📷 Tool 04 · Single-Atom Detection SNR

Imaging SNR Calculator

Fluorescence imaging budget for single atoms in optical tweezers, photon scattering rate, collection efficiency, camera noise model, and detection fidelity for EMCCD and sCMOS systems.

Single-atom fluorescence imaging is the fundamental read-out mechanism for all tweezer and optical-lattice experiments, every circuit ends with counting atoms. A photon is scattered off the trapped atom onto a camera; the challenge is collecting enough photons for reliable 0/1 discrimination before the atom heats out of the trap. This calculator computes signal-to-noise ratio (SNR) for your specific imaging setup (species, NA, detuning, camera) and the resulting single-atom detection fidelity, plus the new Histogram Fidelity section which shows how to extract detection fidelity from the bimodal fluorescence histogram, the same analysis used in Evered et al. 2023, Manetsch et al. 2025, and Wu et al. 2022.
Common mistake, SNR is not fidelity. A high photon-count SNR is only a proxy. Detection fidelity is the overlap error of the dark-count and bright-count histograms after choosing a threshold, and it can be limited by atom loss, blinking, camera tails, or imperfect state preparation even when the raw SNR looks excellent.
Picture: collect enough fluorescence photons that the bright-atom histogram is separated from the empty-trap histogram. The useful question is not "how many photons did I collect?" but "how much do the two histograms overlap?"

More NA helps twice

Sentence: Larger NA increases collected photons without increasing atom heating.

ηcoll = (1 - cosθ)/2

Failure mode: assuming higher camera gain improves photon collection.

Thresholds are physics

Sentence: The optimal threshold trades false positives against false negatives.

F = 1 - (P0→1 + P1→0)/2

Failure mode: quoting SNR when state loss dominates.

Survival is separate

Sentence: A measurement can identify an atom correctly while still ejecting it.

Freadout ≠ Psurvival

Failure mode: multiplying every error into one unlabeled number.

01 Imaging Beam

Atom & Imaging Transition
0.01×I_sat1×I_sat100×I_sat
⁸⁷Rb D2 line: λ = 780.241 nm · Γ/2π = 6.065 MHz · Isat = 1.67 mW/cm²
Scattering rate
photons / s
— photons / ms
Saturation parameter
I / Isat
Branching / linewidth
Γ/2π (MHz)
τ = 1/Γ
R_max at saturation
photons / s (s→∞, Δ=0)
= Γ/2

Two-level scattering rate

On resonance (Δ=0) a two-level atom saturates at Γ/2. Off resonance the Lorentzian denominator reduces the rate:

$$R_{\rm sc} = \frac{\Gamma}{2} \cdot \frac{s}{1 + s + (2\Delta/\Gamma)^2}$$ $s = I/I_{\rm sat}$, $\Gamma$ = natural linewidth, $\Delta$ = laser detuning

02 Collection & Camera

Objective & Optics
NA 0.10NA 0.50NA 0.95
0.1 ms1 ms10 ms100 ms
Camera Noise Model
100×1000×
01 e⁻ (sCMOS)60 e⁻ (EMCCD CIC)
EMCCD: excess noise factor √2 applied; read noise divided by EM gain.

03 Results

Signal-to-Noise Ratio
15102050100+
Collection efficiency η
η = ηgeo × QE × T
ηgeo = —
Photons scattered (total)
in exposure time
Photons detected (signal)
Nsig = R_sc × t × η
Detection fidelity
Min exposure for SNR=10
target SNR = 10
Min exposure for SNR=50
target SNR = 50
Noise Budget
Source σ² (photons²) Fraction Contribution
Set parameters above to compute noise budget.
🔬
Hood Lab imaging setup: We image 87Rb and 133Cs single atoms in optical tweezers using a NA = 0.6 objective (Mitutoyo G Plan Apo) and an EMCCD camera (Andor iXon). Typical imaging: 780 nm D2 line, detuning Δ = −2Γ, saturation s ≈ 2, exposure 2 ms. We collect ~600 photons from Rb at 99% survival probability with gray molasses applied during imaging. The critical figure is not peak SNR but the histogram separation between the 0-atom and 1-atom peaks — we target >8σ separation to achieve detection fidelity > 99.9%.

