Loophole-Free Bell Tests
Locality, detection and freedom-of-choice loopholes, closed.
A Bell violation is only as convincing as the assumptions behind it. For four decades every experiment left a gap a determined local realist could squeeze through: detectors that missed most particles, settings chosen too slowly, statistics that assumed independent trials. In 2015 three groups closed the main gaps at once, and the analysis tools they used now underpin device-independent cryptography.
What you’ll learn
- List the assumptions that turn Bell's theorem into an experimental test, and the loophole each one opens.
- Derive the detection-efficiency threshold for CHSH with a maximally entangled state, and explain how Eberhard's approach lowers it toward .
- Prove the Clauser–Horne inequality from deterministic local strategies and explain why it needs no fair-sampling assumption.
- Bound the probability that any local model, even one with memory, fakes a violation, using the Azuma–Hoeffding inequality.
- Compare the 2015 Delft, Vienna and NIST experiments and explain what remains open (freedom of choice, superdeterminism).
From theorem to experiment
Bell's theorem is a statement about probability distributions. An experiment produces a finite list of events recorded by imperfect devices at a finite distance. To conclude "no local hidden-variable model" you must make sure the hidden variables could not have exploited any gap between the two. Each assumption of the theorem becomes a potential loophole:
| Loophole | The local-realist escape | How it is closed |
|---|---|---|
| Locality (communication) | Information about Bob's setting reaches Alice's device, at or below light speed, before her outcome is fixed | Fast random setting choices; each side's choice and measurement spacelike separated from the other side's |
| Detection (fair sampling) | The hidden variable decides whether a particle is detected, so the detected subset is biased | Efficiency above threshold with every trial counted (Clauser–Horne/Eberhard inequalities), or heralded "event-ready" trials |
| Freedom of choice (measurement independence) | The settings are correlated with through a common cause | Physical random number generators spacelike separated from the source; settings from humans or distant stars push any common cause far back. It cannot be excluded entirely |
| Memory | Devices adapt to earlier rounds, so trials are not independent | Fresh random settings every round and statistics valid for adaptive models (martingales) |
| Coincidence time | Hidden variables shift detection times so that pairing clicks by time windows biases the sample | Pre-defined time slots (pulsed trials) instead of coincidence windows |
In the Delft experiment the two labs were 1.3 km apart. Within how many microseconds must each side choose its setting and record its outcome, so that no light-speed signal carrying the other side's setting can arrive first?
Show the answer
4.3 μs
. Random-number generation and single-shot spin readout both had to fit inside this window. For labs tens to a couple of hundred meters apart, the window shrinks to well under a microsecond.
The detection loophole, quantitatively
Real detectors miss particles. Let each side detect its particle with probability , independently, and let the analysis count every trial, recording a non-detection as the outcome . Discarding non-detections instead means post-selecting the detected subset, and then a local model can choose which particles get detected.
Both detect: the quantum correlator. Only one detects: the other contributes , leaving a marginal. Neither detects: .
For a maximally entangled state every local marginal vanishes, .
Three terms of enter with and one with , so the constant contributes .
Insert and simplify. Below this efficiency, CHSH with a maximally entangled state cannot be violated at all (Garg & Mermin, 1987).
Eberhard (1993) found the way down. Use a non-maximally entangled state together with an inequality that counts single detections. As the required efficiency approaches (with no background counts). The photonic loophole-free tests of 2015 used this strategy with overall detection efficiencies around 75–80%. The right inequality is the Clauser–Horne (CH) inequality, which refers only to "click" events:
Let mean "Alice's detector registers a click when her setting is " and anything else, including no click; same for Bob. Check: if , ; if , .
Average the deterministic bound over . Every probability refers to all trials: nothing is post-selected, and non-detections are simply outcome .
With perfect detection the CH expression is just a rescaled CHSH value, so a maximally entangled state gives . With losses, the CH form stays valid without any extra assumption; the price is that it is violated only for high enough .
A photonic Bell test has an overall detection efficiency of 78% per side and counts every trial. With a maximally entangled state and CHSH, no violation is possible. Which change could make a detection-loophole-free violation possible?
- A brighter source of maximally entangled pairs
- Discarding trials in which only one detector clicks
- A less entangled state, analyzed with a Clauser–Horne/Eberhard-type inequality
- Moving Alice and Bob farther apart
Show the answer
C. A less entangled state, analyzed with a Clauser–Horne/Eberhard-type inequality
is below the threshold for CHSH with maximally entangled states, but above Eberhard's limit of , which is approached with weakly entangled states and an inequality, like CH, that counts every trial.
2015: three loophole-free tests
In 2015 three experiments closed the locality and detection loopholes simultaneously, with fast random setting choices and statistical analyses that do not assume independent trials.
| Feature | Delft (Hensen et al.) | Vienna (Giustina et al.) | NIST (Shalm et al.) |
|---|---|---|---|
| Systems | Electron spins of NV centers in diamond, 1.3 km apart | Polarization-entangled photons | Polarization-entangled photons |
| Entanglement | Heralded: each spin emits a photon entangled with it, and a joint photon detection at a middle station swaps the entanglement onto the spins ("event-ready") | Directly from a parametric down-conversion source | Directly from a parametric down-conversion source |
| Detection loophole | Closed by high-fidelity single-shot spin readout in every heralded trial | Closed by superconducting single-photon detectors and an Eberhard-type inequality | Closed the same way, with a CH/Eberhard-type inequality |
| Statistics | 245 heralded trials, , |
Why heralding works. Delft lost almost all of its photons, yet closed the detection loophole. Which trials count is decided by the herald, which the setting choices cannot influence, and every heralded trial produces an outcome. Losses lower the rate of heralds; they do not bias the sample.
