The second law states that entropy obeys a local paraconservation law. That is, entropy is “nearly” conserved.
By that we mean something very specific:
 (2.1) 
The structure and meaning of equation 2.1 is very similar to equation 1.2, except that it has an inequality instead of an equality. It tells us that the entropy in a given region can increase, but it cannot decrease except by flowing into adjacent regions.
As usual, the local law implies a corresponding global law, but not conversely; see the discussion at the end of section 1.2.
Entropy is absolutely essential to thermodynamics … just as essential as energy.
Entropy is defined in terms of statistics, as we will discuss in a moment. In some situations, there are important connections between entropy, energy, and temperature … but these do not define entropy. The first law (energy) and the second law (entropy) are logically independent.
If the second law is to mean anything at all, entropy must be welldefined always. Otherwise we could create loopholes in the second law by passing through states where entropy was not defined.
We do not define entropy via dS = dQ/T or anything like that, first of all because (as discussed in section 7.2) there is no statefunction Q such that dQ = TdS, and more importantly because we need entropy to be well defined even when the temperature is unknown, undefinable,^{1} irrelevant, or zero.
Entropy is related to information. Essentially it is the opposite of information, as we see from the following scenarios.
As shown in figure 2.1, suppose we have three blocks and five cups on a table.
To illustrate the idea of entropy, let’s play the following game: Phase 0 is the preliminary phase of the game. During phase 0, the dealer hides the blocks under the cups however he likes (randomly or otherwise) and optionally makes an announcement about what he has done. As suggested in the figure, the cups are transparent, so the dealer knows the exact microstate at all times. However, the whole array is behind a screen, so the rest of us don’t know anything except what we’re told.
Phase 1 is the main phase of the game. During phase 1, we are required to ascertain the position of each of the blocks. Since in this version of the game, there are five cups and three blocks, the answer can be written as a threesymbol string, such as 122, where the first symbol identifies the cup containing the red block, the second symbol identifies the cup containing the black block, and the third symbol identifies the cup containing the blue block. Each symbol is in the range zero through four inclusive, so we can think of such strings as base5 numerals, three digits long. There are 5^{3} = 125 such numerals. (More generally, in a version where there are N cups and B blocks, there are N^{B} possible microstates.)
We cannot see what’s inside the cups, but we are allowed to ask yes/no questions, whereupon the dealer will answer. Our score in the game is determined by the number of questions we ask; each question contributes one bit to our score. Our objective is to finish the game with the lowest possible score.
Remark on terminology: Any microstates that have zero probability are classified as inaccessible, while those that have nonzero probability are classified as accessible.
These examples have certain restrictions in common:  More generally: 
For starters, we have been asking yes/no questions.  Binary questions are not universally required; by way of contrast you can consider the threeway measurements in reference 13. 
Also, so far we have only considered scenarios where all accessible microstates are equally probable.  If the accessible microstates are not equally probable, we need a more sophisticated notion of entropy, as discussed in section 2.6. 
Subject to these restrictions, if we want to be sure of identifying the correct microstate, we should plan on asking a sufficient number of questions S such that 2^{S} is greater than or equal to the number of accessible microstates. 
To calculate what our score will be, we don’t need to know anything about energy; all we have to do is count states (specifically, the number of microstates consistent with what we know about the situation). States are states; they are not energy states.
If you wish to make this sound more thermodynamical, you can assume that the table is horizontal, and the blocks are noninteracting, so that all possible configurations have the same energy. But really, it is easier to just say that over a wide range of energies, energy has got nothing to do with this game.
The point of all this is that we define the entropy of a given situation according to the number of questions we have to ask to finish the game, starting from the given situation. Each yes/no question contributes one bit to the entropy.
The central, crucial idea of entropy is that it measures how much we don’t know about the situation. Entropy is not knowing.
Here is a card game that illustrates the same points as the cup game. The only important difference is the size of the state space: roughly eighty million million million million million million million million million million million states, rather than 125 states. That is, when we move from 5 cups to 52 cards, the state space gets bigger by a factor of 10^{66} or so.
Consider a deck of 52 playing cards. By reordering the deck, it is possible to create a large number (52 factorial) of different configurations.
