Zhang–Yeung Construction

• The Zhang–Yeung construction is an auxiliary-variable method that produced the first non-Shannon entropy inequality in a four-variable setting.
• It employs the Copy Lemma to extend joint distributions by introducing a copy variable that maintains marginal consistency and enforces conditional independence.
• Iterated applications and symmetrization of the technique refine entropy bounds, influencing network coding and the study of the entropic region.

Searching arXiv for recent and foundational papers on the Zhang–Yeung construction, Copy Lemma, and non-Shannon information inequalities. The Zhang–Yeung construction is the auxiliary-variable method that yielded the first genuinely non-Shannon entropy inequality in four random variables. In the modern formulation given in "Exploring the Entropic Region" (Csirmaz, 15 Sep 2025), the construction is distilled into the Copy Lemma: one enlarges a joint distribution by adjoining a suitably constrained “copy” variable, applies Shannon-type inequalities in the extended space, and then substitutes the copy constraints to obtain a new inequality on the original variables. In "Non-Shannon Information Inequalities in Four Random Variables" (Dougherty et al., 2011), this technique is presented as the key idea behind the first example of a non-Shannon information inequality, and as the prototype for many later inequalities in four variables.

1. Origin and defining role

In 1998, Zhang and Yeung gave the first example of a "non-Shannon" information inequality in four variables (Dougherty et al., 2011). The construction arose from the observation that although any unconstrained information inequality in three or fewer random variables can be written as a linear combination of instances of Shannon's inequality $I(A;B\mid C)\ge 0$, this ceases to hold for four variables (Dougherty et al., 2011). The Zhang–Yeung method therefore marks the point at which Shannon’s basic inequalities were shown to be insufficient for characterizing the entropic region.

The modern tutorial treatment identifies the construction not merely with a single inequality, but with a general method for obtaining new entropy inequalities (Csirmaz, 15 Sep 2025). The tutorial states that the Copy Lemma was distilled from the original Zhang–Yeung construction which produced the first non-Shannon inequality, and that its iterated version, effects of symmetrizations, and connections with polyhedral vertex enumeration are central to exploring the entropic region (Csirmaz, 15 Sep 2025).

A plausible implication is that the term “Zhang–Yeung construction” refers both to the original proof strategy for the first non-Shannon inequality and to the broader auxiliary-variable paradigm later abstracted as the Copy Lemma.

2. The Zhang–Yeung inequality

For four random variables $A,B,C,D$, the tutorial gives the original Zhang–Yeung inequality in an equivalent “Ingleton-style” form as (Csirmaz, 15 Sep 2025)

$[A,B,C,D] +(A,B\|C)+(A,C\|B)+(B,C\|A)\ge 0,$

where

$[A,B,C,D]=H(A,B\|C)+H(A,B\|D)+H(C,D)-H(A,B).$

In expanded Shannon-entropic terms this becomes

$$H(A,B\|C)+H(A,B\|D)+H(C,D)+H(A,C\|B)+H(B,C\|A)\ge H(A,B).$$

The 2011 exposition gives two equivalent formulations of the original Zhang–Yeung non-Shannon inequality (Dougherty et al., 2011). In its first published form it reads

$$2\,I(C;D)\le I(A;B)+I\bigl(A;C,D\bigr)+3\,I(C;D\mid A)+I(C;D\mid B).$$

Equivalently, one may write it in symmetric conditional-mutual-information form as

$$I(A;B)\le 2\,I(A;B\mid C)+I(A;C\mid B)+I(B;C\mid A)+I(A;B\mid D)+I(C;D).$$

These formulations encode the same underlying phenomenon: a valid entropy inequality for four variables that is not implied by Shannon’s monotonicity and submodularity alone (Dougherty et al., 2011). The tutorial further emphasizes that Zhang and Yeung’s original proof exhibited a particular four-variable distribution whose entropy vector violates all inequalities implied by Shannon alone, that is, it violates Ingleton, but nonetheless satisfies the Zhang–Yeung inequality (Csirmaz, 15 Sep 2025).