04 SNR vs Exposure Time

SNR as a function of imaging duration
Current settings SNR = 10 target SNR = 50 target

05 Fluorescence Histogram — Single-Atom Detection

Each frame adds simulated imaging shots drawn from the real Poisson distributions for your settings. Green = atom present (N_sig photons) · Gray = empty tweezer (N_bg background) · dashed = optimal threshold θ. Detection fidelity = 1 − P(miss) − P(false alarm).


05 Physics & Formulas

Geometric collection fraction η_geo

A high-NA objective collects photons into a cone defined by the half-angle θ_max = arcsin(NA/n), where n is the refractive index of the medium (n=1 in air/vacuum).

$$\eta_{\rm geo} = \frac{1 - \sqrt{1 - {\rm NA}^2}}{2} \quad \text{[exact, for } n=1\text{]}$$ $$\approx \frac{{\rm NA}^2}{4} \quad \text{[paraxial, NA} \ll 1\text{]}$$ Example: NA $= 0.6 \Rightarrow \eta_{\rm geo} = 10\%$; NA $= 0.8 \Rightarrow \eta_{\rm geo} = 20\%$

Note that isotropic emission into 4π steradians means even NA=0.95 captures only ~26% of photons geometrically. High-NA objectives (0.5–0.85) are typical for tweezer experiments.

Noise variance per pixel per frame

Each camera type has a distinct noise model. The total variance σ²_total determines the denominator of SNR = N_sig / σ_total.

EMCCD (excess noise factor $F = \sqrt{2}$): $$\sigma^2 = F^2(N_{\rm sig} + N_{\rm bg} + N_{\rm dark}) + (\sigma_{\rm read}/G)^2 = 2(N_{\rm sig} + N_{\rm bg} + N_{\rm dark}) + (\sigma_{\rm read}/G)^2$$ sCMOS (no multiplication noise): $$\sigma^2 = N_{\rm sig} + N_{\rm bg} + N_{\rm dark} + \sigma_{\rm read}^2$$ Ideal: $\sigma^2 = N_{\rm sig} + N_{\rm bg} + N_{\rm dark}$
where $G$ = EM gain, $\sigma_{\rm read}$ = read noise (e$^-$ RMS)

For EMCCD the read noise is divided by EM gain G, making it negligible at high gain (G ≳ 50). The penalty is the √2 excess noise factor that increases shot noise variance by 2×. For sCMOS cameras (1–2 e⁻ read noise), read noise is the dominant term at low photon counts.

Binary atom detection

We distinguish "atom present" (Poisson mean μsig) from "no atom" (Poisson mean μbg) by choosing a threshold θ. The optimal threshold minimises total error probability.

For Gaussian approximation ($\mu \gg 1$): $$P_{\rm miss} = \Phi\!\left(\frac{-(\mu_{\rm sig} - \theta)}{\sqrt{\sigma^2_{\rm sig}}}\right) \quad \text{[false negative]}$$ $$P_{\rm false} = 1 - \Phi\!\left(\frac{\theta - \mu_{\rm bg}}{\sqrt{\sigma^2_{\rm bg}}}\right) \quad \text{[false positive]}$$ Optimal $\theta \approx (\mu_{\rm sig} + \mu_{\rm bg})/2$
Detection fidelity: $\mathcal{F} = 1 - (P_{\rm miss} + P_{\rm false})/2$.
If the bright and dark peaks have equal variance and ${\rm SNR}=(\mu_{\rm sig}-\mu_{\rm bg})/\sigma$, then $\mathcal{F}\approx\Phi({\rm SNR}/2)$. The calculator below uses the bright and dark variances separately.

For SNR ≥ 10, fidelity exceeds 99.9%. The key insight is that fidelity scales as erfc(SNR), doubling SNR dramatically reduces error. Background suppression (e.g. dark-field or EIT imaging) improves fidelity by reducing μ_bg without reducing μ_sig.

