---
title: 'TriTR: A Polysemous Research Label'
url: https://www.emergentmind.com/topics/tritr
type: topic
---

# TriTR: A Polysemous Research Label

Searching arXiv for “TriTR” to verify the term’s usage across fields.
TriTR is not a single standardized term in the arXiv literature. The label appears in several technically unrelated contexts, including edge-local differential privacy for subgraph counting, frequency-bin photonic quantum gates, Jefferson Lab tritium target engineering, four-dimensional tensor-network renormalization, high-spin organic triradicals, and LHCb VELO tracking [2507.06508] [1712.03992] [2304.09998] [2412.14104] [2103.12713] [2207.03936]. In each case, the same string denotes a distinct object: an algorithm, an optical gate, a target system, a tensor-network prescription, a molecule, or a reconstruction procedure. The term is therefore best treated encyclopedically as a polysemous research label rather than as a unitary concept.

## 1. Disambiguation and domain-specific meanings

A common source of confusion is the assumption that TriTR denotes a unique framework. The available arXiv usage instead shows six distinct meanings, each anchored in a specific disciplinary vocabulary and problem setting.

| Usage of “TriTR” | Research area | Meaning in context |
|---|---|---|
| TriTR | Differential privacy on graphs | Two-round triangle counting algorithm |
| TriTR | Photonic quantum information | Balanced \(3\times 3\) frequency tritter |
| TriTR | Jefferson Lab Hall A | TRitium TRansfer sealed-cell target system |
| TriTR | Tensor renormalization | Triad representation shorthand in 4D ATRG |
| TriTR | Organic open-shell chemistry | Triradical 3 |
| TriTR | High-energy detector software | “Search by triplet” tracking algorithm |

These usages share little beyond the mnemonic prominence of “tri-”, which refers variously to triangles, tritters, tritium, triads, triradicals, or triplets. This suggests that TriTR functions primarily as a local acronym or shorthand whose interpretation must be fixed by field and citation context.

## 2. TriTR in edge-local differential privacy: two-round triangle counting

In graph analytics, TriTR is a two-round algorithm for triangle counting under edge-local differential privacy within the Noisy Adjacency Matrix framework [2507.06508]. Each user \(u\) holds an adjacency row \(a_u\in\{0,1\}^n\), applies an \(\epsilon_1\)-edge-LDP randomizer, zeros out entries with index \(\ge u\) to avoid duplication, and sends a perturbed half-row \(\tilde a_u\) to the server. The server symmetrizes and debiases the resulting matrix. Under Warner’s randomized response, each off-diagonal entry is reconstructed as
\[
\hat a_{ij}=\frac{\tilde a_{ij}(e^{\epsilon_1}+1)-1}{e^{\epsilon_1}-1},
\]
while under Laplace perturbation one treats \(\tilde a_{ij}=a_{ij}+Lap(1/\epsilon_1)\). The resulting noisy adjacency matrix \(\hat A\) satisfies three stated properties: \(E[\hat A]=A\), \(\hat A\) is symmetric with zeros on the diagonal, and all off-diagonal entries \(\hat a_{ij}\) are independent.

The second round uses the released \(\hat A\) to accumulate local triangle contributions. For each user,
\[
sum_u=\sum_{i<j} a_{ui}a_{uj}\,\hat a_{ij},
\]
so that \(E[sum_u]=2 f_u^{\triangle}\), where \(f_u^{\triangle}\) is the number of triangles containing \(u\). A second randomizer, exemplified by Laplace noise after clamping to \([-\Delta f_u,+\Delta f_u]\), produces \(\hat T_u\), and the server returns
\[
\hat f^{\triangle}(G)=\frac{1}{6}\sum_{u=1}^n \hat T_u.
\]
By construction, GNAM is \(\epsilon_1\)-edge-LDP, the second round is \(\epsilon_2\)-edge-LDP, and sequential composition yields \((\epsilon_1+\epsilon_2)\)-edge-LDP.

