— Rawls's Difference Principle
The intuition and its axiom
Rawls (1971) proposed a theory of justice built around a single structural principle:
inequalities in a just society are permissible only if they benefit the least
advantaged members. A distribution $D$ is just only if
$\Delta W_{\min}(D) \geq 0$ — the welfare of the worst-off does not decrease.
This is the Difference Principle.
The foundation Rawls provided was the veil of ignorance: a thought experiment in
which rational agents choose principles of justice without knowing which position
in society they will occupy. Behind the veil, not knowing whether you will be
born the least or most advantaged, you would choose to maximize the position of
the worst-off as a form of rational insurance. The Difference Principle follows
from the self-interest of epistemically constrained agents.
The problem is that the veil of ignorance is an axiom, not a theorem. Rawls
assumed it as the foundation and derived the Difference Principle as a consequence.
This means the Difference Principle is only as secure as the plausibility of the
veil as a thought experiment — and critics from both the utilitarian and libertarian
traditions have found it wanting.
— The paper's claim
The Difference Principle does not need the veil of ignorance. It is a geometric
consequence of standard Lagrangian optimization applied to Shannon's welfare
function under equal contexts. Rawls's intuition was correct; his foundation
was weaker than necessary.
— Theorem (Shannon-Rawls)
The Difference Principle as KKT condition
Let $V(D,C)$ be Shannon's welfare function satisfying properties 1–6. Consider
the optimization problem
$$\max \sum_{e \in E} V(D_e,\, C) \quad \text{subject to} \quad C_e = C \;\; \forall e$$
(equal contexts: every entity operates under the same context $C$). The
Karush-Kuhn-Tucker (KKT) conditions for this problem require
$$\frac{\partial V}{\partial D_e} = \lambda \quad \forall e$$
— the same Lagrange multiplier for every entity. By strict concavity of $V$
(property 2), this condition uniquely implies $D_e^* = D^*$ for all $e$.
The optimal distribution is equal across entities, and therefore
$$\Delta W_{\min}(D^*) \geq 0.$$
The Difference Principle is not an axiom. It is what happens when you apply
standard Lagrangian optimization to Shannon's welfare function under equal contexts.
No veil of ignorance required.
— What the result means
Three implications
For Rawlsian political philosophy
Rawls's intuition was correct but its foundation was weaker than necessary. The
veil of ignorance served as a thought experiment designed to produce a result
that follows from the mathematics alone, without needing the epistemic fiction.
The Difference Principle survives the removal of its axiom — and is strengthened
by surviving it. A result derived from standard optimization under well-defined
properties is more robust than one derived from a hypothetical deliberative
procedure, because it does not depend on the contested psychology of rational
agents behind the veil.
For welfare economics
The Bergson-Samuelson social welfare function and the Rawlsian maximin are not
competing axiom systems requiring a philosophical choice between them. They are
different special cases of the same Shannon structure. Cooperation geometry
($\Omega_1$, full inclusion) under equal contexts produces Rawlsian outcomes
— the Difference Principle emerges automatically. Extraction geometry ($\Omega_3$,
partial inclusion with excluded entities) produces Pareto-inferior outcomes with
concentrated welfare loss among excluded entities — the exact pathology that
utilitarian social welfare functions have been criticized for permitting.
The geometry unifies the debate.
For AI alignment
An AI system trained to maximize $\sum_e V(D_e, C)$ under equal context
constraints will automatically protect its least-advantaged users — without being
explicitly instructed to do so, and without the system having any representation
of Rawlsian principles. This is alignment geometry, not alignment constraint.
The protection of the worst-off is not a rule layered on top of the optimization;
it is a structural property of the optimization itself. This has direct implications
for how aligned systems should be designed: not by adding Rawlsian side-constraints
to misaligned objectives, but by using Shannon's welfare function with full
entity inclusion from the start.
— The geometry
What changes when contexts are unequal
The Shannon-Rawls theorem requires equal contexts. What happens when contexts
are unequal — when different entities operate under structurally different
circumstances, as they do in any real society or system?
Equal contexts
Shannon optimization converges to equal distribution $D^* = D^*_e$ for all
$e$. The Difference Principle emerges as a KKT condition. Rawlsian justice
is the geometric solution.
Unequal contexts
Entities with lower context $C_e$ have steeper welfare gradients by strict
concavity of $V$. An aligned system prioritizes improving their contexts
— not by constraint, but because marginal welfare gain is highest there.
The key result under unequal contexts: strict concavity of $V$ guarantees that
marginal welfare returns are highest for the lowest-context entities. An aligned
system — one maximizing $\sum_e V(D_e, C_e)$ — will naturally direct resources
toward the worst-off, not because it is instructed to, but because the gradient
is steepest there.
This is not a separate argument for redistribution. It is the same geometry as
the equal-context case, generalized. The Difference Principle (protect the
worst-off) is the equal-context limit of a more general result: aligned systems
always direct their optimization pressure toward the steepest welfare gradients,
which by concavity are always at the bottom of the context distribution.
— The unified picture
Rawls gave us the intuition in the equal-context case. Shannon gives us the
geometry in both cases. The veil of ignorance was a philosophical instrument
for reaching a mathematical result that the mathematics produces on its own.
The result is now detached from its scaffold — and stands on firmer ground for it.