Photon recoil vs trap depth

Every scattered photon imparts a recoil kick ℏk. At scattering rate R_sc the atom heats at rate dE/dt = E_rec × R_sc (in a σ⁺/σ⁻ Sisyphus beam, this is partially cancelled; in a σ beam near resonance it accumulates).

$$E_{\rm rec} = \frac{\hbar^2 k^2}{2m}$$ $^{87}$Rb D2: $E_{\rm rec}/k_{\rm B} \approx 181$ nK, so one absorption plus one emission deposits $\sim 2E_{\rm rec}/k_B \approx 362$ nK
Heating rate: $\dot{T} \approx 2E_{\rm rec} R_{\rm sc}/k_{\rm B}$ [absorption + spontaneous emission, no cooling]
Max imaging time ($U_0$ = trap depth): $t_{\rm max} \sim U_0/(E_{\rm rec} R_{\rm sc})$
Example: $U_0/k_{\rm B} = 1$ mK, $R_{\rm sc} = 50$ kHz $\Rightarrow t_{\rm max} \approx 55$ ms

In practice, tweezer imaging uses repump + imaging beams in a gray molasses or Λ-enhanced dark SPOT configuration to scatter >1000 photons without losing the atom. This pushes fidelity above 99.5% (Bergamini 2004; Liu 2019).


06 Imaging Histogram & Detection Fidelity

What is the imaging histogram?

Run your experiment ~500–2000 times. Each shot, sum the camera electrons over the region of interest (typically a 3–7 px ROI around one tweezer site). Plot the distribution of those counts. You get a bimodal histogram: a narrow "dark" peak when the site is empty (only background scatter + camera noise), and a broader "bright" peak when an atom is present (atom fluorescence + background). The overlap between the two peaks is the fundamental limit on how well you can distinguish atom present from atom absent, this is your imaging fidelity.

Anatomy of a single-atom imaging histogram
Detected signal (counts / electrons) Number of shots Dark peak (no atom) μ_dark Bright peak (atom present) μ_bright θ (threshold) Miss False alarm σ_dark σ_bright SNR = (μ_bright − μ_dark) / σ
θ = threshold. Shots below θ → "no atom"; shots above θ → "atom". Orange = missed atoms (false negative). Red = false alarms (false positive). The overlap area determines error rates.

Peak models

For $N_{\rm det} \gg 1$ detected photons, both peaks are well-approximated by Gaussians (Central Limit Theorem on Poisson photon statistics + Gaussian read noise):

$$P_{\rm dark}(n) = \mathcal{N}(\mu_{\rm dark},\,\sigma_{\rm dark}^2)$$ $$\mu_{\rm dark} = N_{\rm bg} + N_{\rm dark} + \sigma_{\rm read}^2/G^2$$ $$\sigma_{\rm dark}^2 = F^2 \mu_{\rm dark} + (\sigma_{\rm read}/G)^2$$ $$P_{\rm bright}(n) = \mathcal{N}(\mu_{\rm bright},\,\sigma_{\rm bright}^2)$$ $$\mu_{\rm bright} = N_{\rm sig} + \mu_{\rm dark}$$ $$\sigma_{\rm bright}^2 = F^2 \mu_{\rm bright} + (\sigma_{\rm read}/G)^2$$ $F=\sqrt{2}$ (EMCCD excess noise), $F=1$ (sCMOS/ideal)

Why σ_bright > σ_dark

Shot noise grows as $\sqrt{N}$. More photons in the bright peak → larger absolute fluctuations → the bright peak is always wider than the dark peak. This asymmetry means the optimal threshold is not exactly at the midpoint — it shifts slightly toward the brighter side to equalize miss and false-alarm probabilities.

Error probabilities & fidelity

Given threshold $\theta$, define two error types:

$P_{\rm miss}(\theta) = \Phi\!\left(\frac{\theta - \mu_{\rm bright}}{\sigma_{\rm bright}}\right) \quad$ [bright peak below θ]

$P_{\rm false}(\theta) = 1 - \Phi\!\left(\frac{\theta - \mu_{\rm dark}}{\sigma_{\rm dark}}\right) \quad$ [dark peak above θ]

$$\mathcal{F}_{\rm detect} = 1 - \frac{P_{\rm miss} + P_{\rm false}}{2}$$ Optimal threshold: minimise $P_{\rm miss} + P_{\rm false}$.
Equal $\sigma$: $\theta_{\rm opt} = (\mu_{\rm bright} + \mu_{\rm dark})/2$.
Unequal $\sigma$: solve numerically.