3. Probabilistic mechanism: the copy construction

The core probabilistic step is to enlarge an arbitrary joint distribution by adjoining a new variable with prescribed marginal and conditional-independence properties. In the tutorial presentation, one starts from an arbitrary joint distribution

$\xi=(A,B,C,D)$

of four discrete random variables, fixes a subset $D=\{C,D\}$, and forms a new random variable $C'$ which is a “copy” of $A,B,C,D$0 over the subset $A,B,C,D$1 in the sense that (Csirmaz, 15 Sep 2025):

1. the joint law of $A,B,C,D$2 is unchanged;
2. $A,B,C,D$3 has the same marginal law as $A,B,C,D$4;
3. $A,B,C,D$5 is conditionally independent of $A,B,C,D$6 given $A,B,C,D$7.

The resulting five-variable distribution $A,B,C,D$8 has an entropy vector that underlies the Copy Lemma, and from it one obtains the Zhang–Yeung inequality by a short Shannon-type argument (Csirmaz, 15 Sep 2025).

The 2011 account gives the distilled four-variable form of this idea as follows: given jointly distributed random variables $A,B,C,D$9, one can introduce a new variable $[A,B,C,D] +(A,B\|C)+(A,C\|B)+(B,C\|A)\ge 0,$0 such that (Dougherty et al., 2011):

• the marginal $[A,B,C,D] +(A,B\|C)+(A,C\|B)+(B,C\|A)\ge 0,$1 has exactly the same joint law as $[A,B,C,D] +(A,B\|C)+(A,C\|B)+(B,C\|A)\ge 0,$2;
• $[A,B,C,D] +(A,B\|C)+(A,C\|B)+(B,C\|A)\ge 0,$3.

One then says that $[A,B,C,D] +(A,B\|C)+(A,C\|B)+(B,C\|A)\ge 0,$4 is a $[A,B,C,D] +(A,B\|C)+(A,C\|B)+(B,C\|A)\ge 0,$5-copy of $[A,B,C,D] +(A,B\|C)+(A,C\|B)+(B,C\|A)\ge 0,$6 over $[A,B,C,D] +(A,B\|C)+(A,C\|B)+(B,C\|A)\ge 0,$7 (Dougherty et al., 2011). In words, $[A,B,C,D] +(A,B\|C)+(A,C\|B)+(B,C\|A)\ge 0,$8 “looks like” $[A,B,C,D] +(A,B\|C)+(A,C\|B)+(B,C\|A)\ge 0,$9 to the pair $[A,B,C,D]=H(A,B\|C)+H(A,B\|D)+H(C,D)-H(A,B).$0, yet $[A,B,C,D]=H(A,B\|C)+H(A,B\|D)+H(C,D)-H(A,B).$1 brings no new information about $[A,B,C,D]=H(A,B\|C)+H(A,B\|D)+H(C,D)-H(A,B).$2 once $[A,B,C,D]=H(A,B\|C)+H(A,B\|D)+H(C,D)-H(A,B).$3 is known (Dougherty et al., 2011).

This probabilistic coupling is the defining structural move of the Zhang–Yeung construction. A plausible implication is that the construction should be understood less as a specific symbolic derivation than as a method of extending the variable set so that ordinary Shannon inequalities become strong enough to imply genuinely new bounds on the original coordinates.