Technical note: There is a separation of variables. We choose to consider only the part of the system that describes the ordering of the cards. We assume these variables are statistically independent of other variables, such as the spatial position and orientation of the cards. This allows us to understand the entropy of this subsystem separately from the rest of the system, for reasons discussed in section 2.8.Also, unless otherwise stated, we assume the number of cards is fixed at 52 ... although the same principles apply to smaller or larger decks, and sometimes in an introductory situation it is easier to see what is going on if you work with only 8 or 10 cards.
Phase 0 is the preliminary phase of the game. During phase 0, the dealer prepares the deck in a configuration of his choosing, using any combination of deterministic and/or random procedures. He then sets the deck on the table. Finally he makes zero or more announcements about the configuration of the deck.
Phase 1 is the main phase of the game. During phase 1, our task is to fully describe the configuration, i.e. to determine which card is on top, which card is second, et cetera. We cannot look at the cards, but we can ask yes/no questions of the dealer. Each such question contributes one bit to our score. Our objective is to ask as few questions as possible. As we shall see, our score is a measure of the entropy.
One configuration of the card deck corresponds to one microstate. The microstate does not change during phase 1.
The macrostate is the ensemble of microstates consistent with what we know about the situation.
At this point we know that the deck is in some microstate, and the microstate is not changing … but we don’t know which microstate. It would be foolish to pretend we know something we don’t. If we’re going to bet on what happens next, we should calculate our odds based on the ensemble of possibilities, i.e. based on the macrostate.
Our best strategy is as follows: By asking six wellchosen questions, we can find out which card is on top. We can then easily describe every detail of the configuration. Our score is six bits.
This illustrates that the entropy is a property of the ensemble, i.e. a property of the macrostate, not a property of the microstate. Cutting the deck the second time changed the microstate but did not change the macrostate. See section 2.4 and especially section 2.7 for more discussion of this point.
Note that we are not depending on any special properties of the “reference” state. For simplicity, we could agree that our reference state is the factorystandard state (cards ordered according to suit and number), but any other agreedupon state would work just as well. If we know deck is in Moe’s favorite state, we can easily rearrange it into Joe’s favorite state. Rearranging it from one known state to to another known state does not involve any entropy.
As a variation on the game described in section 2.3, consider what happens if, at the beginning of phase 1, we are allowed to peek at one of the cards.
In the case of the standard deck, example 1, this doesn’t tell us anything we didn’t already know, so the entropy remains unchanged.
In the case of the cut deck, example 3, this lowers our score by six bits, from six to zero.
In the case of the shuffled deck, example 6, this lowers our score by six bits, from 226 to 220.
The reason this is worth mentioning is because peeking can (and usually does) change the macrostate, but it cannot change the microstate. This stands in contrast to cutting an alreadycut deck or shuffling an alreadyshuffled deck, which changes the microstate but does not change the macrostate. This is a multiway contrast, which we can summarize as follows:
Macrostate Changes  Macrostate Doesn’t Change  
Microstate Usually Changes: 

 
Microstate Doesn’t Change: 


This gives us a clearer understanding of what the macrostate is. Essentially the macrostate is the ensemble, in the sense that specifying the ensemble specifies the macrostate and vice versa. Equivalently, we can say that the macrostate is a probability distribution over microstates.
In the simple case where all the microstates are equiprobable, the ensemble is simply the set of all microstates that are consistent with what you know about the system.
In a poker game, there is only one deck of cards. Suppose player Alice has peeked but player Bob has not. Alice and Bob will then play according to very different strategies. They will use different ensembles – different macrostates – when calculating their next move. The deck is the same for both, but the macrostate is not.
We see that the physical state of the deck does not provide a complete description of the macrostate. The players’ knowledge of the situation is also relevant, since it affects how they calculate the probabilities. Remember, as discussed in item 4 and in section 2.7, entropy is a property of the macrostate, not a property of the microstate, so peeking can – and usually does – change the entropy.
To repeat: Peeking does not change the microstate, but it can have a large effect on the macrostate. If you don’t think peeking changes the ensemble, I look forward to playing poker with you!
If you want to understand entropy, you must first have at least a modest understanding of basic probability. It’s a prerequisite, and there’s no way of getting around it. Anyone who knows about probability can learn about entropy. Anyone who doesn’t, can’t.
