Semi-Merged Diphoton Systems in Higgs Decays

• Semi-merged diphoton systems are defined by an intermediate photon separation, combining a resolved and a merged photon cluster in exotic decay chains.
• Advanced ECAL clustering and shower-shape analyses, complemented by machine learning mass regression, enable accurate reconstruction of these complex photon signatures.
• These systems provide new search avenues at the LHC for light boson decays and extended resonance models while offering precise efficiency and background modeling insights.

A semi-merged diphoton system is a composite photon-like object arising from the decay of a light boson—commonly in exotic decay chains such as $H \to \mathcal{AA} \to \gamma\gamma\gamma\gamma$—where one $\mathcal{A} \to \gamma\gamma$ decay produces two resolved photon candidates, while the other yields a highly collimated photon pair reconstructed as a single merged photon cluster in the electromagnetic calorimeter (ECAL). This regime is distinguished by an opening angle, $\Delta R_{\gamma\gamma}$, intermediate between the limits set by ECAL granularity and cluster-separation criteria, specifically defined for $1 < m_{\mathcal{A}} < 15$ GeV in the context of LHC searches for new light resonances (Collaboration, 1 Jan 2026, Collaboration, 2024, Bi et al., 2015).

1. Kinematic Regimes and Definition

For a scalar $\mathcal{A}$ produced in Higgs decays, the photon pair from $\mathcal{A} \to \gamma\gamma$ acquires an opening angle $$\Delta R_{\gamma\gamma} \simeq 2 m_{\mathcal{A}} / p_T^{\mathcal{A}}$$ in the lab frame. Detailed studies by CMS show:

• $m_{\mathcal{A}} \lesssim 1$ GeV: $\Delta R_{\gamma\gamma} \lesssim 0.02$–$0.035$ (typical ECAL Molière radius); photons form a nearly indistinguishable shower overlap ("fully merged").
• $1 < m_{\mathcal{A}} < 2$ GeV: $\Delta R_{\gamma\gamma} \sim 0.05$–$0.09$; partially resolved with strong overlap.
• $m_{\mathcal{A}} \gtrsim 10$ GeV: $\Delta R_{\gamma\gamma} \gtrsim 0.15$–$0.3$; photons are fully resolved as separate clusters.

The “semi-merged” diphoton regime is operationally defined for $0.01 < \Delta R_{\gamma\gamma} < 0.20$, corresponding to intermediate photon separations. In this interval, one $A \to \gamma\gamma$ decay is reconstructed as two resolved photons ("resolved leg") and the other as a merged photon-like object ("merged leg"). For high-mass parents ($m_X$ in $X \to \phi\phi \to 4\gamma$), semi-merged diphotons occur for $\alpha \equiv m_\phi / m_X = 0.5\%$–$2.5\%$ (Collaboration, 2024, Bi et al., 2015).

2. ECAL Clustering and Photon Identification

Photon clustering in CMS ECAL follows a seed-based algorithm:

• Seed crystals are required to have $E_T > 230$ MeV.
• Basic clusters aggregate energy from adjacent crystals within local $$\Delta\eta \oplus \Delta\phi \lesssim 0.05 \oplus 0.15$$ for superclusters (for $E_T > 10$ GeV).
• For merged configurations, two collimated photon showers coalesce into an extended energy distribution, which standard PF (particle-flow) algorithms may not split.

Discrimination between single photons, merged diphotons, and hadronic backgrounds employs several shower-shape variables:

• $R9 = E_{3\times3} / E_{5\times5}$
• $$\sigma_{i\eta i\eta} = \sqrt{\sum_i w_i (\eta_i - \bar{\eta})^2}$$
• $H/E$ (ratio of hadronic to electromagnetic energy)
• Charged-hadron isolation, $I_\text{ch}$, and electron veto

Merged photon-like PF candidates that satisfy loose photon ID (tight $R9$, $\sigma_{i\eta i\eta}$, $H/E < 0.05$) but are not split into two PF photons are classified as the merged leg in semi-merged event selections (Collaboration, 1 Jan 2026).

3. Event Selection and Categorization

Selection of semi-merged diphoton topologies in dedicated searches (e.g., $H \to \mathcal{AA} \to 4\gamma$) is performed as follows:

• Trigger requires diphoton events with $m_{\gamma\gamma} > 55$ GeV, leading $p_T > 30$ GeV, subleading $p_T > 18$ GeV; all photons must satisfy $|\eta| < 1.44$ (restricted to ECAL barrel).
• Offline, exactly three photon-like PF candidates must pass preselection ($H/E < 0.05$, $R9 > 0.8$, $\sigma_{i\eta i\eta} < 0.0107$, $I_\text{ch} < 2$ GeV, no pixel seed).
• Triphoton invariant mass $90 < m_{\gamma\gamma\gamma} < 180$ GeV.
• Resolved $A_2$ candidate comprises the closest photon pair; merged $A_1$ leg's mass is reconstructed via machine learning regression.