Since then, loophole-free violations have been reported with atoms about 400 m apart (Munich, 2017) and with superconducting qubits in two cryostats joined by a 30 m cryogenic link (ETH Zurich, 2023), whose nanosecond-scale operations make spacelike separation possible over such a short distance.
Statistics, memory and freedom of choice
A finite experiment never proves a negative; it rejects the local-realist hypothesis at some -value. The naive analysis (" is five standard deviations above 2") assumes independent, identically distributed trials. A local model with memory may correlate its behavior across rounds, so the honest question is: what is the largest probability with which any local model, adaptive or not, produces data at least this extreme?
Play rounds of the CHSH game with fresh, uniformly random questions each round; if round is won.
Whatever the devices remember, the round- questions are independent of the past. Conditioned on the history, a local model is again a one-round local strategy.
Its expected increment, given the past, is , and each increment ranges over an interval of length .
The Azuma–Hoeffding inequality for bounded increments.
An 80% win rate over 1000 rounds rejects every local model, memory included, at better than 1%. Real experiments use sharper test-statistic martingales, but the logic is the same.
Freedom of choice can never be closed completely: a common cause in the shared past of the source and the setting generators cannot be excluded, only pushed back in time. The BIG Bell Test (2018) let about 100,000 human volunteers supply setting bits. "Cosmic Bell tests" chose settings from the color of starlight emitted about 600 years earlier (2017), and from light of quasars that left them about 7.8 and 12.2 billion years ago (2018). A local model would need correlations planted that long ago.
Key takeaways
- Each assumption of Bell's theorem is a loophole in practice: locality, fair sampling, measurement independence, independence of trials (memory) and coincidence timing.
- Detection: CHSH with a maximally entangled state needs ; weakly entangled states with CH/Eberhard inequalities need only ; heralded schemes sidestep losses entirely.
- Statistics must hold against adaptive local models: with fresh random questions, (Azuma–Hoeffding).
- Delft, Vienna and NIST (2015) closed locality and detection simultaneously; freedom of choice can only be pushed back (humans, stars, quasars).
- Loophole-free violations underpin device-independent randomness and QKD, and the field was recognized by the 2022 Nobel Prize.
Check yourself
In the interactive lesson these are scored with your confidence, and what you miss comes back for review.
For CHSH with a maximally entangled state, symmetric detection efficiency and every non-detection counted as outcome , which efficiencies allow a violation?
Show the answer
B.
The observed value is , which exceeds 2 iff .
Why does the bound remain valid when the devices remember earlier rounds?
- Because devices with memory are still deterministic
- Because the analysis assumes independent, identically distributed trials
- Because memory can only help quantum devices
- Because each round's questions are fresh and independent of the past, so conditioned on the history the win probability is at most
Show the answer
D. Because each round's questions are fresh and independent of the past, so conditioned on the history the win probability is at most
Given any history, the next round is again a one-shot local strategy facing independent uniform questions, so its conditional win probability is at most . The centered win count is then a supermartingale with bounded increments, and Azuma–Hoeffding applies.
Which loophole can in principle never be closed completely, only pushed further into the past?
- Detection (fair sampling)
- Locality
- Freedom of choice (measurement independence)
- Memory
Show the answer
C. Freedom of choice (measurement independence)
A common cause in the shared past of the source and the setting generators cannot be excluded by any experiment. Human choices and cosmic photons only force such a cause to act earlier and more conspiratorially.
The Delft experiment lost almost all of its photons, yet closed the detection loophole. Why?
- Which trials count is fixed by the herald, which the settings cannot influence, and every heralded trial yields an outcome via high-fidelity spin readout
- The photons at the middle station were detected with 100% efficiency
- Lost photons were replaced by random outcomes in the analysis
- NV centers obey a different Bell inequality than photons
Show the answer
A. Which trials count is fixed by the herald, which the settings cannot influence, and every heralded trial yields an outcome via high-fidelity spin readout
The detection loophole needs the selection of counted trials to depend on the settings or hidden variables in a way the analysis ignores. With event-ready heralding, selection does not depend on the settings and no heralded trial is discarded.
Further reading
- B. Hensen et al., "Loophole-free Bell inequality violation using electron spins separated by 1.3 kilometres", Nature 526, 682 (2015).
- M. Giustina et al., "Significant-loophole-free test of Bell's theorem with entangled photons", Phys. Rev. Lett. 115, 250401 (2015).
- L. K. Shalm et al., "Strong loophole-free test of local realism", Phys. Rev. Lett. 115, 250402 (2015).
- P. H. Eberhard, "Background level and counter efficiencies required for a loophole-free Einstein-Podolsky-Rosen experiment", Phys. Rev. A 47, R747 (1993).
- J.-Å. Larsson, "Loopholes in Bell inequality tests of local realism", J. Phys. A 47, 424003 (2014).