Fairness, Memory and Regeneration in Time

A random record does not reveal its whole future at once. It opens one more page: the result of a wager, a new observation, a return, or the birth of a child. We revise judgments accordingly and sometimes decide whether to continue. Probability is no longer concerned only with how one distribution allocates chances to outcomes, but with how judgments change as the observed world unfolds.

For finitely many experiments, successive conditional averaging seems to settle matters. If the next step offers no advantage given the information already available, a game remains fair. An infinite record exposes a difficulty in that simple judgment: an average relation valid at every finite time need not survive in the path's eventual outcome. No particular step has secretly changed the rules. The issue is whether a finite-time promise has been carried to infinity without justification.

A player who chooses when to leave changes the observation experiment again. A fixed time and waiting for the first win are different choices; the latter uses information acquired along the way. To determine which probability relations survive, we must specify what the player knows then, together with the waiting and losses that may occur before stopping. Time is not merely passing: it becomes a decision based on observations.

The same difficulty hides in the word present. For a coin with unknown bias, its latest outcome need not summarize all earlier observations; two records with the same ending can demand different next-step predictions. We seek not a label that makes the past disappear, but a state retaining the information through which the past still predicts the future. Whether the present suffices is something a model and its conditional probabilities must establish.

Once an adequate state has been found, returning acquires another meaning: before the path’s lifetime ends, reaching the same position at a legitimate random time may restore the same future rule. But it does not guarantee that the time will ever come. A walker whose individual steps favor no direction may still leave its starting point forever. Local fairness, identical future rules and eventual path behavior must be distinguished.

When a return does occur, long stretches of time can instead be read as cycles between returns. One cycle ends and another wait begins, allowing simpler repeated experiments to describe a long record. Yet choosing a cycle by its arrival number and choosing a location on the clock do not reveal the same typical object. Long cycles occupy more time, making the observation procedure part of the probability problem.

Restarting need not continue along a single line. Each individual can produce several children, which independently use the same reproduction rule, giving time a genealogical structure. Averaging over possible families and observing growth along a surviving tree can yield very different pictures. Rare large families may carry a substantial share of the mean; changing observational weights can make those contributions visible.

What connects these problems is not a common answer. Conditional fairness cannot replace integrability, the Markov property does not ensure return, and independent renewal cycles do not automatically make arbitrary time windows independent. Together they demand attention to a distinction: a rule describes how to take another step, whereas judgments about a whole path also require us to understand retained information, chosen times and the exact conditions under which an experiment restarts.

Generations and step numbers provide orderly clocks. Arrivals along real time leave gaps of varying length. Does the length of a gap already observed change how much longer we must wait? When do restarting at an arrival and restarting at an arbitrary time give the same future? A count distribution within one fixed window is insufficient: arrival locations, waiting intervals and relationships between windows must be constructed together.

If an arrival changes the system's state as well, the next waiting rule may change with it. Writing such rules is only a beginning. We must check that they define an actual time record rather than squeeze infinitely many changes into finite time. If they do, continuation needs a further rule. These are difficulties introduced by a continuous clock, and intuition from a discrete clock cannot remove them in advance.

Fair games, spatial walks, repeated waits and random families have different histories and require different proofs. Bringing them together in this volume invents no common history of discovery. It asks how we can make reliable judgments about a future revealed through finite information, and how we recognize where those judgments cease to apply.

We begin with a modest agreement: given what is already known, the mean gain of the next step is zero. We shall see how far that finite promise of fairness can constrain an infinite path.