The concept of "Balance" has an important role in treatment effects studies. However, it is often misunderstood and misinterpreted. In this article, we investigate the concept of "Balance", its essence and function. We find that instead of one, there are three different types of balance: the conditional (or unconditional) independence between treatment variable and covariates (Type I balance), the similarity of empirical distributions between treated group and control group (Type II balance), the balancing property of propensity score (Type III balance). The conditional (unconditional) independence is powered by the unconfoundedness assumption, while the similarity of empirical distributions is backed by the overlap assumption. The propensity score has balancing property as an innate quality, not in relation to unconfoundedness assumption. Analyzing several related papers, we find several incorrect views caused by mixing different types of balance, such as "In a randomized experiment, covariates must be balanced", "The object of observational study is to mimic a completely randomized experiment", "Covariates imbalance signifies failure of unconfoundedness assumption", "Normalized difference can measure the balancing property of propensity score", "Normalized difference and Imai-Ratkovic overidentification test serve same purpose", "Two sample t-test is suitable for checking covariates balance". We conduct four Monte Carlo simulations, and their results confirm our viewpoints.