Extended Information Causality Overview

• Extended information causality is a family of generalizations that broadens the original principle by incorporating varied input correlations, higher-dimensional alphabets, and non-standard random-access setups.
• The extended frameworks yield tighter constraints on nonlocal correlations, leading to improved Bell inequalities, channel-capacity formulations, and bounds consistent with quantum limits like the Tsirelson threshold.
• These approaches introduce innovations such as correlated inputs and redundancy-corrected measures, addressing limitations of the original formulation and guiding both theoretical insights and experimental testing.

Searching arXiv for recent and foundational papers on extended information causality and closely related formulations. Extended Information Causality denotes a family of strengthenings and reformulations of the information-causality principle introduced by Pawłowski et al. as an operational constraint on nonlocal correlations. In its original form, the principle states that if Alice sends $m$ classical bits to Bob, then Bob’s total accessible information about Alice’s previously unknown data cannot exceed $m$; in the standard random-access-code setting this is expressed as

$$I \equiv \sum_{K=0}^{N-1} I(a_K:\beta \mid b=K)\le m.$$

Subsequent work extended this idea in several distinct directions: correlated inputs, distributed functions beyond index retrieval, higher-dimensional alphabets, redundant-information quantifiers that avoid random-access-code asymmetries, multipartite communication tasks, channel-capacity formulations without concatenation, and communication-complexity lower bounds. Across these variants, the common objective is to constrain physically admissible correlations more sharply than no-signaling while remaining valid for quantum theory (0905.2292, Yu et al., 2022, Jain et al., 2023, Pollyceno et al., 2022, Jain et al., 1 Jun 2026).

1. Original principle and its operational meaning

The original information-causality scenario is a bipartite random-access task. Alice holds a string of independent random bits or dits, Bob receives an index specifying which component he should guess, Alice sends a classical message through a channel of capacity $\mathcal C$ or $m$ bits, and the parties may share arbitrary no-signaling resources. The foundational constraint is

$\sum_{i=0}^{n-1} I(a_i;g\mid b=i)\le \mathcal C,$

or, in the binary formulation of the original paper,

$I=\sum_{K=0}^{N-1} I(a_K:\beta\mid b=K)\le m.$

Its physical content is that classical communication bounds Bob’s total accessible information about Alice’s data, even in the presence of pre-shared nonlocal correlations (0905.2292).

This formulation strictly strengthens no-signaling. When $m=0$, the inequality reduces to $I=0$, so Bob cannot gain any information about Alice’s private data without communication. For $m=1$, Bob may recover one chosen bit perfectly, but not more than one bit’s worth of information in total. The original paper showed that both classical and quantum theory satisfy the principle, whereas maximally nonlocal no-signaling resources such as PR boxes violate it. In the two-bit, one-bit-communication protocol, a PR box allows Bob to recover either $m$0 or $m$1 perfectly, yielding $m$2, in contradiction with $m$3 (0905.2292).

The foundational significance of information causality in quantum foundations came from its relation to Bell nonlocality. In the isotropic family

$m$4

the recursive protocol of Pawłowski et al. implies that for sufficiently large depth $m$5, information causality is violated whenever $m$6, equivalently $m$7. This reproduces the quantum Tsirelson threshold on that family without invoking Hilbert-space formalism directly (0905.2292).

2. Why the original formulation was extended

The original statement is operationally clear but structurally narrow. Several later papers identify the same limitations from different angles. One limitation is that the usual formulation presumes independent inputs on Alice’s side. Another is that it is tied to the random-access-code success structure, with a one-to-one relation between Alice’s data and Bob’s target queries. A third is technical: many derivations of Bell constraints relied on concatenation protocols that are cumbersome and scenario-specific (Miklin et al., 2021, Jain et al., 2023, Miklin et al., 10 Feb 2026).

A particularly explicit criticism appears in the redundant-information reformulation. There the random-access-code viewpoint is described as too rigid and task-specific, because Bob’s information may be distributed across several indicators $m$8 in a way that is not naturally represented by a bit-by-bit retrieval game. The same paper also argues that the random-access-code criterion can be biased by the chosen labeling or symmetry of the box: in the $m$9 Bell scenario, the original IC detects some extremal boxes but not symmetry-equivalent ones, even though the underlying “potential information” should be the same (Yu et al., 2022).

An entropic reformulation sharpened the same point from a different direction. It derived a generalized inequality

$$I \equiv \sum_{K=0}^{N-1} I(a_K:\beta \mid b=K)\le m.$$0

which does not assume the $$I \equiv \sum_{K=0}^{N-1} I(a_K:\beta \mid b=K)\le m.$$1 are independent; the familiar information-causality inequality is then recovered when $$I \equiv \sum_{K=0}^{N-1} I(a_K:\beta \mid b=K)\le m.$$2. This showed that the dependence on independent inputs is not intrinsic to the basic informational logic of the principle (Al-Safi et al., 2011).

These observations motivate the phrase “extended information causality” in the literature. It does not refer to a single canonical generalization. Rather, it denotes a set of related proposals that preserve the communication-budget intuition while broadening the admissible tasks, input distributions, or information measures.