The paper emphasizes both statistical efficiency and explicit communication trade-offs. TriTR requires each node to download the entire \(\hat A\), with download cost \(O(n^2)\) floats, user time \(O(n+d_{\max}^2)\), and server time \(O(n^2)\). Its theoretical mean-squared error is bounded by
\[
\mathrm{MSE}\le O\Bigl(\frac{n\,d_{\max}^3}{\epsilon_1^2}+\frac{|E|}{\epsilon_1^2\epsilon_2^2}+\frac{n\,d_{\max}^2}{\epsilon_2^2}\Bigr),
\]
where the core unbiased estimator has dominating term
\[
\mathrm{MSE}_1=\frac{1}{9}\sigma^2\sum_{i<j} b_{ij}^2 \le O\bigl(n\,d_{\max}^3/\epsilon_1^2\bigr),
\]
with \(b_{ij}=(A^2)_{ij}\). Relative to TriOR and TriMTR, TriTR accepts higher download cost in exchange for the best accuracy in the reported study.

Empirically, the reported relative errors on Facebook and CA-AstroPH are \(1.85\times 10^{-2}\) and \(3.03\times 10^{-2}\) at total \(\epsilon=1.0\), improving to \(7.82\times 10^{-3}\) and \(1.44\times 10^{-2}\) at \(\epsilon=2.0\). On Facebook, TriTR runs in approximately \(5.6\,\mathrm{s}\) versus \(78\,\mathrm{s}\) for \(RR_{\triangle}\) over 20 trials. In this literature, “TriTR” therefore denotes a privacy-preserving estimator that combines an unbiased noisy matrix release with a second, local triangle-aggregation round.

## 3. TriTR as a frequency-bin photonic tritter

In photonic quantum information processing, TriTR denotes the balanced \(3\times 3\) frequency tritter realized by electro-optic modulation and Fourier-transform pulse shaping [1712.03992]. The ideal operation is the 3-point discrete Fourier transform,
\[
U_{3\times3}=\frac{1}{\sqrt3}
\begin{pmatrix}
1 & 1 & 1 \\
1 & e^{2\pi i/3} & e^{4\pi i/3} \\
1 & e^{4\pi i/3} & e^{2\pi i/3}
\end{pmatrix},
\]
which evenly redistributes amplitude from any one input frequency bin into all three output bins with the appropriate Fourier-phase factors. In the paper’s framing, it is the \(3\times 3\) extension of the Hadamard gate and implements a balanced mixing of three paths with uniform amplitudes.

The experimental architecture uses two identical electro-optic modulators before and after a line-by-line pulse shaper. Each EOM is driven by the sum of two pure sinewaves at the fundamental \(\Delta\omega\) and its second harmonic \(2\Delta\omega\). For the reported tritter, \(\Delta\omega/2\pi=18.1\,\mathrm{GHz}\), so the tones are at \(18.1\,\mathrm{GHz}\) and \(36.2\,\mathrm{GHz}\). The physical transformation is modeled as
\[
V = F\,D_3\,F^\dagger\,D_2\,F\,D_1\,F^\dagger,
\]
with \(F\) a discrete Fourier transform on \(M\gg 3\) modes, \(D_1,D_3\) diagonal in time, and \(D_2\) diagonal in frequency. Numerical optimization constrains each EOM phase to a two-term Fourier series and the pulse shaper to arbitrary per-line phase, then maximizes success probability \(\mathcal P\) subject to \(\mathcal F\ge 0.9999\). The reported solution uses time-shifted, two-tone phase functions with amplitude \(\le 0.7\pi\) rad peak and a spectral phase on about 16 central modes.

Performance is characterized by success probability and fidelity in the truncated three-mode subspace:
\[
\mathcal P=\frac{1}{3}\sum_{m,n=0}^2 |V_{mn}|^2,
\qquad
\mathcal F=\frac{1}{9\mathcal P}\Bigl|\sum_{m,n=0}^2 V_{mn}U_{mn}^*\Bigr|^2.
\]
Experimental reconstruction of \(V_{mn}\) combines single-tone power measurements for \(|V_{mn}|\) with coherent superposition inputs for relative phases. Averaging five independent reconstructions yields \(\mathcal P=0.9730\pm 0.0002\) and \(\mathcal F=0.9989\pm 0.0004\). The device is described as the first ever demonstrated for frequency modes.

