1: \begin{abstract}

2:   A comparison-based search algorithm lets a user find a target item $t$ in a database by answering queries of the form, ``Which of items $i$ and $j$ is closer to $t$?''

3:   Instead of formulating an explicit query (such as one or several keywords), the user navigates towards the target via a sequence of such (typically noisy) queries.

4:   We propose a scale-free probabilistic oracle model called $\gamma$-CKL for such similarity triplets $(i,j;t)$, which generalizes the CKL triplet model proposed in the literature. %\cite{tamuz2011adaptively} 

5:   % indep control over power and dim

6:   The generalization affords independent control over the discriminating power of the oracle and the dimension of the feature space containing the items.

7:   %and show that it compares favourably with the state of the art when scaling to high embedding dimensions.

8:   %Using the properties of a scale-free oracle, 

9:   We develop a search algorithm with provably exponential rate of convergence under the $\gamma$-CKL oracle, thanks to a backtracking strategy that deals with the unavoidable errors in updating the belief region around the target.

10:   We evaluate the performance of the algorithm both over the posited oracle and over several real-world triplet datasets.

11:   We also report on a comprehensive user study, where human subjects navigate a database of face portraits.

12:   %Our results are supported by an evaluation on established comparison datasets, a comprehensive user-study, and synthetic simulation results.

13: \end{abstract}

14: