithmic and Data-Driven Cut Optimization

Optimization of selection cuts involves maximizing a figure of merit—most commonly the signal significance, $Z$—over the space of cut configurations. Two widely used algorithmic techniques are:

Random Grid Search (RGS): This method samples possible cut-points from the signal distribution and evaluates the significance $Z$ for each configuration. In addition to one-sided cuts, RGS implements two-sided (interval/“box”) and staircase cuts (unions of AND-combined cuts), capturing complex, non-convex regions in observable space:

• Two-sided: $x_j^{(i)} < x_j < x_j^{(k)}$
• Staircase: $$C(x) = \bigvee_{\ell} \bigwedge_j \big[x_j > x_j^{(\ell)}\big]$$

The optimal cut-set is that which maximizes $Z$, e.g., defined in the large-sample limit as

$$Z = \sqrt{2\,b\left[(1+s/b)\ln(1+s/b) - s/b\right]}$$

where $s$ and $b$ are passing signal and background counts (Bhat et al., 2017).

Neural Gradient Optimization (CABIN): Recent methods parametrize cuts as differentiable functions (notably sigmoidal approximations to step functions) that can be optimized via gradient descent, allowing direct control over target efficiencies and smoothness across efficiency working points. Logical AND/OR operations are differentiably implemented, and regularization stabilizes the evolution of thresholds (Hance et al., 12 Feb 2025).

3. Physical Motivation and Empirical Calibration

Optimal selection cuts must balance physical realism, detector limitations, and the statistical needs of hypothesis separation.

• Physical constraints: In pedestrian or active-matter systems, universal or physics-motivated boundaries derived from empirical phase space analysis (e.g., $\omega_{c}(v)=C\,v^{-0.8}$) define feasible regions, rejecting noise/artifacts and enabling anomaly detection and model training confined to physically plausible data (Wang et al., 18 Jun 2025).
• Signal/background separation: In collider physics, cuts on $p_T$, invariant mass, angular variables ($\Delta\phi, \Delta R, \cos\theta^*$), and event topology are chosen to reflect characteristic features of the signal or distinctive background kinematics. For instance, BSM resonance searches select windows around resonance mass, high $p_T$ for boosted objects, and enforce separation in $\Delta R$ to exploit event geometry (Queiroz et al., 31 Dec 2025).
• Data cleaning and systematic correction: In heavy-ion or cosmology studies, cut placement affects the measurement of statistical observables (such as charge cumulants or anisotropies) and is carefully calibrated to correct for or minimize acceptance, edge, or finite-volume effects (Karsch et al., 2015, Hausegger et al., 2024).

4. Quantitative Examples and Benchmark Case Studies

Well-documented applications demonstrate the technical impact of optimized cut choices:

• Pedestrian datasets: In (Wang et al., 18 Jun 2025), robust filtering using $|\omega| \leq 1.0 v^{-0.8}$ (for 30 Hz sampling) or $|\omega| \leq 0.6 v^{-0.8}$ (for 3 Hz decimated data) across diverse datasets supports universal data cleaning, outlier flagging, and dynamic constraint imposition in robotics.
• Collider resonance search (FCC-hh): For heavy $Z'$ boson searches, optimal cuts include hard single-lepton $p_T$ ($p_T > 0.42 M_{Z'}$), tight pseudorapidity ($|\eta| < 4$), and narrow mass window selection ($m_{\ell\ell} \in [M_{Z'} \pm 1-2 \Gamma_{Z'}]$), with efficiency and background rejection explicitly tabulated for each benchmark (see below) (Queiroz et al., 31 Dec 2025).

$M_{Z'}$ (TeV)	$p_T^{\min}$ (GeV)	$|\cos\theta^*|^{\max}$	$m_{\ell\ell}$ window (GeV)	$\Delta R^{\min}$	Signal eff. (\%)	Background eff. (\%)
4	1694	0.51	[3838, 4240]	1.44	14.8	0.89
8	3347	0.42	[7679, 8481]	4.98	3.0	0.20
12	4781	0.48	[11859, 13183]	1.72	4.5	0.013

• Higgsino searches: Soft-dilepton plus monojet analyses at LHC14 benefit from staged, empirically tuned C1–C4 cuts. Notably, a novel angular veto on the azimuthal orientation between the two leptons and missing energy reduces the $Z\to\tau\tau + j$ irreducible background by a factor of 52 (at a cost of ∼40% signal reduction, but netting a 13× gain in S/B relative to legacy collinear-mass cuts) (Baer et al., 2021, Baer et al., 2022).

5. Statistical Robustness, Data-Driven Validation, and Acceptance Correction

• Low-statistics regimes may induce artificial significance spikes if background passing counts ($b$) are very low; constraints such as $b > 3$ are typically imposed to ensure the validity of $Z$ as significance (Bhat et al., 2017).
• Acceptance effects introduced by cuts, especially non-uniform or resonance/mass-dependent windows, must be corrected using unbinned maximum likelihood or efficiency-corrected moments via Monte Carlo normalization. This is critical in analyses extracting angular distributions or higher-order cumulants (Pauli et al., 2018, Karsch et al., 2015).
• Correlation and selection bias require explicit evaluation via pre- and post-cut correlation matrices, ROC-scans, and pairwise scatter plots, especially when optimizing for both statistical power and shape information in distributions (critical for dark-sector discrimination) (Dienes et al., 2014).

6. Systematic Effects, Power Corrections, and Advanced Strategies

Imposing fiducial or isolation cuts can induce parametrically enhanced power corrections in subtraction-based NNLO/N$^3$LO calculations, most notably for $q_T$ and $N$-jettiness slicing. Selections such as $p_{T,\min}$ and photon isolation (fixed-cone vs. smooth-cone) affect the convergence rate of power corrections:

• Fixed-cone isolation is optimal for minimizing leading-power corrections; additional induced terms are strictly suppressed for $q_T < E_T^{\rm iso}$.
• Smooth-cone isolation leads to slower convergence; careful tuning of the profile (minimizing the power $n$, maximizing $\epsilon$) is required (Ebert et al., 2019).

In advanced implementations, differential slicing and projection-to-Born techniques are recommended to exactly cancel cut-induced power corrections, enabling experimental-like selections in perturbative predictions without uncontrolled systematic error proliferation.

7. Applications Across Domains and Future Perspectives

Optimal kinematic and angular cuts are universally relevant:

• Trajectory analysis: Universal envelope filtering in agent-based pedestrian flows, robotics, and motion capture (Wang et al., 18 Jun 2025).
• Collider searches and measurements: Automated and human-guided cut optimization in discovery-oriented analyses (BSM, SUSY, dark sectors), background suppression, anomaly detection, and shape-preserving signal extraction (Bhat et al., 2017, Baer et al., 2021, Queiroz et al., 31 Dec 2025, Dienes et al., 2014).
• Statistical fluctuation measurements: Consistent cumulant and moment computation in heavy-ion and lattice QCD studies, effective volume mapping via $p_{t,\min}$ analog (Karsch et al., 2015).
• Cosmological parameter extraction: Selection strategies in redshift tomography ensure precise dipole measurements with minimal boundary-induced bias (Hausegger et al., 2024).

The continued development of algorithmically optimized and physically calibrated cut strategies, integrated with rigorous statistical and acceptance correction protocols, underpins both the sensitivity and interpretability of experimental and observational inferences in modern physics.