CERN Accelerating science

Production of a dijet resonance, A, in a proton-proton collision. The A resonance decays to two resonances B and C, which in turn each decay to a jet with anomalous substructure arising from multiple subjets.

The $p$-values as a function of the injected signal cross sections for the different analysis procedures for two different signals: (upper) the 2-prong \XtoYY signal with $m_\mathrm{X}=3\TeV$, $m_\mathrm{Y}=170\GeV$, and $M_\mathrm{Y'}=170\GeV$, and (lower) 3-prong \Wp signal with $M_\mathrm{W'}=3\TeV$ and $M_\mathrm{B'}=400\GeV$. Significance values larger than 7$\sigma$ are denoted with downwards facing triangles.

The $p$-values as a function of the injected signal cross sections for the different analysis procedures for two different signals: (upper) the 2-prong \XtoYY signal with $m_\mathrm{X}=3\TeV$, $m_\mathrm{Y}=170\GeV$, and $M_\mathrm{Y'}=170\GeV$, and (lower) 3-prong \Wp signal with $M_\mathrm{W'}=3\TeV$ and $M_\mathrm{B'}=400\GeV$. Significance values larger than 7$\sigma$ are denoted with downwards facing triangles.

The dijet invariant mass spectrum and resulting background fit to the data for \vae~(upper left), \cwola~(upper right), \TNT~(middle left), \cathode~(middle right), \cathodeb~(lower left), and \quak~(lower right). The shapes of two benchmark signals are shown for the \vae method; the signal shapes for the other methods are similar. For all methods besides the \vae, separate selections are applied for different signal mass hypotheses and the resulting mass spectra are fit separately. The figures show the fitted and observed dijet mass distribution in the signal window of each selection, which results in a discontinuous distribution. The spectra in the $\alpha$ signal regions (indicated by the vertical dotted lines) are shown for the weakly supervised methods and a similar selection of signal regions are shown for the \quak method.

The dijet invariant mass spectrum and resulting background fit to the data for \vae~(upper left), \cwola~(upper right), \TNT~(middle left), \cathode~(middle right), \cathodeb~(lower left), and \quak~(lower right). The shapes of two benchmark signals are shown for the \vae method; the signal shapes for the other methods are similar. For all methods besides the \vae, separate selections are applied for different signal mass hypotheses and the resulting mass spectra are fit separately. The figures show the fitted and observed dijet mass distribution in the signal window of each selection, which results in a discontinuous distribution. The spectra in the $\alpha$ signal regions (indicated by the vertical dotted lines) are shown for the weakly supervised methods and a similar selection of signal regions are shown for the \quak method.

The dijet invariant mass spectrum and resulting background fit to the data for \vae~(upper left), \cwola~(upper right), \TNT~(middle left), \cathode~(middle right), \cathodeb~(lower left), and \quak~(lower right). The shapes of two benchmark signals are shown for the \vae method; the signal shapes for the other methods are similar. For all methods besides the \vae, separate selections are applied for different signal mass hypotheses and the resulting mass spectra are fit separately. The figures show the fitted and observed dijet mass distribution in the signal window of each selection, which results in a discontinuous distribution. The spectra in the $\alpha$ signal regions (indicated by the vertical dotted lines) are shown for the weakly supervised methods and a similar selection of signal regions are shown for the \quak method.

The dijet invariant mass spectrum and resulting background fit to the data for \vae~(upper left), \cwola~(upper right), \TNT~(middle left), \cathode~(middle right), \cathodeb~(lower left), and \quak~(lower right). The shapes of two benchmark signals are shown for the \vae method; the signal shapes for the other methods are similar. For all methods besides the \vae, separate selections are applied for different signal mass hypotheses and the resulting mass spectra are fit separately. The figures show the fitted and observed dijet mass distribution in the signal window of each selection, which results in a discontinuous distribution. The spectra in the $\alpha$ signal regions (indicated by the vertical dotted lines) are shown for the weakly supervised methods and a similar selection of signal regions are shown for the \quak method.