The same paper stresses translation invariance across frequency bins. Because the EOM-plus-shaper design is invariant under overall frequency shift, the C-band can be partitioned into many interleaved triplets and processed in parallel. In the two-mode case, up to 33 parallel gates spaced four frequency modes apart were shown with no observable degradation in fidelity; the same principle is said to apply to the tritter. Numerical studies in Appendix B further indicate that by adding one more microwave harmonic per additional mode, the same three-element architecture can realize balanced DFT gates up to \(d=7\) with \(\mathcal F\times\mathcal P>0.97\). In this domain, TriTR is thus a high-fidelity qutrit gate in the frequency basis.

## 4. TriTR as the Jefferson Lab TRitium TRansfer target system

In nuclear and hadronic structure measurements at Jefferson Lab, TriTR denotes the sealed-cell gas-target system developed for the 2018 Hall A program [2304.09998]. Its central goal was to provide well-characterized, low-pressure tritium and \(^3\mathrm{He}\) targets of sufficient areal density, approximately \(50\text{–}150\,\mathrm{mg/cm^2}\), for inclusive and coincidence \((e,e'p)\) measurements. The scientific motivation was to exploit the isospin symmetry of the mirror nuclei \(^3\mathrm{H}\) and \(^3\mathrm{He}\) to extract neutron structure functions, form factors, and short-range correlation observables with minimal nuclear-model dependence.

The target assembly mounts a five-position, motor-actuated ladder inside the standard Hall A scattering chamber. Four sealed, low-pressure gas cells \((\mathrm{H}_2,\mathrm{D}_2,^3\mathrm{H},^3\mathrm{He})\) occupy the upper positions, an empty dummy cell sits below for background measurements, and a solid-foil ladder occupies the lowest slots for optics and absolute cross-section normalization. Each cell is built from three OEM-machined aluminum pieces of 7075-T851: a top-hat entrance-window cap, a tubular main body of approximately \(25\,\mathrm{cm}\) length and \(1.27\,\mathrm{cm}\) inner diameter, and a stainless-steel fill tube with tritium-rated valve. Entrance and exit windows are \(0.25\,\mathrm{mm}\) thick, side walls are \(0.50\,\mathrm{mm}\), and aluminum-to-aluminum joints use 1100-O annular seals with \(12.3\,\mathrm{kN}\)-rated fasteners.

The design fill pressure is \(3.44\,\mathrm{MPa}\) \((\approx 34.4\,\mathrm{bar})\) at room temperature \((\sim 293\,\mathrm{K})\), corresponding to approximately \(32\,\mathrm{cm^3}\) of gas in each \(25\,\mathrm{cm}\) cell. During beam operations, the cell bodies are cryogenically cooled to approximately \(40\,\mathrm{K}\) to limit thermal stress from a \(22.5\,\mu\mathrm{A}\), \(2\times 2\,\mathrm{mm^2}\)-rastered electron beam. Using
\[
n=\frac{P}{k_B T},
\]
the ideal-gas number density at \(34.4\,\mathrm{bar}\) and \(293\,\mathrm{K}\) is \(n\approx 8.5\times 10^{26}\,\mathrm{m^{-3}}\). The reported areal densities are \(70.8\,\mathrm{mg/cm^2}\) for \(\mathrm{H}_2\), \(142.2\,\mathrm{mg/cm^2}\) for \(\mathrm{D}_2\), \(84.8\,\mathrm{mg/cm^2}\) for \(^3\mathrm{H}\), and \(53.2\,\mathrm{mg/cm^2}\) for \(^3\mathrm{He}\), corresponding to approximately \(4\text{–}8\times 10^{-2}\,\mathrm{g/cm^2}\) of liquid-equivalent thickness.

Radiological control is a defining feature. Tritium cells were assembled and leak-checked to \(10^{-9}\,\mathrm{mbar\cdot L/s}\) helium-equivalent at Savannah River Tritium Enterprises under inert atmosphere, then transported and installed as sealed units, with no active tritium plumbing at Jefferson Lab. The scattering chamber maintains \(10^{-6}\,\mathrm{mbar}\) base pressure, the entire target stack resides within a double-walled chamber continuously purged with dry nitrogen, and tritium permeation barriers, vacuum interlocks, and redundant gas alarms assure that even in a breach scenario any release remains below \(10^{-7}\,\mathrm{Ci}\). Rigorous pre-fill bakeout reduced \(\mathrm{H_2O}\) to \(<0.1\,\mathrm{ppm}\), limiting isotopic exchange.