In extended resonance searches ($X \to \phi\phi$), semi-merged objects are classified by CNNs as either diphoton, single photon, or hadron, using a normalized $30 \times 30$ ECAL energy image. Mass regression CNNs, taking crystal energy images and candidate $\eta$, predict $(m/E)_\text{pred}$ to determine the cluster mass (Collaboration, 2024).

4. Mass Reconstruction and Machine Learning Techniques

Mass reconstruction for merged diphoton objects employs dedicated machine learning models:

• In $H \to \mathcal{AA}$ searches, a graph neural network (GNN) regresses $m_\text{reg}$ of the merged leg, producing linear response ($\langle m_\text{reg} \rangle \approx m_\text{true}$) across $0$–$18$ GeV with typical resolution $\sigma(m_\text{reg}) \simeq 0.5$–$1.5$ GeV and scale uncertainty $1.6$–$6\%$.
• For $X \to \phi\phi$, the mass regression CNN outputs $(m/E)_\text{pred}$; cluster mass is $$m_\Gamma = E_\text{cluster} \times (m/E)_\text{pred}$$. Resolution matches simulation within $23\%$; energy-scale uncertainty per cluster is $0.5\%$.

Selection efficiency for semi-merged topologies rises with $m_{\mathcal{A}}$—$\epsilon(m_{\mathcal{A}}) \approx 2.8$–$5.6\%$ for $m_{\mathcal{A}} = 1,3,5,10,15$ GeV—and the CNN classifier achieves $\sim$55% efficiency for true merged $\gamma\gamma$ (Collaboration, 1 Jan 2026, Collaboration, 2024).

5. Backgrounds and Statistical Modeling

Dominant backgrounds include:

• QCD multijet events with jets faking photons (“$\gamma$+jets”, “jet+jet”)
• Prompt $\gamma\gamma$+jet production

Background shapes are extracted from multiple sideband regions in the two-dimensional plane of merged vs. resolved masses, $(m_1, m_2)$. Empirical functions (dijet-like, modified dijet, diphoton, power×exp, four-parameter) are fit to the invariant mass spectra in binned $\alpha$ categories, with discrete profiling and floating nuisance parameters (Collaboration, 2024).

Validation compares predicted 2D backgrounds to data in sideband regions, with residuals fit by Chebyshev polynomials to assign shape uncertainties. All systematic sources (luminosity, trigger, ID efficiency, energy scale/resolution, ML calibration, background function choice) are treated as nuisance parameters in final profile-likelihood fits (Collaboration, 1 Jan 2026, Collaboration, 2024).

6. Theoretical Interpretations and Model Significance

The semi-merged diphoton signature naturally arises in models with new light scalars coupling to the Higgs or heavy sector. One illustrative example is the $H' + s$ extension of the Standard Model (Bi et al., 2015), with $$q \bar{q} \to H_0' (750~\text{GeV}) \to ss \to (\gamma\gamma)(\gamma\gamma)$$ and $m_s < 2.6$ GeV. Highly boosted $s$ yields photon pairs with $\Delta R_{\gamma\gamma} \simeq 2 m_s / 375$; for $m_s \lesssim 3.4$ GeV, the pair is fully merged in the ECAL, satisfying $\Delta R_{\gamma\gamma} < 0.018$.

Experimental constraints—dijet cross section bounds, photon-jet searches, ECAL granularity—are respected by restricting the relevant parameter space; prediction efficiency factors $\epsilon_\text{ID} \simeq 0.6$–$0.9$ and production rates $\sigma_\text{eff} \simeq 3$–$13$ fb are demonstrated to explain observed anomalies without contradicting negative searches in the broader parameter space.

7. Experimental Limits and Outlook

Dedicated analyses at CMS have set stringent limits on the cross section times branching ratio: $$\sigma(pp \to H)\mathcal{B}(H \to \mathcal{AA} \to 4\gamma)$$ is constrained to $0.264$–$0.005$ pb at $95\%$ CL for $m_{\mathcal{A}} = 1$–$15$ GeV—the strongest bounds to date in the $1$–$5$ GeV regime (Collaboration, 1 Jan 2026). Analogous searches for $X \to \phi\phi \to 4\gamma$ set 95% CL bounds from $0.03$ to $1.06$ fb for $m_X = 300$–$3000$ GeV and $\alpha = 0.5\%$–$2.5\%$ (Collaboration, 2024). No significant excess has been observed; future searches will benefit from enhanced ECAL granularity, improved machine learning mass regression, and further event topology exploitation, which will extend sensitivity into new regimes of collimated photon emission.