---
title: Information Conservation Score (ICS)
url: https://www.emergentmind.com/topics/information-conservation-score-ics
type: topic
---

# Information Conservation Score (ICS)

Searching arXiv for the specified paper and closely related context.
Information Conservation Score (ICS), also called conserved active information $I^\oplus$, is a symmetric extension of active information that compares a baseline law $q$ on a search space $X$ with an informed or target-biased law $p$. Its purpose is to quantify net information gain or loss across the entire search space rather than only on a designated target event. In the formulation introduced in "Conserved active information" [2512.21834], ICS is designed to respect No-Free-Lunch conservation by pairing the usual active information on an event $A$ with the corresponding quantity on its complement $A^c$. This yields a measure that distinguishes regimes in which increased target probability reflects greater global disorder from regimes in which sufficiently strong bias imposes order.

## 1. Formal definition and notation

Let $X$ be a finite, or more generally measurable, search space. Let $q$ denote the baseline or prior law on $X$, and let $p$ denote the informed or target-biased law. For any event $A \subseteq X$, write
$$
q(A)=\sum_{x\in A}q(x), \qquad p(A)=\sum_{x\in A}p(x),
$$
with $\log$ taken as $\ln$ or any other convenient base [2512.21834].

The usual active information of an event $A$ is
$$
I^+(A)=\log\frac{p(A)}{q(A)}.
$$
This quantity records how much more, or less if negative, probability mass the informed search $p$ places on $A$ relative to the blind search $q$.

ICS is defined by symmetrizing this event-level quantity across the partition $\{A,A^c\}$:
$$
I^\oplus(A)=I^+(A)-I^+(A^c)
=\log\frac{p(A)}{q(A)}-\log\frac{p(A^c)}{q(A^c)}
=\log\frac{p(A)\,q(A^c)}{q(A)\,p(A^c)}.
$$
The defining feature is that $I^\oplus(A)$ does not treat increased mass on $A$ in isolation. It explicitly encodes the compensating change on $A^c$, so the target event is evaluated together with the rest of the search space.

## 2. Symmetric extension and conservation principle

The motivation for $I^\oplus$ is that traditional active information measures gain or loss of bias on $A$ alone, while saying nothing about what happens on $A^c$. The conserved construction addresses the longstanding critique that an apparent advantage on a target may be reported without tracking the complementary redistribution of probability mass. In the exposition, this is tied directly to No-Free-Lunch-style reasoning: if an algorithm does better on $A$, it must do worse on $A^c$, so the two terms are paired in the symmetric combination $I^\oplus(A)=I^+(A)-I^+(A^c)$ [2512.21834].

The basic conservation statement is the partition-wise identity
$$
I^\oplus(A)+I^\oplus(A^c)=0.
$$
By direct substitution,
$$
\Bigl[\log\frac{p(A)}{q(A)}-\log\frac{p(A^c)}{q(A^c)}\Bigr]
+
\Bigl[\log\frac{p(A^c)}{q(A^c)}-\log\frac{p(A)}{q(A)}\Bigr]
=0.
$$
The same exposition introduces local densities
$$
i^\oplus(x):=\log\frac{q(x)}{p(x)},
$$
and states that summing these over the entire space using counting measure gives a total that is zero if and only if $q\equiv p$. The proof sketch further notes that the global sum $\sum_{x\in X}\log\frac{q(x)}{p(x)}$ vanishes if and only if $\prod_x(q(x)/p(x))=1$, equivalently $\sum_x\log q(x)=\sum_x\log p(x)$; in particular, if $q=p$ point-wise then the total is zero.

The significance of this construction is not merely algebraic. It formalizes the claim that target-specific improvement cannot be evaluated independently of the compensating reallocation elsewhere in the space.

## 3. Global formulation and total information

Alongside the event-wise form, the exposition defines a global conserved active information comparing the full laws $q$ and $p$. If $\mu$ is any dominating measure, such as counting measure on a finite $X$, and $q(x)=dq/d\mu$, $p(x)=dp/d\mu$, then
$$
I^\oplus(q\|p)=\int_X \log\frac{q(x)}{p(x)}\,\mu(dx)=H(p)-H(q),
$$
where the total information is
$$
H(r)=\int_X (-\log r(x))\,\mu(dx).
$$
Thus the global quantity is the un-weighted integral of $\log(q/p)$ [2512.21834].

This formulation is distinct from KL divergence. The abstract explicitly states that $I^\oplus$ reveals regimes hidden from KL divergence, particularly regimes in which strong knowledge reduces global disorder. The global identity $I^\oplus(q\|p)=H(p)-H(q)$ makes that contrast operational: the sign of $I^\oplus$ tracks whether the informed law has increased or decreased the total information relative to the baseline.

The exposition also states that one may switch freely between the event-wise form $I^\oplus(A)$ and the global form $I^\oplus(q\|p)$. In practice, the event-wise form isolates a target event, while the global form compares the full laws directly.

## 4. Canonical sign regimes

Two elementary examples are used to characterize the sign structure of $I^\oplus$ and to separate disorder-increasing mild knowledge from order-imposing strong knowledge [2512.21834].