Our notion of entropy is completely dependent on having a notion of microstate, and on having a procedure for assigning a probability to each microstate.
In some special cases, the procedure involves little more than counting the “allowed” microstates, as discussed in section 8.6. This type of counting corresponds to a particularly simple, flat probability distribution, which may be a satisfactory approximation in special cases, but is definitely not adequate for the general case.
For simplicity, the cup game and the card game were arranged to embody a clear notion of microstate. That is, the rules of the game specified what situations would be considered the “same” microstate and what would be considered “different” microstates. Such games are a model that is directly and precisely applicable to physical systems where the physics is naturally discrete, such as systems involving only the nonclassical spin of elementary particles (such as the demagnetization refrigerator discussed in section 10.10).
For systems involving continuous variables such as position and momentum, counting the states is somewhat trickier. The correct procedure is discussed in section 11.3.
The point of all this is that the “score” in these games is an example of entropy. Specifically: at each point in the game, there are two numbers worth keeping track of: the number of questions we have already asked, and the number of questions we must ask to finish the game. The latter is what we call the the entropy of the situation at that point.
Remember that the macrostate is the ensemble of microstates. In the ensemble, probabilities are assigned taking into account what the observer about the situation. The entropy is a property of the macrostate.
At each point during the game, the entropy is a property of the macrostate, not of the microstate. The system must be in “some” microstate, but we don’t know which microstate, so all our decisions must be based on the macrostate.
The value any given observer assigns to the entropy depends on what that observer knows about the situation, not what the dealer knows, or what anybody else knows. This makes the entropy somewhat contextdependent or even subjective. Some people find this irksome or even shocking, but it is real physics. For physical examples of contextdependent entropy, and a discussion, see section 11.8.
Note that entropy has been defined without reference to temperature and without reference to heat. Room temperature is equivalent to zero temperature for purposes of the cup game and the card game; theoretically there is “some” chance that thermal agitation will cause two of the cards to spontaneously hop up and exchange places during the game, but that is really, really negligible.
Nonexperts often try to define entropy in terms of energy. This is a mistake. To calculate the entropy, I don’t need to know anything about energy; all I need to know is the probability of each relevant state. See section 2.6 for details on this.
Entropy is not defined in terms of energy, nor vice versa.
In some cases, there is a simple mapping that allows us to identify the ith microstate by means of its energy Ê_{i}. It is often convenient to exploit this mapping when it exists, but it does not always exist.
In pop culture, entropy is often associated with disorder. There are even some textbooks that try to explain entropy in terms of disorder. This is not a good idea. It is all the more disruptive because it is in some sense half true, which means it might pass superficial scrutiny. However, science is not based on halftruths.
Small disorder generally implies small entropy. However, the converse does not hold, not even approximately; A highlydisordered system might or might not have high entropy. The spin echo experiment (section 10.7) suffices as an example of a highly disordered macrostate with relatively low entropy.
Before we go any farther, we should emphasize that entropy is a property of the macrostate, not of the microstate. In contrast, to the extent that “disorder” can be measured at all, it can be measured on a microstatebymicrostate basis. Therefore, whatever the “disorder” is measuring, it isn’t entropy. (A similar microstate versus macrostate argument applies to the “energy dispersal” model of entropy, as discussed in section 8.8.) As a consequence, the usual textbook illustration – contrasting snapshots of orderly and disorderly scenes – cannot be directly interpreted in terms of entropy. To get any value out of such an illustration, the reader must make a sophisticated leap:
The disorderly snapshot must be interpreted as representative of an ensemble with a very great number of similarlydisorderly microstates. The ensemble of disorderly microstates has high entropy. This is a property of the ensemble, not of the depicted microstate or any other microstate.  The orderly snapshot must be interpreted as representative of a very small ensemble, namely the ensemble of similarlyorderly microstates. This small ensemble has a small entropy. Again, entropy is a property of the ensemble, not of any particular microstate (except in the extreme case where there is only one microstate in the ensemble, and therefore zero entropy). 
To repeat: Entropy is defined as a weighted average over all microstates. Asking about the entropy of a particular microstate (disordered or otherwise) is asking the wrong question. As a matter of principle, the question is unanswerable.