4. The Copy Lemma as an abstract formulation

The tutorial isolates the copy operation in polymatroidal form (Csirmaz, 15 Sep 2025). Let $[A,B,C,D]=H(A,B\|C)+H(A,B\|D)+H(C,D)-H(A,B).$4 be a finite set of indices, let $[A,B,C,D]=H(A,B\|C)+H(A,B\|D)+H(C,D)-H(A,B).$5 be a polymatroid on $[A,B,C,D]=H(A,B\|C)+H(A,B\|D)+H(C,D)-H(A,B).$6, write $[A,B,C,D]=H(A,B\|C)+H(A,B\|D)+H(C,D)-H(A,B).$7, and let $[A,B,C,D]=H(A,B\|C)+H(A,B\|D)+H(C,D)-H(A,B).$8. After adjoining a disjoint copy $[A,B,C,D]=H(A,B\|C)+H(A,B\|D)+H(C,D)-H(A,B).$9 of $$H(A,B\|C)+H(A,B\|D)+H(C,D)+H(A,C\|B)+H(B,C\|A)\ge H(A,B).$$0, with involution $$H(A,B\|C)+H(A,B\|D)+H(C,D)+H(A,C\|B)+H(B,C\|A)\ge H(A,B).$$1 swapping each $$H(A,B\|C)+H(A,B\|D)+H(C,D)+H(A,C\|B)+H(B,C\|A)\ge H(A,B).$$2 with its twin $$H(A,B\|C)+H(A,B\|D)+H(C,D)+H(A,C\|B)+H(B,C\|A)\ge H(A,B).$$3 and fixing everything in $$H(A,B\|C)+H(A,B\|D)+H(C,D)+H(A,C\|B)+H(B,C\|A)\ge H(A,B).$$4 and in $$H(A,B\|C)+H(A,B\|D)+H(C,D)+H(A,C\|B)+H(B,C\|A)\ge H(A,B).$$5, one defines an $$H(A,B\|C)+H(A,B\|D)+H(C,D)+H(A,C\|B)+H(B,C\|A)\ge H(A,B).$$6-copy of $$H(A,B\|C)+H(A,B\|D)+H(C,D)+H(A,C\|B)+H(B,C\|A)\ge H(A,B).$$7 over $$H(A,B\|C)+H(A,B\|D)+H(C,D)+H(A,C\|B)+H(B,C\|A)\ge H(A,B).$$8 to be a polymatroid $$H(A,B\|C)+H(A,B\|D)+H(C,D)+H(A,C\|B)+H(B,C\|A)\ge H(A,B).$$9 satisfying three conditions (Csirmaz, 15 Sep 2025):

• Extension: $$2\,I(C;D)\le I(A;B)+I\bigl(A;C,D\bigr)+3\,I(C;D\mid A)+I(C;D\mid B).$$0 for every $$2\,I(C;D)\le I(A;B)+I\bigl(A;C,D\bigr)+3\,I(C;D\mid A)+I(C;D\mid B).$$1.
• Symmetry on the copy: $$2\,I(C;D)\le I(A;B)+I\bigl(A;C,D\bigr)+3\,I(C;D\mid A)+I(C;D\mid B).$$2 for every $$2\,I(C;D)\le I(A;B)+I\bigl(A;C,D\bigr)+3\,I(C;D\mid A)+I(C;D\mid B).$$3.
• Conditional independence:

$$2\,I(C;D)\le I(A;B)+I\bigl(A;C,D\bigr)+3\,I(C;D\mid A)+I(C;D\mid B).$$4

The Copy Lemma then states that if $$2\,I(C;D)\le I(A;B)+I\bigl(A;C,D\bigr)+3\,I(C;D\mid A)+I(C;D\mid B).$$5 is entropic, almost-entropic, or linear-representable, then for every choice of $$2\,I(C;D)\le I(A;B)+I\bigl(A;C,D\bigr)+3\,I(C;D\mid A)+I(C;D\mid B).$$6 and every partition $$2\,I(C;D)\le I(A;B)+I\bigl(A;C,D\bigr)+3\,I(C;D\mid A)+I(C;D\mid B).$$7, the polymatroid $$2\,I(C;D)\le I(A;B)+I\bigl(A;C,D\bigr)+3\,I(C;D\mid A)+I(C;D\mid B).$$8 admits an entropic, almost-entropic, or linear $$2\,I(C;D)\le I(A;B)+I\bigl(A;C,D\bigr)+3\,I(C;D\mid A)+I(C;D\mid B).$$9-copy over $$I(A;B)\le 2\,I(A;B\mid C)+I(A;C\mid B)+I(B;C\mid A)+I(A;B\mid D)+I(C;D).$$0, respectively (Csirmaz, 15 Sep 2025).

