Foundations of quantum mechanics · Theory
Information Causality Defeats Every Finite NPA Level
arXiv:2611.21371v1 ↗PDF ↗Submitted 16 November 2026
- Question
- Information causality says that if Alice sends Bob \(m\) classical bits, the information Bob can gain about her independent database bits, summed over which bit he is asked for, is at most \(m\). Quantum correlations obey it. The Navascués–Pironio–Acín (NPA) hierarchy approximates the quantum set from outside by levels of semidefinite constraints. Does some fixed finite level already enforce information causality in every Bell scenario, however many measurement settings it has?
- Result
- No. For every level \(L\ge2\) the authors build a correlation with two outcomes and \(N\) settings per party that satisfies all level-\(L\) constraints, in which both parties obtain 0 only when their settings agree, with probability \(s_L>1/(185L^3)\) at each equal pair, independent of \(N\). One use of it plus one bit from Alice yields information at least \((1-e^{-1})s_L\log_2N\), which exceeds the one-bit budget once \(N\) is large. At fixed \(L\) the best one-use, one-bit score grows as \(\log N\), with leading coefficient exactly \(s_L\). The violation survives when every outcome has positive probability and when every restriction to a bounded number of settings admits a local model.
The paper in detail
With independent copies of a Popescu–Rohrlich box, Alice can send one bit that lets Bob read whichever bit he wants from an arbitrarily long database, without Alice learning which bit he asked for. Information causality was proposed in 2009 to rule this out: the information Bob gains about the requested bit, summed over all possible requests, may not exceed the number of bits sent. Quantum mechanics obeys it, and it recovers Tsirelson's bound for the CHSH inequality. The NPA hierarchy describes quantum correlations from outside: at level \(L\) the observed probabilities must be entries of a positive semidefinite moment matrix indexed by products of at most \(L\) measurement projectors. Earlier work found violations of information causality at level 1 and in the almost-quantum set, which sits between levels 1 and 2. The question here is whether any fixed finite level suffices in every scenario.
The answer rests on a simple correlation. Each party has \(N\) settings and two outcomes; both get 0 with probability \(s_L\) when the settings agree and never otherwise, and each party alone gets 0 with probability \(3s_L\). Alice's database bits are each 1 with probability \(r\). She measures at the position of the first 1 and sends her outcome; Bob measures at the requested position \(i\) and outputs 1 when both his outcome and her message are 0. That event is impossible unless Alice measured at \(i\), so it certifies that \(R_i=1\), an event of prior probability \(r\) that carries \(\log_2(1/r)\) bits. Summing over requests gives at least \(s_L[1-(1-r)^N]\log_2(1/r)\), which at \(r=1/N\) is at least \((1-e^{-1})s_L\log_2N\). Everything therefore hinges on keeping \(s_L\) fixed as \(N\) grows.
That is the hard part. The natural vector model pairs Alice's and Bob's setting directions, giving joint-zero moments \(\delta_{xy}\), but its identity vector has squared norm \(4N\), so normalising it sends the matched probability to zero. The authors use the fact that level \(L\) sees only products of at most \(L\) projectors. They replace the identity vector by another vector that reproduces every inner product with nonempty products up to that length, prove that the resulting moment functional is still positive, and bound the new vector's norm independently of \(N\) by giving the projectors random signs on shared modes and choosing mode weights by polynomial interpolation. The normalisation \(C_L<185L^3\) becomes \(1/s_L\).
The rest of the paper maps the boundaries. Any nonsignalling correlation used once with one bit scores at most \(\log_2N+1\), because Alice's setting and bit can be sent classically together with a pre-shared table of responses; the fixed-level optimum is therefore \(\Theta(\log N)\), with leading coefficient \(s_L\). Quantum and even commuting-operator correlations without joint zeros at unequal settings have average matched probability at most \(1/N\), which kills the logarithmic growth, and at level \(L\) exact certificates obey \(s_L\le(2L-3)^{-2}\), so the level must grow roughly as \(\sqrt N\) to restore the quantum bound. Mixing in product noise whose zero probability falls as a power of \(N\) keeps the violation unbounded with all probabilities positive, while uniform noise caps it. Finally, the Pompili–Weilenmann correlations obey information causality yet lie outside every level \(L\ge2\), so in these scenarios the two criteria are incomparable.
Transferable methods
Replacing the identity vector of a moment matrix
A level-\(L\) moment matrix only records inner products among vectors \(wR\) for products \(w\) of at most \(L\) projectors. The identity vector enters through its overlaps with the others and through its own norm, which sets the normalisation. Swapping it for a vector with the same overlaps but a much smaller norm raises every probability without changing any tested relation, provided positivity survives.
