How do two jazz improvisers balance individual voices with a shared rhythm?

Identity and togetherness split across dimensions

The paper's central result is a double dissociation. If two improvisers had fused into a single system, they should have been as indistinguishable in pitch as they were in rhythm; if they had stayed independent, rhythm should have tracked the individual as reliably as pitch did. Neither happened. Each duo realized both organizations at once, carried by different musical dimensions.

The evidence comes from 986 jazz improvisations in total, but the interaction analyses rest on a narrower set: 6 duos (12 professional improvisers), 54 takes, 530 solos, and 11,433 note events, of which 10,903 had a defined predictive context. These were the solo-trading sections of pulsed performances of Autumn Leaves, where the two players strictly alternated solos over a shared form. The authors treat this strict alternation as a conversational turn-taking structure that makes interaction measurable note by note.

Surprise and uncertainty as note-level measurements

The study does not ask listeners how the music felt. It computes two quantities for every note. Surprise is indexed by information content, the negative log probability of the note actually played given its preceding context; a high value means the realized note was unlikely under the learned model of the jazz idiom. Uncertainty is indexed by entropy, the spread of the model's predictive distribution; a high value means the context allowed many plausible continuations. Both are expressed in bits.

These quantities came from a hierarchical Bayesian statistical-learning model trained on the Weimar Jazz Database of the Jazzomat project, 456 professionally transcribed monophonic solos by 78 musicians. That corpus was used only to estimate population-level regularities of the jazz idiom and was never fitted to the performances used for hypothesis testing. The model then continued to update its statistics online while listening to the duo recordings, as a listener accumulating experience would. Because the two partners within a take were evaluated at essentially the same model state, comparisons between them are not affected by this gradual updating.

The model was run separately on three event streams: pitch, rhythm, and a joint stream encoding each tone as a compound pitch-rhythm symbol. This separation matters, because the paper's main contrast is precisely between what the pitch stream and the rhythm stream carry.

Pitch fluctuations identify the performer

Pitch-based fluctuations behaved as stable signatures of individual identity. Within a single take, the two partners' note-level uncertainty values differed significantly in 90% of takes, whereas the same performer compared across consecutive takes differed in only 35%. This contrast held in all six duos.

The separation was also large and persistent. Within a take, distances between solos by two different performers were 2.14 times greater than distances between solos by the same performer, using mean Jensen-Shannon divergence of 0.261 versus 0.125. The distance to the partner's solos was as large one turn later as five turns later (0.270 versus 0.254), so turn-taking never closed the gap. Performer identity could be decoded from single solos with a leave-one-take-out cross-validation AUC of 0.87, above chance in 6 of 6 duos. A mixed-model variance decomposition assigned 74.3% of the variance in uncertainty to performer identity and only 0.6% to take; for surprise the split was 18.6% versus 1.3%.

One caveat belongs here rather than at the end: the two members of every duo played different instruments, so part of this melodic signature may reflect instrument-specific affordances for pitch rather than personal style alone. The authors note that this concern does not touch the rhythmic findings, whose defining feature is the absence of individual specificity.

Rhythm fluctuations belong to the performance

Rhythm-based fluctuations showed the reverse organization. Within a take, the two partners were largely indistinguishable: only 39% of take-level comparisons reached significance, despite the same note-level statistical power that detected pitch differences in 90% of takes. Instead, rhythmic fluctuations changed from performance to performance. The same performer differed from themselves across consecutive takes in 60% of comparisons, significantly more often than the two partners differed within a take.

The decoding and variance results point the same way. Performer identity could not be decoded from rhythmic fluctuations (AUC = 0.52), and PERMANOVA attributed significant variance to take in every duo but to performer in none, with take exceeding performer in 6 of 6 duos (mean R-squared 0.56 versus 0.01).

The decisive test was whether the shared state belonged to the actual interaction or merely to the pair as a fixed combination of two people. The authors compared each solo's distance to the partner's next solo in the same take with its distance to the same partner instrument's solo at the same turn position in a different take, a pseudo-duet of the same two people who were not, at that moment, playing together. Rhythmic distances were twice as large in the pseudo-duets (0.116 versus 0.238), and the difference persisted when pseudo-duets were drawn only from the immediately adjacent take (0.116 versus 0.161), which rules out slow session-wide change as the sole explanation. Pitch showed no such take-specificity (0.275 versus 0.280).

The shared rhythmic state drifts collectively

If each take establishes a shared rhythmic state, the next question is whether that state stays put. The strict alternation makes this testable: odd separations of turns always compare a performer with their partner, while even separations always compare a performer with themselves. The two series can therefore be read directly as distance to the other and distance to one's own past.