The proof sketch in the tutorial distinguishes three cases (Csirmaz, 15 Sep 2025). In the entropic case, one starts with a joint distribution realizing $$I(A;B)\le 2\,I(A;B\mid C)+I(A;C\mid B)+I(B;C\mid A)+I(A;B\mid D)+I(C;D).$$1 and defines a new joint law on $$I(A;B)\le 2\,I(A;B\mid C)+I(A;C\mid B)+I(B;C\mid A)+I(A;B\mid D)+I(C;D).$$2 by

$$I(A;B)\le 2\,I(A;B\mid C)+I(A;C\mid B)+I(B;C\mid A)+I(A;B\mid D)+I(C;D).$$3

This leaves the marginal on $$I(A;B)\le 2\,I(A;B\mid C)+I(A;C\mid B)+I(B;C\mid A)+I(A;B\mid D)+I(C;D).$$4 unchanged, identifies the marginal on $$I(A;B)\le 2\,I(A;B\mid C)+I(A;C\mid B)+I(B;C\mid A)+I(A;B\mid D)+I(C;D).$$5 with that on $$I(A;B)\le 2\,I(A;B\mid C)+I(A;C\mid B)+I(B;C\mid A)+I(A;B\mid D)+I(C;D).$$6, and enforces the required conditional-independence relation (Csirmaz, 15 Sep 2025). The linear case is described via vector spaces $$I(A;B)\le 2\,I(A;B\mid C)+I(A;C\mid B)+I(B;C\mid A)+I(A;B\mid D)+I(C;D).$$7, a direct-sum complement of the subspace spanned by $$I(A;B)\le 2\,I(A;B\mid C)+I(A;C\mid B)+I(B;C\mid A)+I(A;B\mid D)+I(C;D).$$8, and a generic basis of the complement to represent $$I(A;B)\le 2\,I(A;B\mid C)+I(A;C\mid B)+I(B;C\mid A)+I(A;B\mid D)+I(C;D).$$9 (Csirmaz, 15 Sep 2025). The almost-entropic case follows by continuity (Csirmaz, 15 Sep 2025).

This abstraction is significant because it separates the structural content of the Zhang–Yeung method from the particulars of the original four-variable example. The construction thus becomes a reusable extension principle within the polymatroidal and entropy-functional setting.

5. Derivation of the inequality from one copy step

The tutorial gives a concise derivation of the Zhang–Yeung inequality from a single Copy Lemma application and one Shannon-type inequality (Csirmaz, 15 Sep 2025). One begins with $\xi=(A,B,C,D)$0 and applies the Copy Lemma with $\xi=(A,B,C,D)$1 and $\xi=(A,B,C,D)$2, obtaining an almost-entropic extension $\xi=(A,B,C,D)$3 on $\xi=(A,B,C,D)$4 such that:

• $\xi=(A,B,C,D)$5;
• $\xi=(A,B,C,D)$6;
• $\xi=(A,B,C,D)$7.

In the extended polymatroid $\xi=(A,B,C,D)$8, the following six-term Shannon-type inequality holds for any five-tuple of sets (Csirmaz, 15 Sep 2025):

$\xi=(A,B,C,D)$9

The tutorial labels this inequality as (MMRV) and notes that it is a Shannon-type inequality in the enlarged space (Csirmaz, 15 Sep 2025).

Substituting the copy constraints yields

$D=\{C,D\}$0

while $D=\{C,D\}$1 is unchanged (Csirmaz, 15 Sep 2025). The result is exactly

$D=\{C,D\}$2

which is the Zhang–Yeung inequality (Csirmaz, 15 Sep 2025).

The 2011 exposition presents a related simplified one-copy derivation using a variable $D=\{C,D\}$3 that is a $D=\{C,D\}$4-copy of $D=\{C,D\}$5 over $D=\{C,D\}$6 (Dougherty et al., 2011). It proves two intermediate inequalities, combines them, cancels the $D=\{C,D\}$7 terms, and obtains the symmetric form

$D=\{C,D\}$8

(Dougherty et al., 2011). That paper states that the Zhang–Yeung inequality can actually be derived from just one auxiliary variable (Dougherty et al., 2011).

Taken together, these accounts show that the Zhang–Yeung construction is a Shannon-type argument in an extended variable space, with the novelty residing in the auxiliary-variable constraints rather than in any new primitive inequality.