Note the following contrast:
Entropy is a property of the macrostate. It is defined as an ensemble average.  Disorder, to the extent it can be defined at all, is a property of the microstate. (You might be better off focusing on the surprisal rather than the disorder, as discussed in section 2.7.) 
The number of orderly microstates is very small compared to the number of disorderly microstates. That’s because when you say the system is “ordered” you are placing constraints on it. Therefore if you know that the system is in one of those orderly microstates, you know the entropy cannot be very large.
The converse does not hold. If you know that the system is in some disorderly microstate, you do not know that the entropy is large. Indeed, if you know that the system is in some particular disorderly microstate, the entropy is zero. (This is a corollary of the more general proposition that if you know what microstate the system is in, the entropy is zero. it doesn’t matter whether that state “looks” disorderly or not.)
Furthermore, there are additional reasons why the typical textbook illustration of a messy dorm room is not a good model of entropy. For starters, it provides no easy way to define and delimit the states. Even if we stipulate that the tidy state is unique, we still don’t know whether a shirt on the floor “here” is different from a shirt on the floor “there”. Since we don’t know how many different disorderly states there are, we can’t quantify the entropy. (In contrast the games in section 2.2 and section 2.3 included a clear rule for defining and delimiting the states.)
Examples of low entropy and relatively high disorder include, in order of increasing complexity:
Technical note: There is a separation of variables, analogous to the separation described in section 2.3. We consider only the part of the system that describes whether the coins are in the “heads” or “tails” state. We assume this subsystem is statistically independent of the other variables such as the position of the coins, rotation in the plane, et cetera. This means we can understand the entropy of this subsystem separately from the rest of the system, for reasons discussed in section 2.8.
Randomize the coins by shaking. The entropy at this point is five bits. If you open the box and peek at the coins, the entropy goes to zero. This makes it clear that entropy is a property of the ensemble, not a property of the microstate. Peeking does not change the disorder. Peeking does not change the microstate. However, it can (and usually does) change the entropy. This example has the pedagogical advantage that it is small enough that the entire microstatespace can be explicitly displayed; there are only 32 = 2^{5} microstates.
Ordinarily, a wellshuffled deck of cards contains 225.581 bits of entropy, as discussed in section 2.3. On the other hand, if you have peeked at all the cards after they were shuffled, the entropy is now zero, as discussed in section 2.4. Again, this makes it clear that entropy is a property of the ensemble, not a property of the microstate. Peeking does not change the disorder. Peeking does not change the microstate. However, it can (and usually does) change the entropy.
Many tricks of the cardsharp and the “magic show” illusionist depend on a deck of cards arranged to have much disorder but little entropy.
There is a longrunning holy war between those who try to define entropy in terms of energy, and those who try to define it in terms of disorder. This is based on a grotesquely false dichotomy: If entropyasenergy is imperfect, then entropyasdisorder “must” be perfect … or vice versa. I don’t know whether to laugh or cry when I see this. Actually, both versions are highly imperfect. You might get away with using one or the other in selected situations, but not in general.
The right way to define entropy is in terms of probability, as we now discuss. (The various other notions can then be understood as special cases and/or approximations to the true entropy.)
The idea of entropy set forth in the preceding examples can be quantified quite precisely. Entropy is defined in terms of statistics.^{2} For any classical probability distribution P, we can define its entropy as:
S[P] := 
 P_{i} log(1/P_{i}) (2.2) 
where the sum runs over all possible outcomes and P_{i} is the probability of the ith outcome. Here we write S[P] to make it explicit that S is a functional that depends on P. For example, if P is a conditional probability then S will be a conditional entropy. Beware that people commonly write simply S, leaving unstated the crucial dependence on P.
Equation 2.2 is the faithful workhorse formula for calculating the entropy. It ranks slightly below Equation 26.6, which is a more general way of expressing the same idea. It ranks above various lessgeneral formulas that may be useful under morerestrictive conditions (as in section 8.6 for example). See chapter 21 and chapter 26 for more discussion of the relevance and range of validity of this expression.
Subject to mild restrictions, equation 2.2 applies to physics as follows: Suppose the system is in a given macrostate, and the macrostate is well described by a distribution P, where P_{i} is the probability that the system is in the ith microstate. Then we can say S is the entropy “of the system”.
