Entanglement Cannot Send a Message. Here Is the Two-Line Proof.

I'll start with a number. In the 2015 Delft Bell test, the two entangled electron spins sat in labs 1.3 km apart [1]. Light covers that distance in $1.3\times10^{3}\ \text{m} / 3.00\times10^{8}\ \text{m/s} = 4.3\ \mu\text{s}$. Suppose the popular story were true and measuring one spin "instantly affects" the other. Then the experimenter at one end could send a message to the other end at least 4.3 μs faster than light. How much information about Bob's choice of measurement reaches Alice's detector per entangled pair? The answer is exactly 0 bits. That is not a small number limited by noise. It is zero, for every quantum state and for every operation Bob can do.

That is the no-signalling theorem. Ghirardi, Rimini and Weber published it in 1980 [2], and it takes two lines once the right object is written down. My thesis is about teaching order. Any popular account of entanglement should derive no-signalling, or at least state it, before it uses the word "spooky". An account that skips it teaches readers a false picture of a quantum telephone, and later explanations spend their effort taking that picture apart. I hold this view at 0.85, and the rest of this post is my reason for it.

The two lines

Alice and Bob share any two-part state $\rho_{AB}$, pure or mixed, entangled or not. Every prediction Alice can test with her own apparatus, without Bob's records, comes from her reduced density matrix:

$$\rho_A = \mathrm{Tr}_B\,\rho_{AB}.$$

The probability of outcome $a$ in any measurement she makes, with POVM elements $E_a$, is $p(a) = \mathrm{Tr}(E_a\rho_A)$. So the question "can Bob signal to Alice?" turns into "can Bob change $\rho_A$?"

Bob can do anything quantum mechanics allows on his own system: a unitary, a projective measurement, a noisy POVM, coupling to an ancilla in his lab. All of these are covered by a set of Kraus operators $K_k$ acting on B, with $\sum_k K_k^\dagger K_k = I$. Alice does not know which outcome $k$ Bob got, so her description after his operation averages over $k$. Here is line one:

$$\rho_A' = \mathrm{Tr}_B\Big[\sum_k (I\otimes K_k)\,\rho_{AB}\,(I\otimes K_k^\dagger)\Big] = \mathrm{Tr}_B\Big[\big(I\otimes \textstyle\sum_k K_k^\dagger K_k\big)\,\rho_{AB}\Big].$$

The step uses the cyclic property of the partial trace for operators that act only on B. Here is line two:

$$\rho_A' = \mathrm{Tr}_B\big[(I\otimes I)\,\rho_{AB}\big] = \rho_A .$$

That completes the proof. Bob's operation leaves Alice's density matrix unchanged, so every probability she can measure is unchanged too. The only physics used is that Bob's operator has the form $I\otimes K$, meaning it acts on his system, plus completeness of his measurement. Entanglement played no role in the argument, and that is the point: the most entangled state there is gets no exemption.

This has the same structure as the duality argument in the Ising post, where four lines fixed the critical temperature and nothing more. Here two lines rule out signalling and nothing more. In the section on objections I am careful about what "nothing more" excludes.

A singlet, worked to the last digit

General proofs persuade few people. A worked example persuades more. Take the singlet

$$|\psi^-\rangle = \tfrac{1}{\sqrt2}\big(|{\uparrow\downarrow}\rangle - |{\downarrow\uparrow}\rangle\big).$$

Before Bob does anything, Alice's reduced state is $\rho_A = \tfrac12|{\uparrow}\rangle\langle{\uparrow}| + \tfrac12|{\downarrow}\rangle\langle{\downarrow}| = I/2$.

Bob measures along z. With probability 1/2 he gets up, and Alice's conditional state is $|{\downarrow}\rangle$. With probability 1/2 he gets down, and hers is $|{\uparrow}\rangle$. Averaged over his outcomes:

$$\rho_A' = \tfrac12|{\downarrow}\rangle\langle{\downarrow}| + \tfrac12|{\uparrow}\rangle\langle{\uparrow}| = \begin{pmatrix} 1/2 & 0 \\ 0 & 1/2 \end{pmatrix}.$$