Operationally, beam heating induced a roughly linear density reduction \(\Delta\rho/\rho\simeq \alpha I\). The reported coefficients at \(22.5\,\mu\mathrm{A}\) are \(\alpha_T\approx -0.00431\,\mu\mathrm{A}^{-1}\), corresponding to a \(-9.7\%\) density drop for \(^3\mathrm{H}\), \(\alpha_D\approx -0.0040\,\mu\mathrm{A}^{-1}\), and \(\alpha_{He}\approx -0.0028\,\mu\mathrm{A}^{-1}\). Tritium \(\beta\)-decay with \(t_{1/2}=12.3\,\mathrm{y}\) increases \(^3\mathrm{He}\) contamination by \(5.5\%/\mathrm{yr}\); the first cell showed a \(4\%\) average T loss over six months, and the second cell exhibited \(4.1\pm 0.2\%\) \(\mathrm{H}_2\) contamination from HT–\(\mathrm{H}_2\) exchange. After corrections using simultaneous \(^3\mathrm{He}\) and separate \(\mathrm{H}_2\) runs, the combined target-thickness uncertainty was reported as \(\lesssim 1.1\%\).

The system enabled several measurements. MARATHON used the \(^3\mathrm{He}/^3\mathrm{H}\) DIS cross-section ratio and the super-ratio
\[
\mathcal R \equiv \frac{2F_2^p+F_2^n}{F_2^p+2F_2^n}
\]
to extract
\[
F_2^n/F_2^p = \frac{R_{^3He/^3H}-2\mathcal R}{\mathcal R-R_{^3He/^3H}},
\]
with total uncertainties \(\lesssim 3\%\) across \(0.3<x<0.8\). In quasi-elastic knockout, the ratio \(\sigma_{^3He(e,e'p)}/\sigma_{^3H(e,e'p)}\) at \(Q^2=2\,\mathrm{GeV^2}\) was measured from \(p_{miss}=40\text{–}500\,\mathrm{MeV/c}\) under PWIA-favored kinematics \((\theta_{rq}<37.5^\circ)\), yielding \(1.00\pm 0.02\) at low \(p_{miss}\) and approaching unity in the SRC-dominated region. Inclusive \((e,e')\) at \(x>1.4\), \(Q^2\approx 1.4\text{–}1.9\,\mathrm{GeV^2}\) gave \(\sigma_{^3H}/\sigma_{^3He}=0.854\pm 0.010\), implying \(N_{np}/N_{pp}\simeq 4.3^{+0.5}_{-0.4}\) from inclusive data and \(3.6^{+3.1}_{-1.1}\) from \((e,e'p)\). Within Hall A practice, TriTR therefore names an enabling target technology rather than an analysis code or a physics observable.

## 5. TriTR in tensor-network renormalization: triad representation for 4D ATRG

In the tensor-network literature summarized here, TriTR is used as shorthand for the triad representation applied to the four-dimensional Anisotropic Tensor Renormalization Group [2412.14104]. The starting point is the ATRG unit-cell tensor after bond swapping along the \(\hat 1\)-direction,
\[
\Gamma_{I,J}=\bigl(A\,X\,\sigma\,Y\,D\bigr)_{i_1(n+\hat1)\cdots i_4(n+\hat1)\;j_1(n)\cdots j_4(n)},
\]
with multi-indices \(I\) and \(J\). The triad representation inserts four oversampled isometries \(U^A,U^X,U^Y,U^D\), each of shape \(\chi\times\chi\times r\chi\), and approximates the unit-cell tensor by a product of four 4-leg tensors \(E,F,G,H\) and four 3-leg legs \(I,J,K,L\). For example,
\[
E_{\,i_1(n+\hat1)\,k_A\,i_4(n+\hat1)\,\alpha}
=
\sum_{i_2,i_3}
A_{\,i_1(n+\hat1)\,i_2\,i_3\,i_4(n+\hat1)\,\alpha}\;
U_{\,i_2\,i_3\,k_A}^{*\,A},
\qquad
I_{\,i_2(n+\hat1)\,i_3(n+\hat1)\,k_A}=U_{\,i_2\,i_3\,k_A}^{\,A},
\]
with analogous definitions for the remaining corners.