For equal peak widths and ${\rm SNR} = (\mu_{\rm bright}-\mu_{\rm dark})/\sigma$:
$\mathcal{F} \approx \Phi\!\left(\tfrac{\rm SNR}{2}\right)$. With unequal widths, compute $P_{\rm miss}$ and $P_{\rm false}$ separately.
SNR → Fidelity rule of thumb:
For equal-width Gaussian peaks and SNR defined as peak separation divided by one-peak σ: SNR = 5 → F ≈ 99.4% and SNR = 10 → F > 99.9999%. Real cameras often have unequal bright/dark widths, so histogram fitting is the safer final metric.
Interactive Histogram Simulator
Preset:
SNR
(μ_bright−μ_dark)/σ_rms
Detection fidelity
1 − (P_miss+P_false)/2
P_miss (false negative)
bright below θ
P_false (false positive)
dark above θ
Optimal threshold θ_opt
minimises P_miss + P_false
Dark peak (no atom) Bright peak (atom) Miss zone (orange) False alarm zone (red) Threshold θ

Alkali (⁸⁷Rb, ¹³³Cs)

Imaging transition: D2 cycling transition (780 nm for Rb, 852 nm for Cs). Closed cycling transition with repumper keeps the atom in the imaging cycle. R_sc → Γ/2 at saturation (~3 MHz for Rb87).

Photon counts: With NA=0.5, η≈5%, 5ms exposure → ~750 detected photons. σ_bright ≈ 40e⁻ (EMCCD with excess noise). Well-separated from dark peak (~50 e⁻).

Main challenge: Photon recoil heating (~181 nK per one-way recoil for Rb87; roughly twice that per scatter cycle without cooling). At R_sc = 3×10⁶/s, the atom heats by ~1 mK/s. Limit imaging to <10 ms or use concurrent Sisyphus/gray molasses cooling (recapture imaging, EIT imaging).

Background: Stray light from near-resonant beams contributes to dark peak. Typical dark peak: μ_dark = 50–200 e⁻ depending on background suppression.

Alkaline-earth-like (¹⁷¹Yb, ⁸⁸Sr)

Imaging transition: Broad ¹P₁ transition (399 nm for Yb, 461 nm for Sr, Γ/2π = 29/32 MHz). Very high R_sc ≥ Γ/2 at low saturation. Alternative: narrow ³P₁ (556 nm Yb, 689 nm Sr) for gentler imaging.

J=0 ground state advantage: Qubit states (nuclear spin) are largely spectator degrees of freedom for electronic cycling transitions such as ¹S₀ → ¹P₁. Atoms in the metastable ³P₀ state are invisible to 399/461 nm, enabling shelved (non-destructive) readout of one qubit basis while leaving the other undisturbed.

Magic wavelength: 759 nm tweezer is magic for ¹S₀/³P₀, no differential light shift on the qubit. The imaging beam (399 nm) shifts ¹S₀ but not ³P₀, so imaging the un-shelved population gives state discrimination.

Histogram: Narrower dark peak (lower background at UV wavelengths far from tweezer), tight bright peak. Excellent separation even at 1ms exposure. Atom Computing Yb171: reports >99.9% imaging fidelity.

Two-Gaussian fit procedure

Given your measured histogram H(n), fit the sum of two Gaussians:

$$H(n) = A_0\,\mathcal{N}(n;\,\mu_{\rm dark},\sigma_{\rm dark}^2) + A_1\,\mathcal{N}(n;\,\mu_{\rm bright},\sigma_{\rm bright}^2)$$ 6 free parameters: $A_0, \mu_{\rm dark}, \sigma_{\rm dark}, A_1, \mu_{\rm bright}, \sigma_{\rm bright}$. Fit with scipy.optimize.curve_fit or a Gaussian mixture model (sklearn.mixture.GaussianMixture).

Quality check: is $A_0/A_1 \approx$ expected loading fraction? Is $\sigma_{\rm bright}/\sigma_{\rm dark} \approx \sqrt{\mu_{\rm bright}/\mu_{\rm dark}}$? (this tests whether shot noise dominates as expected)

When peaks are NOT clean Gaussians

Real histograms show deviations from two clean Gaussians in several situations:

