Most mathematical models of athletic training require the quantification of training intensity and quantity or 'dose'. We aim to summarize both the methods available for such quantification, particularly in relation to cycle sport, and the mathematical techniques that may be used to model the relationship between training and performance. Endurance athletes have used training volume (kilometres per week and/or hours per week) as an index of training dose with some success. However, such methods usually fail to accommodate the potentially important influence of training intensity. The scientific literature has provided some support for alternative methods such as the session rating of perceived exertion, which provides a subjective quantification of the intensity of exercise; and the heart rate-derived training impulse (TRIMP) method, which quantifies the training stimulus as a composite of external loading and physiological response, multiplying the training load (stress) by the training intensity (strain). Other methods described in the scientific literature include 'ordinal categorization' and a heart rate-based excess post-exercise oxygen consumption method. In cycle sport, mobile cycle ergometers (e.g. SRM and PowerTap) are now widely available. These devices allow the continuous measurement of the cyclists' work rate (power output) when riding their own bicycles during training and competition. However, the inherent variability in power output when cycling poses several challenges in attempting to evaluate the exact nature of a session. Such variability means that average power output is incommensurate with the cyclist's physiological strain. A useful alternative may be the use of an exponentially weighted averaging process to represent the data as a 'normalized power'. Several research groups have applied systems theory to analyse the responses to physical training. Impulse-response models aim to relate training loads to performance, taking into account the dynamic and temporal characteristics of training and, therefore, the effects of load sequences over time. Despite the successes of this approach it has some significant limitations, e.g. an excessive number of performance tests to determine model parameters. Non-linear artificial neural networks may provide a more accurate description of the complex non-linear biological adaptation process. However, such models may also be constrained by the large number of datasets required to 'train' the model. A number of alternative mathematical approaches such as the Performance-Potential-Metamodel (PerPot), mixed linear modelling, cluster analysis and chaos theory display conceptual richness. However, much further research is required before such approaches can be considered as viable alternatives to traditional impulse-response models. Some of these methods may not provide useful information about the relationship between training and performance. However, they may help describe the complex physiological training response phenomenon.