6. Iteration, symmetrization, and relation to entropy space

The tutorial states that the Copy Lemma may be applied repeatedly (Csirmaz, 15 Sep 2025). Starting from $D=\{C,D\}$9 on $C'$0, one copies a subset $C'$1 over $C'$2 to obtain $C'$3 on $C'$4, then chooses $C'$5 and $C'$6, produces $C'$7 on $C'$8, and so on (Csirmaz, 15 Sep 2025). Each step preserves the entropic, almost-entropic, or linear property, and a carefully chosen chain

$C'$9

can force new linear constraints among the original coordinates of $A,B,C,D$00, that is, new non-Shannon inequalities (Csirmaz, 15 Sep 2025).

Symmetrization plays a complementary role. At each copy step one can enforce the canonical involution $A,B,C,D$01 to be an exact automorphism of the new polymatroid (Csirmaz, 15 Sep 2025). More generally, any symmetry of the original $A,B,C,D$02 that leaves the “over” set $A,B,C,D$03 globally invariant can be lifted to the copy by intertwining with $A,B,C,D$04, and one may average over the entire symmetry group so as to reduce the number of independent variables in the final linear program that extracts inequalities (Csirmaz, 15 Sep 2025).

The geometric interpretation given in the 2011 paper situates these constructions in entropy space (Dougherty et al., 2011). Shannon inequalities alone carve out a polyhedral cone $A,B,C,D$05, while the Zhang–Yeung bound slices away a “pyramid” outside the Ingleton subcone, proving that the true entropic region $A,B,C,D$06 is strictly smaller than the Shannon cone (Dougherty et al., 2011). The same paper states that Zhang–Yeung was the first demonstration that the space of all quadruple-variable entropy vectors is not cut out by Shannon’s basic inequalities alone and that it showed $A,B,C,D$07 is not polyhedral (Dougherty et al., 2011).

A plausible implication is that the Zhang–Yeung construction is important not only for producing individual inequalities but also for providing a systematic mechanism for refining outer approximations to the entropic region.

7. Significance, limitations, and later perspective

The 2011 paper emphasizes several consequences of the Zhang–Yeung construction (Dougherty et al., 2011). It gave the first non-Shannon inequality, spawned infinite families of inequalities, and motivated work on network coding, including improved upper bounds on coding capacity as demonstrated on specific networks (Dougherty et al., 2011). It also notes that the auxiliary-variable method highlights how new conditional-independence assumptions allow one to bootstrap Shannon’s inequalities into genuinely new information laws (Dougherty et al., 2011).

The 2025 tutorial places the method among three known approaches for obtaining new entropy inequalities and compares their strengths and limitations (Csirmaz, 15 Sep 2025). It states that another method, derived from the principle of maximum entropy, has the Copy Lemma as a special case, but that none of the two presented variants is known to generate more inequalities than the iterated Copy Lemma (Csirmaz, 15 Sep 2025). It further states that the Ahlswede–Körner method employs a hidden application of the Copy Lemma, that the underlying lemma alone cannot generate new inequalities, and that this makes the Ahlswede–Körner method strictly weaker than the Copy Lemma (Csirmaz, 15 Sep 2025).

The same tutorial concludes with open questions and research problems (Csirmaz, 15 Sep 2025). This suggests that, despite its foundational status, the Zhang–Yeung construction remains part of an active research program concerned with the structure of the entropic region, the power of iterated copying, and the comparative strength of auxiliary-variable methods.

A common misconception is that the Zhang–Yeung inequality is merely an exotic reformulation of Shannon’s inequalities. The cited accounts reject this directly: it cannot be derived from ordinary Shannon-type inequalities alone, although its proof becomes a pure Shannon-type argument after enlarging the variable set and imposing copy-lemma constraints (Dougherty et al., 2011). Another misconception is that the original construction is exhausted by the single 1998 inequality. The later tutorial treatment instead presents it as the source of a general Copy Lemma framework whose iterated and symmetrized forms yield many additional non-Shannon inequalities (Csirmaz, 15 Sep 2025).