Mastodawn

A/B Testing for Alien Life
이 논문은 외계 생명체 탐색에서 발생하는 인식론적 장벽, 특히 거짓 양성 문제를 다룬다. 저자는 확산적(비정보적) 사전 확률을 사용하는 기존 방법이 생명체 탐지에 매우 비관적인 결과를 초래함을 보이며, 이를 극복하기 위한 새로운 방법으로 두 그룹을 나누어 AB 테스트를 수행하는 전략을 제안한다. 이 방법은 생명체 존재율이 다른 두 그룹을 비교하면서 공통된 혼란 변수율을 유지하여, 상대적으로 적은 표본 수로도 강력한 생명체 탐지 증거를 얻을 수 있음을 시사한다. 다만, 이 AB 테스트 방식은 설계상 독특한 도전 과제를 동반한다.

https://arxiv.org/abs/2605.02969

#astrobiology #bayesianstatistics #abtesting #lifedetection #epistemology

The Catastrophic Consequences of Agnosticism for Life Searches and a Possible Workaround

Planned and ongoing searches for life, both biological and technological, confront an epistemic barrier concerning false positives - namely, that we don't know what we don't know. The most defensible and agnostic approach is to adopt diffuse (uninformative) priors, not only for the prevalence of life, but also for the prevalence of confounders. We evaluate the resulting Bayes factors between the null and life hypotheses for an idealized experiment with $N_{pos}$ positive labels (biosignature detections) among $N_{tot}$ targets with various priors. Using diffuse priors, the consequences are catastrophic for life detection, requiring at least ${\sim}10^4$ (for some priors ${\sim}10^{13}$) surveyed targets to ever obtain "strong evidence" for life. Accordingly, an HWO-scale survey with $N_{tot}{\sim}25$ would have no prospect of achieving this goal. A previously suggested workaround is to forgo the agnostic confounder prior, by asserting some upper limit on it for example, but we find that the results can be highly sensitive to this choice - as well as difficult to justify. Instead, we suggest a novel solution that retains agnosticism: by dividing the sample into two groups for which the prevalence of life differs, but the confounder rate is global. We show that a $N_{tot}=24$ survey could expect 24% of possible outcomes to produce strong life detections with this strategy, rising to $\geq50$% for $N_{tot}\geq76$. However, AB-testing introduces its own unique challenges to survey design, requiring two groups with differing life prevalence rates (ideally greatly so) but a global confounder rate.

arXiv.org