For the Bernoulli-baseline example, let $X=\{0,1\}$, baseline $q=\mathrm{Bern}(r)$ with $q(1)=r$ and $q(0)=1-r$, and informed law $p=\mathrm{Bern}(s)$ with $p(1)=s$ and $p(0)=1-s$. Then
$$
I^+(\{1\})=\log\frac{s}{r}, \qquad I^+(\{0\})=\log\frac{1-s}{1-r},
$$
and the conserved information on the event $\{1\}$ is
$$
I^\oplus(\{1\})=\log\frac{s}{r}-\log\frac{1-s}{1-r}
=\log\frac{s(1-r)}{r(1-s)}.
$$
The stated sign regimes are the following. If $s<r$, then $I^\oplus(\{1\})<0$: the algorithm hurts $1$ but orders the rest. If $r<s<1-r$, then $I^\oplus(\{1\})<0$: this is a mild bias toward $1$ at the cost of overall disorder. If $s>1-r$, then $I^\oplus(\{1\})>0$: this is a strong bias, termed the “jackpot” regime, and corresponds to net infusion of order.

For the uniform-baseline example, let $|X|=N$ and $q(x)=1/N$. Choose a target set $T\subset X$ of size $|T|=m$, so the baseline target mass is denoted $p:=m/N\ll 1/2$. The example then denotes by $q$ the total mass that the informed law assigns to $T$, yielding
$$
I^\oplus(T)=\log\frac{q}{p}-\log\frac{1-q}{1-p}
=\log\frac{q(1-p)}{p(1-q)}.
$$
Under this notation, the sign regimes are: if $q<p$ or $q>1-p$, then $I^\oplus(T)>0$, described respectively as “target harder but system more ordered” or “jackpot”; if $p<q<1-p$, then $I^\oplus(T)<0$, described as “mild” knowledge with more disorder globally; if $q\in\{p,1-p\}$, then $I^\oplus(T)=0$.

These examples are used to show that increasing target probability does not, by itself, determine whether information has been conserved, merely redistributed, or externally injected.

## 5. Illustrative domains

The exposition gives two application classes that instantiate the same conservation logic in very different settings [2512.21834].

For Markov chains, let $G=(X,E)$ be a finite, connected, $d$-regular graph. The simple random-walk transition matrix is
$$
P(x,y)=
\begin{cases}
1/d,&(x,y)\in E,\\
0,&\text{otherwise}.
\end{cases}
$$
This chain is stated to be irreducible, aperiodic, with unique stationary distribution $\pi(x)=1/N$. If the baseline is an initial law $q$ that places very small mass on a target $T$, and the informed law is the stationary distribution $\pi$, then $q(T)\ll \pi(T)=|T|/N$ is identified as the mild regime: the probability of $T$ is raised, but global disorder is increased, hence $I^\oplus(T)<0$.

For cosmological fine-tuning, the setup consists of a family of physical-constant hypotheses $\{H_\xi\}_{\xi\in\Xi}$, each with likelihood $\mathbb{P}(T\mid H_\xi)$ that the universe is life-permitting, together with a maxent prior $q$ on $\Xi$ and a data-conditioned posterior $p$ that collapses almost entirely onto the small region $\ell\subset\Xi$ that permits life. The tuning probability is
$$
q(T)=\int_\Xi \mathbb{P}(T\mid H_\xi)\,q(d\xi)\ll 1,
$$
whereas the posterior satisfies $p(T)\approx 1$. The exposition then gives
$$
I^+(T)=\log\frac{p(T)}{q(T)}=-\log q(T)\gg 0,
$$
and
$$
I^\oplus(T)=\log\frac{p(T)(1-q(T))}{q(T)(1-p(T))}\approx -\log q(T)>0.
$$
Accordingly, $I^\oplus\gg 0$ is interpreted there as a quantitative witness of “injecting” vast amounts of new information to explain why a life-permitting universe is observed.

These examples are not presented as a single applied methodology. Rather, they show that the same score can be instantiated in stochastic processes, search-style settings, and inference problems involving sharply concentrated posteriors.

## 6. Interpretation, scope, and common misunderstandings

The exposition gives a three-way interpretation of ICS. If $I^\oplus(A)>0$, then biasing toward $A$ net orders all of $X$; in that sense, external information has been injected beyond mere reshuffling. If $I^\oplus(A)<0$, then making $A$ more probable has increased global disorder, corresponding to a No-Free-Lunch redistribution. If $I^\oplus(A)=0$, then the redistribution is neutral: mass has been reshuffled without creating or destroying information [2512.21834].

A common misunderstanding is to treat any increase in $I^+(A)$ as evidence of net information gain. The conserved formulation rejects that inference. The target event and its complement must be considered jointly, because the same increase in $p(A)$ can correspond either to mild knowledge with more disorder globally or to strong knowledge that imposes order. This is precisely the distinction the Bernoulli and uniform-baseline examples are used to formalize.

The exposition also assigns ICS a methodological role in several domains. In search and optimization, it can be used to certify whether an algorithm’s edge on a particular problem is “free,” meaning a reshuffling within the same information budget, or whether it truly requires extra information. In statistical estimation, one can compute $I^\oplus$ for the event “parameter in interval” to judge whether a biased estimator has covertly injected external knowledge. In machine learning, one can ask whether fine-tuning a model on a downstream task creates new information or merely redistributes the pre-trained weights.

The central identity
$$
I^\oplus(A)+I^\oplus(A^c)=0
$$
therefore functions as more than a formal symmetry. It is the operational statement of the Information Conservation Score: a criterion for distinguishing genuine infusion of problem-specific information from relabeling or redistribution of existing uncertainty.

Source: https://www.emergentmind.com/topics/information-conservation-score-ics