  • Atom loss during imaging: Some fraction of bright shots become mid-sequence dark → bright peak develops a low-side tail. Fix: reduce imaging power or duration, or use concurrent cooling to prevent loss.
  • Multiple occupancy: If two atoms load into one site, you get a third peak at ~2×μ_bright. Set loading conditions (MOT density, tweezer depth) to suppress. Good diagnostic: multi-peak structure at integer multiples of μ_bright.
  • Background drift: If background changes shot-to-shot (laser power fluctuations), both peaks broaden. σ_measured² = σ_intrinsic² + σ_drift². Use real-time background subtraction.
  • Non-Poissonian photon statistics: Sub-Poissonian emission (from e.g. shelving protocols) or super-Poissonian (from intensity noise) → peak widths deviate from $\sqrt{\mu}$ scaling. Check: plot σ² vs μ, should be linear with slope 1 (Poisson) or slope F² (EMCCD).
  • Charge-transfer smear (EMCCD): Very fast pixel readout + high scattering rates → charge smear across pixels. Solved by choosing appropriate readout speed or using sCMOS.

Bootstrap fidelity uncertainty

After fitting, estimate statistical uncertainty via bootstrapping: resample your histogram N_shots times with replacement, refit, and take the std of the fidelity distribution. Typical uncertainty: δF ≈ 0.01–0.1% for 1000 shots, 0.001–0.01% for 10,000 shots. Note: systematic error from threshold choice is usually larger than statistical error for well-separated peaks.

Definitions used in neutral-atom QC papers

Different papers define "fidelity" slightly differently. For qubit experiments, the total SPAM (state preparation and measurement) error bundles several contributions:

Total SPAM error: $\varepsilon_{\rm SPAM} = \varepsilon_{\rm prep} + \varepsilon_{\rm meas}$

$\varepsilon_{\rm meas} = (P_{\rm miss} + P_{\rm false})/2$, from histogram overlap [what this section computes]

$\varepsilon_{\rm prep}$, optical pumping error, atom number fluctuations, tweezer loading fraction

$\mathcal{F}_{\rm SPAM} = 1 - \varepsilon_{\rm SPAM} \approx 1 - \varepsilon_{\rm meas} - \varepsilon_{\rm prep}$

In Rb87 Harvard experiments: $\varepsilon_{\rm meas} \lesssim 0.1\%$, $\varepsilon_{\rm prep} \lesssim 0.2\%$ → $\mathcal{F}_{\rm SPAM} > 99.5\%$
⚠️ Common confusion: Evered et al. 2023 (99.5% two-qubit gate) reports gate fidelity after SPAM correction, meaning SPAM errors are subtracted out. The raw circuit fidelity is lower. Always check whether a paper's quoted fidelity is SPAM-corrected or not. For practical fault-tolerant thresholds, you need the raw (uncorrected) fidelity.
🔬
Detection fidelity in practice: In tweezer experiments the dominant infidelity source shifts from photon shot noise (fixed by increasing NA or exposure) to atom loss during imaging (fixed by reducing recoil heating via gray molasses or EIT cooling). For Cs we switch to a 685 nm gray molasses imaging scheme: the cooling rate prevents axial escape while the camera collects photons, enabling ≥ 99.5% detection fidelity with only 1 ms exposure. The limiting factor at that level is preparation fidelity (state initialization), not imaging photon budget.

Key Papers — Single-Atom Imaging & Detection

[1] Schlosser et al., Sub-Poissonian loading of single atoms in a microscopic dipole trap, Nature 411, 1024 (2001) — first single-atom tweezer imaging. DOI
[2] Bakr et al., A quantum gas microscope for detecting single atoms in a Hubbard-regime optical lattice, Nature 462, 74 (2009) — quantum gas microscope, site-resolved imaging. DOI
[3] Fuhrmanek et al., Free-space lossless state detection of a single trapped atom, PRL 106, 133003 (2011) — 99.4% detection fidelity benchmark. DOI
[4] Evered et al., High-fidelity parallel entangling gates on a neutral-atom quantum computer, Nature 622, 268 (2023) — detection fidelity analysis from bimodal histogram (Supplementary Methods). DOI
[5] Tuchendler et al., Energy distribution and cooling of a single atom in an optical tweezer, PRA 78, 033425 (2008) — atom survival and heating during imaging; release-recapture thermometry. DOI
[6] Manetsch et al., A tweezer array with 6100 highly coherent atomic qubits, Nature (2024) — large-scale tweezer array imaging and loading statistics at >99.9% fidelity. DOI

07 References & Further Reading