The isometries are obtained through truncated SVDs of reduced matrices built from the corner tensors. For \(U^A\), the prescription minimizes
\[
\min_{U^A}\;\bigl\|\Gamma-U^A U^{A\dagger}\Gamma\bigr\|,
\]
which leads to an SVD of \(AX\sigma(AX\sigma)^\dagger\). Concretely,
\[
M_{(i_2 i_3),(\bar i_2 \bar i_3)}
=
\bigl(A\,X\,\sigma\bigr)_{i_2,i_3}\,
\bigl(A\,X\,\sigma\bigr)^\dagger_{\bar i_2,\bar i_3},
\]
followed by
\[
M\simeq U^A\,(S^A)^2\,(U^A)^\dagger,
\qquad
U^A:\chi^2\to r\chi.
\]
Exactly analogous constructions produce \(U^X,U^Y,U^D\).

The reported flow for one coarse-graining step along \(\hat 1\) has five stages: bond swapping via RSVD, triad construction through local SVDs, computation of squeezers \(M^{(\mu)},N^{(\mu)}\) for \(\mu=2,3,4\), coarse-grained contraction to form new block tensors \(\Phi,\Psi\), and reshaping into standard ATRG form. The principal algorithmic effect is a contraction-cost reduction from \(O(\chi^9)\) in standard ATRG to \(O(r^2\chi^7)\) after inserting the low-rank triad decomposition. The memory footprint becomes
\[
O\bigl(4\,\chi^3(r\chi)+4\,\chi^2(r\chi)\bigr)=O(r\chi^4),
\]
compared with \(O(\chi^4)\) for ATRG.

Benchmarking on the four-dimensional Ising model is reported in terms of free-energy convergence. For oversampling \(r=7\), the maximum relative deviation between ATRG and Triad-ATRG free energies is \(\lesssim 0.0013\%\). On a single CPU, the wall-clock scaling fits are stated as approximately \(\chi^7\) for ATRG and \(\chi^9\) for Triad-ATRG, with a much smaller prefactor for the latter when \(r\ll \chi\). On two GPUs, the effective scalings are reported as \(\chi^8\) for ATRG and \(\chi^6\) for Triad-ATRG, with a \(3\text{–}5\times\) speed-up at \(\chi=54\) on two NVIDIA GPUs and end-to-end acceleration that grows from \(2\times\) at \(\chi\approx 20\) to approximately \(5\times\) at \(\chi\approx 54\) on two V100 GPUs. In this usage, TriTR denotes a tensor decomposition strategy rather than a named standalone software package.

## 6. TriTR as triradical 3 in open-shell organic chemistry

In synthetic and materials chemistry, TriTR refers to triradical 3, a neutral organic triradical with formula \(\mathrm{C_{33}H_{36}N_7O_4}\) built on a 1,2,4-benzotriazinyl radical core with two nitronyl-nitroxide radical substituents at the 3- and 7-positions [2103.12713]. Electronically, it contains three \(S=\tfrac12\) centers, one on the Blatter radical and two on the nitronyl-nitroxide substituents. Ferromagnetic exchange along two nonequivalent pathways, \(J_1\) and \(J_2\), couples these into an \(S=\tfrac32\) quartet ground state described by
\[
H=-2J_1\,S_1\!\cdot\!S_2-2J_2\,S_1\!\cdot\!S_3.
\]

The molecule is synthesized in one step by Pd(0)-catalyzed radical-radical cross-coupling from di-iodo Blatter radical 5 and nitronyl-nitroxide–\(\mathrm{Au(PPh_3)Cl}\) complex 4 using \(\mathrm{Pd[P(t\text{-}Bu)_3]_2}\) at \(30\,\mathrm{mol\%}\) in dry THF at \(25\,^\circ\mathrm{C}\) for 48 h under Ar. After silica-gel chromatography and pentane wash, the reported yield is \(42\%\) \((93\,\mathrm{mg})\) of pure triradical 3. The article emphasizes that this single-step coupling assembles three spin-\(\tfrac12\) units without over-reduction or side reactions.