Bob measures along x. The singlet has the same form in every basis, so his outcomes $\pm$ leave Alice in $|{\mp}\rangle = (|{\uparrow}\rangle \mp |{\downarrow}\rangle)/\sqrt2$, each with probability 1/2. In the z basis, $|{+}\rangle\langle{+}|$ has off-diagonal entries $+1/2$ and $|{-}\rangle\langle{-}|$ has $-1/2$. The average cancels them:

$$\rho_A' = \tfrac12|{-}\rangle\langle{-}| + \tfrac12|{+}\rangle\langle{+}| = \begin{pmatrix} 1/2 & 0 \\ 0 & 1/2 \end{pmatrix}.$$

The two matrices agree entry by entry. They are equal, not merely close. For any axis $\hat n$, Alice's probability of "up" is $\mathrm{Tr}\big[\tfrac12(I+\hat n\cdot\vec\sigma)\cdot\tfrac12 I\big] = 1/2$. If Bob chose z to mean 0 and x to mean 1, Alice's detector clicks up half the time either way. With $10^{6}$ pairs or $10^{12}$ pairs, the distributions she samples from are identical, so no hypothesis test can tell them apart. The mutual information between Bob's setting and Alice's outcome is exactly 0 bits per pair.

The same fact in the notation of Bell tests: for spin measurements at angles $\theta_x$ (Alice) and $\theta_y$ (Bob) in a plane, the singlet gives

$$P(a,b\,|\,x,y) = \tfrac14\big[1 - ab\cos(\theta_x-\theta_y)\big],\qquad a,b=\pm1.$$

Summing over Bob's outcome gives $\sum_b P(a,b|x,y) = \tfrac14[2 - a\cos(\theta_x-\theta_y)(1-1)] = \tfrac12$. Bob's angle drops out. That cancellation, $(1-1)$, is the no-signalling condition written as a table of numbers.

Where does the famous correlation live, then? In the product $ab$, and only there. The correlation is $E(x,y) = -\cos(\theta_x-\theta_y)$. With $\theta_x\in\{0^\circ, 90^\circ\}$ and $\theta_y\in\{45^\circ, -45^\circ\}$, each correlation has magnitude $\cos 45^\circ = 0.7071$. The CHSH combination is

$$S = \big|E(0,45)+E(0,-45)+E(90,45)-E(90,-45)\big| = 4\times0.7071 = 2\sqrt2 \approx 2.828,$$

and any local hidden-variable model satisfies $S\le 2$. You cannot compute $ab$ until both records are in one place, which takes a classical channel running at light speed or slower. The Delft group measured $S = 2.42 \pm 0.20$ over 245 trials, with $p = 0.039$ against local realism [1]. Every one of those 245 values of $ab$ was computed after the data from both stations had been brought together.

Teleportation shows the same accounting from the other side. The standard protocol (Bennett et al., Phys. Rev. Lett. 70, 1895, 1993) needs two classical bits per qubit. Before those bits arrive, the receiver's qubit is $I/2$ whatever state was sent. The two lines above are the reason the classical bits cannot be skipped.

What the Nobel experiments established, and what they did not

The 2022 Nobel Prize in Physics went to Aspect, Clauser and Zeilinger "for their experiments with entangled photons, establishing the violation of Bell's inequalities and pioneering quantum information science" [3]. The design of those experiments contains the no-signalling logic. Physics World's description of Zeilinger's setup says the settings were arranged so that "no signal travelling slower than the speed of light would be able to transfer information to the other side" before the measurements were registered [3]. The aim was to rule out ordinary communication between the stations as an explanation of the correlations. It was not an attempt to send a message faster than light, and no such message was observed.

What a Bell violation establishes is narrower and stranger than an influence. No model in which each outcome is fixed by local variables, plus whatever reaches it at sublight speed, reproduces the statistics. That claim is about the class of explanations allowed. It is not a claim about a channel.

Two later results pin down how independent these ideas are. Tsirelson (Cirel'son) proved in 1980 that quantum correlations cannot exceed $S = 2\sqrt2$ [4]. Popescu and Rohrlich then asked whether no-signalling is what enforces that ceiling [5]. It is not. They built a hypothetical "PR box" whose correlations reach $S = 4$, the algebraic maximum, while every marginal stays at 1/2, so it is perfectly no-signalling [5]. That makes three distinct levels: local models at 2, quantum mechanics at 2.828, and no-signalling alone permitting 4. Nonlocal correlation and signalling are separate things, and quantum mechanics sits well inside the no-signalling region. Rohrlich later argued that adding a classical limit to relativistic causality and nonlocality recovers Tsirelson's bound [6]. That is an open research programme, not textbook material, and I cite it only to show that the separation is taken seriously at the research level.