3. Correlated-input and non-random-access-code formulations

One influential extension modifies the principle to allow correlated inputs on Alice’s side. For the two-bit case, the reformulated statement is

$$I \equiv \sum_{K=0}^{N-1} I(a_K:\beta \mid b=K)\le m.$$3

This has the same operational interpretation as the original principle, but the second term is conditioned on the previously accounted-for variable $$I \equiv \sum_{K=0}^{N-1} I(a_K:\beta \mid b=K)\le m.$$4. In the general Boolean-function setting, with $$I \equiv \sum_{K=0}^{N-1} I(a_K:\beta \mid b=K)\le m.$$5 and an ordering $$I \equiv \sum_{K=0}^{N-1} I(a_K:\beta \mid b=K)\le m.$$6 of Bob’s inputs, the extended form becomes

$$I \equiv \sum_{K=0}^{N-1} I(a_K:\beta \mid b=K)\le m.$$7

For the binary index function $$I \equiv \sum_{K=0}^{N-1} I(a_K:\beta \mid b=K)\le m.$$8, this reduces exactly to the two-bit correlated-input formula above (Jain et al., 2023, Jain et al., 1 Jun 2026).

A second line of extension removes the random-access-code framework more radically by replacing the sum of target-bit mutual informations with a redundancy-corrected quantity. For Bob variables $$I \equiv \sum_{K=0}^{N-1} I(a_K:\beta \mid b=K)\le m.$$9,

$\mathcal C$0

and in the bivariate case

$\mathcal C$1

Here $\mathcal C$2 is a redundant-information term defined using the construction of Harder, Salge, and Polani. In that framework one projects $\mathcal C$3 onto the convex set $\mathcal C$4 by a Kullback–Leibler minimization and defines

$\mathcal C$5

The resulting quantity satisfies

$\mathcal C$6

which yields $\mathcal C$7 for quantum correlations via the Markov chain $\mathcal C$8 and the bound $\mathcal C$9 (Yu et al., 2022).

These two reformulations differ technically, but they share a common shift of perspective. Information causality is no longer tied to independent database bits and bit-by-bit retrieval. It becomes a statement about the amount of effective, non-double-counted information that Bob can obtain about a family of function values or observables under a fixed communication budget.

4. Bell inequalities and approximation of the quantum set

A central use of extended information causality is the derivation of explicit constraints on observable correlators. One method begins by choosing an encoding and decoding protocol, lower-bounding mutual informations with Fano-type inequalities, expressing guessing probabilities through Bell correlators, and then taking a low-capacity limit of the classical channel. In arbitrary bipartite Bell scenarios this procedure yields polynomial inequalities implied by information causality; in the $m$0 scenario, the standard construction reproduces the Uffink inequality

$m$1

where $m$2 (Jain et al., 2023).

The correlated-input extension produces a tighter one-parameter family in the same $m$3 scenario: $m$4 Setting $m$5 recovers Uffink. In the specific three-box convex-mixture subspace studied in that work, the limit $m$6 recovers the Landau inequality, which exactly characterizes the boundary of the quantum set in the relevant maximally mixed-marginal slice (Jain et al., 2023).

The same program extends to broader Bell families. For $m$7, extended IC yields stronger quadratic inequalities than the original IC-derived ones, and it improves the analytical upper bound on the Collins–Gisin family to

$m$8

for $m$9. This improves the earlier bound

$\sum_{i=0}^{n-1} I(a_i;g\mid b=i)\le \mathcal C,$0

derived from the original formulation (Jain et al., 1 Jun 2026).

Other reformulations reach similar objectives by different routes. A noisy-circuit perspective uses the Evans–Schulman signal-decay theorem to obtain

$\sum_{i=0}^{n-1} I(a_i;g\mid b=i)\le \mathcal C,$1

where $\sum_{i=0}^{n-1} I(a_i;g\mid b=i)\le \mathcal C,$2 is the coding-noise parameter. This implies broad classes of Tsirelson-type inequalities, including the CHSH/Tsirelson bound for $\sum_{i=0}^{n-1} I(a_i;g\mid b=i)\le \mathcal C,$3, and leads to a no-go theorem for reliable nonlocal computation in sufficiently large noisy circuits (Hsu et al., 2010).

A channel-capacity formulation removes concatenation entirely. For a $\sum_{i=0}^{n-1} I(a_i;g\mid b=i)\le \mathcal C,$4-ary symmetric channel, with

$\sum_{i=0}^{n-1} I(a_i;g\mid b=i)\le \mathcal C,$5

the criterion becomes

$\sum_{i=0}^{n-1} I(a_i;g\mid b=i)\le \mathcal C,$6

This reproduces all previously known unbiased-error concatenation bounds and improves some of them. It also applies to Bell scenarios that were previously difficult to treat by concatenation; in the $\sum_{i=0}^{n-1} I(a_i;g\mid b=i)\le \mathcal C,$7 scenario the optimized noisy-channel method yields the bound $\sum_{i=0}^{n-1} I(a_i;g\mid b=i)\le \mathcal C,$8, close to the quantum bound $\sum_{i=0}^{n-1} I(a_i;g\mid b=i)\le \mathcal C,$9 (Miklin et al., 2021).