Two doublet-quartet gaps are defined:
\[
\Delta E_1 \equiv \Delta E_{DQ}=E(S=\tfrac12,\text{ lowest})-E(S=\tfrac32),
\qquad
\Delta E_2 \equiv \Delta E'_{DQ}=E(S=\tfrac12,\text{ second})-E(S=\tfrac32),
\]
with exact expressions
\[
\Delta E_1 = J_1 + J_2 - \sqrt{J_1^2 + J_2^2 - J_1J_2},
\qquad
\Delta E_2 = J_1 + J_2 + \sqrt{J_1^2 + J_2^2 - J_1J_2}.
\]
From quantitative X-band EPR in frozen toluene/\(\mathrm{CHCl_3}\) glass, the fitted values are \(J_1/k_B=280\,\mathrm{K}\) and \(J_2/k_B=79\,\mathrm{K}\), giving \(\Delta E_1\simeq 0.2\,\mathrm{kcal\,mol^{-1}}\) and \(\Delta E_2\simeq 1.2\,\mathrm{kcal\,mol^{-1}}\). From SQUID magnetometry in polystyrene glass, \(J_1/k_B=425\,\mathrm{K}\) and \(J_2/k_B=109\,\mathrm{K}\), giving \(\Delta E_1\simeq 0.30\,\mathrm{kcal\,mol^{-1}}\) and \(\Delta E_2\simeq 1.8\,\mathrm{kcal\,mol^{-1}}\). The room-temperature quartet population is reported as \(70\pm 5\%\), with the explicit Boltzmann estimate yielding approximately \(75\%\) at \(298\,\mathrm{K}\).

Thermal robustness is central to the paper. Under \(\mathrm{N_2}\) with \(5\,^\circ\mathrm{C\,min^{-1}}\) TGA ramp, onset of decomposition at \(1\%\) mass loss occurs at \(166\,^\circ\mathrm{C}\), and maximum mass-loss rate occurs at approximately \(180\,^\circ\mathrm{C}\). Magnetic measurements further support the quartet ground state: SQUID data at \(19\,\mathrm{mM}\) in polystyrene give \(\langle\chi T\rangle_{300\,\mathrm{K}}\simeq 1.0\,\mathrm{emu\cdot K\cdot mol^{-1}}\), \(M_{sat}\simeq 0.74\,\mu_B\) per molecule, and mean-field \(\theta\approx -0.06\,\mathrm{K}\); X-band EPR at \(9.3\,\mathrm{GHz}\) reports zero-field splitting \(D=+77.8\,\mathrm{MHz}\), \(E=22.7\,\mathrm{MHz}\), and \(g\)-tensor principal values \(g_{xx}=2.0075\), \(g_{yy}=2.0047\), \(g_{zz}=2.0065\).

TriTR was also evaporated under ultra-high vacuum onto \(\mathrm{SiO_2/Si(111)}\) with nominal thicknesses \(0.3\text{–}1\,\mathrm{nm}\). XPS gave a C:N ratio of approximately \(84.4:16.6\) at.\%, close to the stoichiometric \(82.5:17.5\) at.\%, supporting intact deposition. Si \(2p\) attenuation indicated Volmer–Weber island growth, while AFM and SEM showed isolated islands of about \(1\,\mathrm{nm}\) height with lateral sizes of tens to hundreds of nanometers and RMS roughness around \(0.3\,\mathrm{nm}\). The films were stable in UHV for at least 17 h but began to decompose within about 4 h in ambient air; thicker drop-cast films degraded more slowly. In this usage, TriTR names a specific high-spin molecule rather than a general molecular class.

## 7. TriTR as “Search by triplet” in VELO track reconstruction

In collider-detector computing, TriTR denotes “Search by triplet,” the local track reconstruction algorithm used for the LHCb Vertex Locator on parallel architectures [2207.03936]. The problem setting is real-time reconstruction of straight charged-particle trajectories in the VELO, a region outside the sizable magnetic field where tracks serve both as seeds for downstream reconstruction and as inputs to vertexing. The algorithm organizes hits module by module and operates in three phases: sort-by-\(\phi\), track seeding plus following, and tracklet filtering.