The dijet invariant mass spectrum and resulting background fit to the data for \vae~(upper left), \cwola~(upper right), \TNT~(middle left), \cathode~(middle right), \cathodeb~(lower left), and \quak~(lower right). The shapes of two benchmark signals are shown for the \vae method; the signal shapes for the other methods are similar. For all methods besides the \vae, separate selections are applied for different signal mass hypotheses and the resulting mass spectra are fit separately. The figures show the fitted and observed dijet mass distribution in the signal window of each selection, which results in a discontinuous distribution. The spectra in the $\alpha$ signal regions (indicated by the vertical dotted lines) are shown for the weakly supervised methods and a similar selection of signal regions are shown for the \quak method.

The dijet invariant mass spectrum and resulting background fit to the data for \vae~(upper left), \cwola~(upper right), \TNT~(middle left), \cathode~(middle right), \cathodeb~(lower left), and \quak~(lower right). The shapes of two benchmark signals are shown for the \vae method; the signal shapes for the other methods are similar. For all methods besides the \vae, separate selections are applied for different signal mass hypotheses and the resulting mass spectra are fit separately. The figures show the fitted and observed dijet mass distribution in the signal window of each selection, which results in a discontinuous distribution. The spectra in the $\alpha$ signal regions (indicated by the vertical dotted lines) are shown for the weakly supervised methods and a similar selection of signal regions are shown for the \quak method.

The dijet invariant mass spectrum and resulting background fit to the data for the $\beta$ signal regions (indicated by the vertical dotted lines) of \cwola~(upper), \TNT~(middle left), \cathode~(middle right), \cathodeb~(lower left), and a similar selection of signal regions of \quak~(lower right). Separate selections are applied for different signal mass hypotheses and the resulting mass spectra are fit separately. The figures therefore show the fitted and observed dijet mass distribution in the signal window of each selection, which results in a discontinuous distribution. The \cathode and \cathodeb methods are not used in the highest mass window of the $\beta$ signal regions due to the limited number of data events. They therefore have one fewer signal region shown than the other methods.

The dijet invariant mass spectrum and resulting background fit to the data for the $\beta$ signal regions (indicated by the vertical dotted lines) of \cwola~(upper), \TNT~(middle left), \cathode~(middle right), \cathodeb~(lower left), and a similar selection of signal regions of \quak~(lower right). Separate selections are applied for different signal mass hypotheses and the resulting mass spectra are fit separately. The figures therefore show the fitted and observed dijet mass distribution in the signal window of each selection, which results in a discontinuous distribution. The \cathode and \cathodeb methods are not used in the highest mass window of the $\beta$ signal regions due to the limited number of data events. They therefore have one fewer signal region shown than the other methods.

The dijet invariant mass spectrum and resulting background fit to the data for the $\beta$ signal regions (indicated by the vertical dotted lines) of \cwola~(upper), \TNT~(middle left), \cathode~(middle right), \cathodeb~(lower left), and a similar selection of signal regions of \quak~(lower right). Separate selections are applied for different signal mass hypotheses and the resulting mass spectra are fit separately. The figures therefore show the fitted and observed dijet mass distribution in the signal window of each selection, which results in a discontinuous distribution. The \cathode and \cathodeb methods are not used in the highest mass window of the $\beta$ signal regions due to the limited number of data events. They therefore have one fewer signal region shown than the other methods.

The dijet invariant mass spectrum and resulting background fit to the data for the $\beta$ signal regions (indicated by the vertical dotted lines) of \cwola~(upper), \TNT~(middle left), \cathode~(middle right), \cathodeb~(lower left), and a similar selection of signal regions of \quak~(lower right). Separate selections are applied for different signal mass hypotheses and the resulting mass spectra are fit separately. The figures therefore show the fitted and observed dijet mass distribution in the signal window of each selection, which results in a discontinuous distribution. The \cathode and \cathodeb methods are not used in the highest mass window of the $\beta$ signal regions due to the limited number of data events. They therefore have one fewer signal region shown than the other methods.

The dijet invariant mass spectrum and resulting background fit to the data for the $\beta$ signal regions (indicated by the vertical dotted lines) of \cwola~(upper), \TNT~(middle left), \cathode~(middle right), \cathodeb~(lower left), and a similar selection of signal regions of \quak~(lower right). Separate selections are applied for different signal mass hypotheses and the resulting mass spectra are fit separately. The figures therefore show the fitted and observed dijet mass distribution in the signal window of each selection, which results in a discontinuous distribution. The \cathode and \cathodeb methods are not used in the highest mass window of the $\beta$ signal regions due to the limited number of data events. They therefore have one fewer signal region shown than the other methods.