Let \(A_x,B_y\) be the outcome-zero projectors and \(R_N\) a reference vector with \(\langle R_N,A_xB_yR_N\rangle=\delta_{xy}\) and single-party moments 3. Choose \(S_L\) with \(\langle S_L,wR_N\rangle=\langle R_N,wR_N\rangle\) for every word \(1\le|w|\le L\), set \(C_L=\|S_L\|^2\), and let \(\varepsilon\) be the evaluation in which both parties always output 1, so \(\varepsilon(\mathbb I)=1\) and \(\varepsilon(w)=0\) for nonempty words. Define \[C_L\,\ell(p)=\langle R_N,pR_N\rangle+\bigl(C_L-\|R_N\|^2\bigr)\varepsilon(p).\]
For a polynomial \(f\) of degree at most \(L\), write \(f=c\mathbb I+g\) with \(\varepsilon(g)=0\). Expanding \(f^\dagger f=|c|^2\mathbb I+\bar cg+cg^\dagger+g^\dagger g\): the scalar term contributes \(|c|^2C_L\); the cross terms involve words of degree at most \(L\), where \(R_N\) may be replaced by \(S_L\); and multiplicativity gives \(\varepsilon(g^\dagger g)=0\), leaving \(\|gR_N\|^2\). Hence \[C_L\,\ell(f^\dagger f)=\|cS_L+gR_N\|^2\ge0,\] so the moment matrix \(\Gamma_{u,v}=\ell(u^\dagger v)\) is positive semidefinite and every probability is a reference moment divided by \(C_L\), giving \(s_L=1/C_L\).
To keep \(C_L\) independent of \(N\), \(S_L\) pairs shared modes \(\alpha\) of weight \(\lambda_\alpha\) with coefficients \(\nu_\alpha\). A product of projectors traces a walk through settings; closed walks require \(\sum_\alpha\nu_\alpha=1\), and open walks give polynomials in the marked weight that vanish at zero, so it suffices that \(\sum_\alpha\nu_\alpha\lambda_\alpha^k=0\) for \(1\le k\le L-1\). Lagrange coefficients at nodes \(\lambda_j=3t_j^2/(2T)\), with \(t_j=j-\tfrac12\) and \(T=(4L^3-L)/3\), evaluate such polynomials at zero, and splitting each coefficient over many equal-weight modes shrinks its contribution to the norm, giving \(C_L<185L^3\). Editorial check: exact rational arithmetic confirms the moment conditions for \(L\le8\), and the resulting \(C_L\) is 879 at \(L=2\) and about \(150L^3\) at \(L=8\).
How to reuse it
When a relaxation tests only low-degree moments, a witness can agree with a model on every tested degree and differ only at higher degree. Searching for such a substitute vector, rather than a whole new model, is an efficient way to show that a fixed level of a semidefinite hierarchy misses a property whose violation needs long operator products.
Equality certificates turned into database reads
An outcome pair that can occur only when the settings agree is a perfect certificate of equality. If Alice's setting encodes the position of the first 1 in her database, each certificate tells Bob of a rare event exactly, worth \(\log_2(1/r)\) bits, and a certificate rate that stays fixed as the number of settings grows yields information growing with \(\log N\).
For request \(i\), Bob outputs \(G_i=1\) with probability \(s\,r(1-r)^{i-1}\), and then \(R_i=1\) with certainty. Editorial step: since \(I(R_i:G_i)=\sum_gP(g)\,D\bigl(P_{R_i|g}\,\|\,P_{R_i}\bigr)\) and every term is nonnegative, the outcome \(G_i=1\) alone contributes \(P(G_i=1)\log_2(1/r)\). Summing over \(i\) gives \[\mathcal I\ge s\bigl[1-(1-r)^N\bigr]\log_2\frac1r,\] and \(1-(1-1/N)^N\ge1-e^{-1}\) at \(r=1/N\).
No nonsignalling correlation can do better than \(\log_2N+1\). Before the inputs arrive, the parties share a table of Alice outcomes at every setting and, for each, Bob responses at every setting. Alice sends her setting \(X\) and message \(M\); Bob reads his response from the table, which reproduces the original transcript. With \(T\) the shared table, \[\sum_jI(R_j:G_j)\le I(R:X,M\mid T)\le H(X,M)\le\log_2N+1.\]
Quantum correlations cannot keep the certificate rate fixed. If joint zeros are impossible at unequal settings, let Alice measure her projectors in order and stop at the first zero. Conditional on Bob's zero at setting \(i\), every earlier projector gives 1 with certainty and leaves the state undisturbed, so Alice's first zero falls at \(i\) with probability at least \(s_i\); these events are disjoint, giving \(\sum_is_i\le1\) and an average of at most \(1/N\). Editorial check: for the uniform first-one database the exact mutual information agrees with \((s/N)\log_2N!\approx s(\log_2N-\log_2e)\) to \(10^{-4}\), so at \(s=7/424\) the one-bit budget is exceeded only when \(\log_2N>62\).
How to reuse it
Hardy-type zero-probability constraints act as certificates, and a certificate rate that does not shrink with the number of inputs translates into a communication advantage. To show that a model respects a communication principle, it is often enough to bound how often such certificates can fire; sequential measurement arguments like the one above do this for quantum models.
Context & limitations
Why it matters. Information causality and the finite NPA levels turn out to be incomparable in a strong sense: no fixed level of semidefinite tests captures what one communicated bit can reveal, and the level needed grows with the number of settings. The same correlation gives a zero-error equality test using \(O(L^3)\) bits for any number of labels, where quantum correlations need \(n\) bits for \(N=2^n\) labels.
Limitations. The violation needs astronomically many settings. With the general bound, a violation is guaranteed only when \((1-e^{-1})\log_2N>185L^3\), about \(2^{2341}\) settings at level two; even the optimized level-two rate \(7/424\) needs \(\log_2N>62\) before its leading term exceeds one bit. The construction's rate, of order \(L^{-3}\), is not matched by the upper bound \(s_L\le(2L-3)^{-2}\) proved for exact certificates. Comparison with other relaxations is limited to the correlations of Pompili and Weilenmann, which lie outside the almost-quantum set yet obey information causality. The paper contains no statement about AI use.