The first phase maps each hit azimuth \(\varphi=\operatorname{atan2}(y,x)\in[-\pi,+\pi]\) to a 16-bit unsigned coordinate
\[
\phi=\Big\lfloor\frac{\varphi+\pi}{2\pi}\times 2^{16}\Big\rfloor,
\]
then sorts hits in each module by ascending \(\phi\). On SIMT hardware this uses parallel insertion sort into shared memory, with worst-case \(O(n^2)\) per module; on CPUs the SIMD variant may use `std::sort`, giving \(O(n\log n)\). Data are stored in Structure-of-Arrays form, with \(\phi\) as `uint16` and coordinates as half-precision floats.

Seeding and following operate on module triplets ordered from outside to inside. A boolean flag array suppresses clones. For a candidate pair in modules \(k\) and \(k-1\), the line is extrapolated to module \(k-2\) using
\[
t=\frac{z_{k-2}-z_{k-1}}{z_k-z_{k-1}},
\qquad
(x_{ex},y_{ex})=
(x_k-x_{k-1},\,y_k-y_{k-1})\,t+(x_{k-1},y_{k-1}),
\]
then candidate hits are found through binary search in \(\phi\) windows and a “pendulum search” that alternates upward and downward in sorted \(\phi\) until a fixed number of unflagged hits is collected. Seeds are extended inward one module at a time, allowing one miss before termination. Remaining 3-hit tracklets are fitted by least squares in \((x,z)\) and \((y,z)\), and accepted if
\[
\chi^2=\sum_{i=1}^3 \bigl|\,\bar p_i-(a z_i+b)\bigr|^2
\]
is below threshold and none of the hits is flagged.

The complexity analysis distinguishes worst-case from practical behavior. Sort-by-\(\phi\) costs \(O(mn^2)\) in the SIMT insertion-sort implementation or \(O(mn\log n)\) in the SIMD variant. Seeding is analyzed as \(O(mn^2\log n)\) in one derivation and \(O(mn^2\log^2 n)\) in the strict worst case quoted by the authors. Following is also \(O(mn^2\log n)\), while tracklet filtering is at most \(O(mn)\). In practice, the seeding term dominates at high occupancy.

The implementation is heavily architecture-aware. Reported GPU-oriented choices include Structure-of-Arrays layout, shared-memory sort buffers, separate kernels for sort, seeding, following, and filtering, explicit `__syncthreads()` barriers between seeding and following, pendulum search to prune candidate sets, fast hardware intrinsic `atan2_fast`, scatter metric \(dx^2+dy^2\), and event-level concurrency via CUDA streams. Approximately 14 streams are said to saturate Turing and V100 devices.

Quantitatively, the full VELO sequence reaches \(592\,\mathrm{kHz}\) on an NVIDIA Quadro RTX 6000, \(560\,\mathrm{kHz}\) on an NVIDIA RTX 2080 Ti, \(278\,\mathrm{kHz}\) on an NVIDIA Tesla V100 (32 GB), \(266\,\mathrm{kHz}\) on an AMD Radeon Instinct MI50, and \(228\,\mathrm{kHz}\) on an NVIDIA Tesla T4. Single-socket CPU throughput is reported as \(65\,\mathrm{kHz}\) on AMD EPYC 7502, \(40\,\mathrm{kHz}\) on IBM Power9 IC922, \(22\,\mathrm{kHz}\) on Intel Xeon E5-2630, and \(18\,\mathrm{kHz}\) on Cavium ThunderX2. The evolution from July 2018 to July 2020 is summarized as an overall \(>3\times\) speedup. Physics performance improved from \(95.4\%\) to \(98.52\%\) overall reconstruction efficiency, while overall fake rate fell from \(3.25\%\) to \(0.86\%\); electron efficiency improved from \(70.3\%\) to \(97.38\%\). In this setting, TriTR is a state-of-the-art local-tracking algorithm tailored to heterogeneous CPU/GPU trigger farms.

Source: https://www.emergentmind.com/topics/tritr