Expressions of this form date back to Boltzmann (reference 14 and reference 15) and to Gibbs (reference 16). The range of applicability was greatly expanded by Shannon (reference 17). See also equation 26.6.
Beware that uncritical reliance on “the” observed microstatebymicrostate probabilities does not always give a full description of the macrostate, because the P_{i} might be correlated with something else (as in section 10.7) or amongst themselves (as in chapter 26). In such cases the unconditional entropy S[P] will be larger than the conditional entropy S[PQ], and you have to decide which is/are physically relevant.
In the games discussed above, it was convenient to measure entropy in bits, because I was asking yes/no questions. Other units are possible, as discussed in section 8.5.
Figure 2.2 shows the contribution to the entropy from one term in the sum in equation 2.2. Its maximum value is approximately 0.53 bits, attained when P(H)=1/e.
Let us now restrict attention to a system that only has two microstates, such as a coin toss, so there will be exactly two terms in the sum. That means we can identify P(H) as the probability of the the “heads” state. The other state, the “tails” state, necessarily has probability P(T)≡1−P(H) and that gives us the other term in the sum, as shown by the red curve in figure 2.3. The total entropy is shown by the black curve in figure 2.3. For a twostate system, it is necessarily a symmetric function of P(H). Its maximum value is 1 bit, attained when P(H)=P(T)=½.
As discussed in section 8.5 the base of the logarithm in equation 2.2 is chosen according to what units you wish to use for measuring entropy. If you choose units of joules per kelvin (J/K), we can pull out a factor of Boltzmann’s constant and rewrite the equation as:
S = −k 
 P_{i} lnP_{i} (2.3) 
Entropy itself is conventionally represented by big S. Meanwhile, molar entropy is conventionally represented by small s and is the corresponding intensive property.
People commonly think of entropy as being an extensive quantity. This is true to a good approximation in many situations, but not all. Some exceptions are discussed in section 11.8 and especially section 11.11.
Although it is often convenient to measure molar entropy in units of J/K/mol, other units are allowed, for the same reason that mileage is called mileage even when it is measured in metric units. In particular, sometimes additional insight is gained by measuring molar entropy in units of bits per particle. See section 8.5 for more discussion of units.
When discussing a chemical reaction using a formula such as
2 O_{3} → 3 O_{2} + Δs (2.4) 
it is common to speak of “the entropy of the reaction” but properly it is “the molar entropy of the reaction” and should be written Δs or ΔS/N (not ΔS). All the other terms in the formula are intensive, so the entropyrelated term must be intensive also.
Of particular interest is the standard molar entropy, s^{0} or S^{0}/N, measured at standard temperature and pressure. The entropy of a gas is strongly dependent on density, as mentioned in section 11.3.
If we have a system characterized by a probability distribution P, the surprisal of the ith state is given by
$_{i} := log(1/P_{i}) (2.5) 
By comparing this with equation 2.2, it is easy to see that the entropy is simply the ensemble average of the surprisal. In particular, it is the expectation value of the surprisal. (See equation 26.7 for the fully quantummechanical generalization of this idea.)
Note the following contrast:
Surprise value is a property of the microstate i.  Entropy is a property of the distribution P as a whole. It is defined as an ensemble average. 
Some consequences of this fact were discussed in item 4 and in section 2.4.
Another obvious consequence is that entropy is not, by itself, the solution to all the world’s problems. Entropy measures a particular average property of the distribution. It is easy to find situations where other properties of the distribution are worth knowing.
Suppose we have subsystem 1 with a set of microstates {(i)} and subsystem 2 with a set of microstates {(j)}. Then in all generality, the microstates of the combined system are given by the Cartesian direct product of these two sets, namely
{(i)}×{(j)} = {(i,j)} (2.6) 
where (i,j) is an ordered pair, which should be a familiar idea and a familiar notation.
We now consider the lessthangeneral case where the two subsystems are statistically independent. That means that the probabilities are multiplicative:
R(i,j) = P(i) Q(j) (2.7) 
Let’s evaluate the entropy of the combined system:
 (2.8) 
where we have used the fact that the subsystem probabilities are normalized.
So we see that the entropy is additive whenever the probabilities are multiplicative, i.e. whenever the probabilities are independent.