5. Higher-dimensional, multipartite, and monogamous extensions

The first explicit higher-dimensional extension replaced bits by dits. In that $I=\sum_{K=0}^{N-1} I(a_K:\beta\mid b=K)\le m.$0-ary random-access setting, Alice holds $I=\sum_{K=0}^{N-1} I(a_K:\beta\mid b=K)\le m.$1, Bob must guess $I=\sum_{K=0}^{N-1} I(a_K:\beta\mid b=K)\le m.$2, and Alice sends $I=\sum_{K=0}^{N-1} I(a_K:\beta\mid b=K)\le m.$3 dits. The principle becomes

$I=\sum_{K=0}^{N-1} I(a_K:\beta\mid b=K)\le m.$4

A generalized PR box achieves perfect success in the basic $I=\sum_{K=0}^{N-1} I(a_K:\beta\mid b=K)\le m.$5 task, implying $I=\sum_{K=0}^{N-1} I(a_K:\beta\mid b=K)\le m.$6 for $I=\sum_{K=0}^{N-1} I(a_K:\beta\mid b=K)\le m.$7, so it violates the extended principle maximally. For isotropic $I=\sum_{K=0}^{N-1} I(a_K:\beta\mid b=K)\le m.$8 boxes, comparison with macroscopic locality showed that for $I=\sum_{K=0}^{N-1} I(a_K:\beta\mid b=K)\le m.$9 the IC threshold coincides with the $m=0$0 boundary at $m=0$1, but for $m=0$2 the IC threshold moves below the macroscopic-locality threshold. Hence there are macroscopically local correlations that violate the $m=0$3-ary extension of IC (Cavalcanti et al., 2010).

Multipartite scenarios required a different type of extension, because the original bipartite formulation can fail to detect even extremal stronger-than-quantum multipartite correlations. In the genuinely multipartite version, $m=0$4 senders communicate with one receiver, who must guess functions $m=0$5. The resulting inequalities are necessary criteria for multipartite information causality, hold for all quantum resources, and exclude some stronger-than-quantum non-signaling boxes. In the tripartite concatenated setting they imply the quadratic criterion

$m=0$6

which is strengthened further either by multiple-copy concatenation or by replacing message entropy with noisy-channel capacities $m=0$7 (Pollyceno et al., 2022).

A related multi-receiver extension yields a monogamy statement. If Alice broadcasts one classical bit to multiple Bobs, then even in the presence of multipartite nonlocal resources the total information gain obeys

$m=0$8

The interpretation given in that work is that multipartite nonlocal resources can be used only for information splitting, not for amplification of the total accessible information. This trade-off generates CHSH-type and higher-order Bell monogamy relations (Hsu, 2011).

6. Conceptual status, limitations, and open structure

Extended information causality is stronger than the original principle in several concrete senses, but its present status is not that of a closed characterization of the quantum set. The redundant-information reformulation closes a notable gap in the $m=0$9 Bell scenario: for the slice $I=0$0, where the old IC detected violations only beyond the Tsirelson bound, the new redundancy-corrected quantity recovers the quantum boundary exactly, numerically. However, for the local-deterministic slice $I=0$1, the old and new IC boundaries coincide and both still leave a gap to the quantum boundary (Yu et al., 2022).

The same qualified picture appears in more recent work on random-access codes. Extended IC yields tighter Bell inequalities and improved nonlocality bounds, yet in entanglement-assisted random-access codes the stronger principle does not improve the best known theory-independent bound on the winning probability. In the isotropic and symmetric setup studied there,

$I=0$2

and the paper proves that correlated inputs do not increase the relevant extended-IC information term: the optimum is attained for independent, uniformly distributed inputs and the uniform-error channel. This suggests that the existing IC bounds are already optimal for that class of protocols (Jain et al., 1 Jun 2026).

General-protocol studies likewise indicate that “more Bell violation” and “more accessible information gain” are not identical orderings. Numerical analysis of the most general binary two-level, two-setting protocols found that quantum-realizable schemes respect information causality, but the boundary for the information-causality bound does not agree with the Tsirelson boundary in general. In some cases the maximal information gain is not attained at the point of maximal CHSH violation; one can even saturate $I=0$3 by making one effective subchannel noiseless and the other completely noisy (Yu et al., 2013).

Another unresolved issue is conceptual rather than operational. Several papers argue that IC is best understood not as a specialized game but as a consequence of deeper informational structure. One entropic formulation derives IC from the existence of an entropy satisfying consistency with classical entropy and monotonicity under local evolution with ancillas, while a thermodynamic line of work derives information-causality-type statements from monotonicity of Bregman divergence and sufficiency conditions on state spaces (Al-Safi et al., 2011, Harremoës, 2020). This suggests that the long-term significance of extended information causality may lie not only in sharper Bell inequalities, but also in clarifying which structural properties of information theory are genuinely responsible for the observed boundary between quantum and post-quantum correlations.