In rhythm, both series climbed together. The distance to the partner rose from 0.111 at one turn apart to 0.126 at three turns and 0.146 at five turns, while the distance to one's own previous solos rose in parallel from 0.116 to 0.134 to 0.149. Across all available separations up to 22 turns, the lag-distance correlation was positive in every duo (mean Spearman rho = +0.17). Because each performer moved away from their own past at the same rate as from their partner, the pair was not diverging; the shared rhythmic state itself drifted gradually across the take.

Pitch showed no such drift (mean Spearman rho = +0.01). The melodic separation between partners was as constant in time as it was wide. The authors are explicit that gradual joint drift alone does not demonstrate interaction, since two performers independently wandering at similar rates could produce a similar profile. That is why the next test targets turn-level dependence.

Solo hand-offs show complementarity, not imitation

At each solo hand-off, the partners showed a systematic push-pull. For each solo the authors computed its mean surprise, removed each take's linear trend, and expressed the residual relative to that performer's own average, then asked whether one performer's deviation on a turn predicted the partner's deviation on the next. Because strict alternation combined with stable level differences can produce spurious negative correlations, they benchmarked the observed coupling against two null models that preserve each individual's playing but break its alignment with the partner: shuffling each performer's solos within their own turn positions, and circularly shifting one performer's sequence against the other's.

Real duos showed reliable negative coupling that both controls abolished. In rhythm, the mean within-duo correlation across consecutive solos was r = -0.230, versus -0.000 under the first control and +0.069 under the second. In pitch, r = -0.151 versus +0.001 and +0.045. In the joint stream the effect was weaker and did not reach significance (r = -0.110). Pooling all hand-offs in the corpus (n = 415), a performer who played above their own average was reliably followed by a partner playing below their own (rhythm r = -0.19; pitch r = -0.10).

The authors call this complementarity rather than imitation, and note that it runs opposite to the convergence emphasized in accounts of interpersonal synchrony and alignment. Their interpretation is that the pair co-regulates the joint flow of unpredictability: when one spends, the other saves, keeping the total stream of surprise within a workable band while the shared state slowly explores. That reading is the authors' explanation, not a directly measured mechanism.

A separate unsigned analysis asked whether the partner's temporally adjacent solo was also the most similar. This held for rhythm only: solos were closer to the partner's next solo than to the partner's solos three or more turns away, but not for pitch. The two analyses are not redundant, because the signed push-pull captures the direction of momentary deviations while the unsigned analysis captures overall proximity.

What the design can and cannot establish

The analyses are observational. They establish that the observed coupling exceeds stringent null models, but not the causal pathway that produces it. All performances were improvisations on a single tune within a single idiom, so whether the same division of roles between identity-bearing and state-bearing dimensions holds across repertoires, genres, and cultures remains untested, as does whether assigning pitch to identity and rhythm to the shared state is universal or idiom-specific.

The reference model was trained on a specific historical corpus of monophonic solos, and different reference corpora could shift absolute values of surprise and uncertainty, although the within-corpus contrasts on which the conclusions rest are less exposed to that shift. The authors also state that the data show how novelty is jointly regulated but not that this regulation is what makes the interaction creative in the evaluative sense; linking these dynamics to judgments of creativity and aesthetic value is left as a next step.

The paper's own framing of the result refines its starting hypothesis. The authors had proposed that creative duos maintain an interpersonal distance in predictive fluctuations that is neither too close nor too far. The results show that this intermediate organization is real but is not implemented as a compromise on a single dimension. It is implemented as a division of roles across dimensions: complete sharing in rhythm, preserved difference in pitch, and active rebalancing at the point of exchange in both.

The dimension that is shared is the one that moves

The pattern the authors highlight is that what is shared moves, and what is not shared is preserved. The dimension in which the partners fully converged, rhythm, was precisely the dimension that travelled across the take and reset between takes, whereas the dimension that carried stable identity, pitch, stayed where it was.

Their suggested functional reading is that a group moving together in some shared dimension can explore new territory without fragmenting, while stable individual differences maintain the diversity of material available for exploration. This is an interpretation offered in the discussion, not a tested outcome of the study.

For readers weighing what this paper changes, the most useful distinction is between the two kinds of question the measures answer. Pitch fluctuations answer who is playing; rhythm fluctuations answer which performance is unfolding. A duo can therefore be maximally close and reliably distinct at the same time, without either dimension settling at an intermediate value. The paper's own next step is to test whether that organization relates to judged creativity, which the present data do not establish.

Sources used in this answer