The strongest objection

The objection I take most seriously goes like this. "No-signalling is a statement about averages. Bell's theorem says something nonlocal is really happening. In Bohmian mechanics there is an explicit instantaneous dependence of one particle's motion on the distant setting, hidden only because the hidden variables are distributed in quantum equilibrium. In textbook collapse, Bob's measurement updates the global state at once. So 'spooky action' names a real feature, and your two lines are a technicality about what can be controlled."

I grant most of it. Bell's theorem, together with loophole-free tests such as Delft's [1], does rule out local causal explanations. Some interpretations do contain a nonlocal dependence at the level of their hidden variables. I am not claiming that entanglement is unremarkable.

The crux is what the word "influence" means to a reader with no training. Popular accounts say "change one particle and the other changes instantly." A stranger reads that as a dependence that can be controlled: twist this one and that one twists. That reading is exactly what the two lines exclude, for every state and every interpretation that reproduces the quantum predictions. Whatever nonlocality Bell finds, it cannot be used to change anything at the other end. Telling readers about an effect that cannot be used, while leaving that fact out, gives them the wrong picture.

Collapse does not rescue the influence reading either. Alice's conditional state after Bob's outcome is pure, but it is conditional on information she does not have. Without that information she holds $I/2$, as computed above. There is also a sharper problem. When the two measurements are spacelike separated, their time order depends on the reference frame. Because $K\otimes I$ and $I\otimes K'$ commute, the joint distribution is the same in either order. "Bob's measurement collapsed Alice's spin" and "Alice's measurement collapsed Bob's" are both available descriptions, and they give identical predictions. An influence whose direction depends on the frame, and which has no observable consequence, should not be the headline.

So I am not asking popular writers to drop the strangeness. I am asking them to state the no-signalling fact first. If readers know that nothing Bob does changes Alice's statistics, they will place the strangeness where it belongs: in correlations that no local model can produce, found only when two records are compared.

Limits of the proof

Where this fails: the proof has four assumptions, and each one marks a regime where its conclusion changes.

1. Linearity. The proof uses the fact that Bob's operation is linear in $\rho$. Nonlinear modifications of the Schrödinger equation break this. Gisin showed that Weinberg's nonlinear quantum mechanics permits superluminal signalling (Phys. Lett. A 143, 1, 1990). No-signalling is therefore a constraint that any proposed extension of quantum mechanics has to satisfy.
2. No perfect copying. If Alice could clone her qubit many times, she could estimate a pure conditional state and tell Bob's z basis from his x basis. Wootters and Zurek proved in 1982 that linearity forbids cloning an unknown state [7]. Their paper answered a faster-than-light scheme of exactly this kind, Herbert's FLASH proposal. The superluminal telephone failed at exactly the step the theorem predicts: Alice's receiver reads $I/2$.
3. Operations that are truly local. The form $I\otimes K$ assumes Bob acts only on his own subsystem. If his apparatus interacts with Alice's system, he can signal, but at the speed of whatever carries the interaction. In relativistic quantum field theory, this assumption becomes microcausality: observables at spacelike separation commute. Recent work argues for the converse, that no-signalling implies microcausality [8].
4. Unconditioned statistics. If Alice sorts her outcomes by Bob's results, her subensembles do depend on his setting. Sorting needs his records, so it needs a classical channel. Many "quantum eraser shows retrocausality" stories fall at this step.

The theorem also proves less than it may seem to. It does not explain why quantum correlations stop at 2.828 rather than 4 [4][5]. A theory with the same no-signalling property could be far more nonlocal than ours.

What follows if I am right

If the argument holds, here is a test anyone can apply to any explainer. Does the explainer imply that Alice, using only her own data, can learn anything about what Bob chose? If it does, the explainer is wrong. The test fails nothing that the Nobel experiments actually showed, and it fails most headlines about them. What would change my mind is a reproducible experiment, with spacelike-separated stations, in which Alice's marginal frequencies depend on Bob's setting by more than their statistical error. That result would not refute my teaching advice. It would refute quantum mechanics, and it would be the biggest physics story in a century. Until it exists, the two lines come first, and the word spooky, if it is used at all, comes after.