The discovery sensitivity for the process $\mathrm{A} \to \mathrm{BC}$, using the anomaly detection methods and a comparison to sensitivity of the inclusive search. In all signal processes, the mass of the heavy resonance is set to $m_\mathrm{A}=3\TeV$. For the BSM daughter particles, the masses of the \PY and \PYpr are set to 170\GeV, while the masses of the \PBpr, $\mathrm{R}$, and \PH are set to 400\GeV. In the upper panel, for each method, the cross section, which would have led to an expected $3\sigma$ ($5\sigma$) excess, is shown as a cross (square) marker. Sensitivities from six anomaly detection methods (six colors) are compared to an inclusive dijet search in which no substructure selection is made (black) and traditional substructure selections targeting 2-prong (dark brown) or 3-prong (tan) decays. The expected 95\% confidence level upper limits from the inclusive search are also shown in the upper panel as a dashed line. For all signal models at least one anomaly detection method is able to achieve an expected $5\sigma$ significance at a cross section at or below the upper limit of the inclusive search. Shown in the lower panel is the ratio of the cross section sensitivity from the inclusive search to the corresponding sensitivity for each method.

The upper limit at 95\% confidence level on the cross section for the process $\mathrm{A} \to \mathrm{BC}$, is shown for each search method applied to a variety of signal models. For a resonance mass $m_\mathrm{A}=3\TeV$ (upper) and $m_\mathrm{A}=5\TeV$ (lower), we show for each signal model (columns), and search method (all colors), the observed limits (crosses), expected limits (squares), and their 68\% expected central intervals (error bars). For the BSM daughter particles, the masses of the \PY and \PYpr are set to 170\GeV, while the masses of the \PBpr, $\mathrm{R}$, and \PH are set to 400\GeV. Limits from the anomaly detection methods (six colors) are compared to those from an inclusive dijet search in which no substructure selection is made (black markers and horizontal lines), traditional substructure selections targeting 2-prong (dark brown) or 3-prong decays (tan), and the observed limit from a previous CMS search \cite{CMS_triboson_hadronic} for the \Wkk model in the all-hadronic channel (gray).

The upper limit at 95\% confidence level on the cross section for the process $\mathrm{A} \to \mathrm{BC}$, is shown for each search method applied to a variety of signal models. For a resonance mass $m_\mathrm{A}=3\TeV$ (upper) and $m_\mathrm{A}=5\TeV$ (lower), we show for each signal model (columns), and search method (all colors), the observed limits (crosses), expected limits (squares), and their 68\% expected central intervals (error bars). For the BSM daughter particles, the masses of the \PY and \PYpr are set to 170\GeV, while the masses of the \PBpr, $\mathrm{R}$, and \PH are set to 400\GeV. Limits from the anomaly detection methods (six colors) are compared to those from an inclusive dijet search in which no substructure selection is made (black markers and horizontal lines), traditional substructure selections targeting 2-prong (dark brown) or 3-prong decays (tan), and the observed limit from a previous CMS search \cite{CMS_triboson_hadronic} for the \Wkk model in the all-hadronic channel (gray).

A flowchart outlining how the samples for the weakly supervised training were constructed in the \TNT method. The two jets in the dijet candidate were randomly assigned labels J1 and J2. For each event, the J1 (J2) jet is placed into either a signal-like or background-like sample based on the autoencoder scores evaluated on the J2 (J1) jet. The samples of signal-like (background-like) J1's and signal-like (background-like) J2's were merged together to construct a single sample of signal-like (background-like) jets. The \TNT classifier is then trained to distinguish between these two samples.

Diagram of the limit-setting procedure for the \XtoYY signal at 3\TeV with the \cathode method. The upper panel shows the estimated signal acceptance times efficiency as a function of the cross section injected in data. The shaded region in the upper panel shows the total statistical and systematic uncertainty in the efficiency. The resulting $N_\text{sig}(\sigma_\text{sig})$ curve and its corresponding uncertainty band from the efficiency are shown in blue in the lower panel. The expected and observed limits on the number of signal events are shown as a horizontal solid black line and green dashed lines, respectively, and connected to the corresponding limits on the cross section (vertical lines). The 68\% confidence level band around the expected limit is displayed similarly.