Entropy and Dimensions (Following Landau and Lifshitz)

Some time ago I wrote about volumes of spheres in multi-dimensional phase space – as needed in integrals in statistical mechanics.

The post was primarily about the curious fact that the ‘bulk of the volume’ of such spheres is contained in a thin shell beneath their hyperspherical surfaces. The trick to calculate something reasonable is to spot expressions you can Tayler-expand in the exponent.

Large numbers ‘do not get much bigger’ if multiplied by a factor, to be demonstrated again by Taylor-expanding such a large number in the exponent; I used this example:

Assuming N is about 1025  then its natural logarithm is about 58 and $Ne^N = e^{\ln(N)+N} = e^{58+10^{25}}$, then 58 can be neglected compared to N itself.

However, in the real world numbers associated with locations and momenta of particles come with units. Calling the unit ‘length’ in phase space $R_0$ the large volume can be written as $aN{(\frac{r}{R_0})}^N = ae^{\ln{(N)} + N\ln{(\frac{r}{R_0})}}$, and the impact of an additional factor N also depends on the unit length chosen.

I did not yet mention the related issues with the definition of entropy. In this post I will follow the way Landau and Lifshitz introduce entropy in Statistical Physics, Volume 5 of their Course of Theoretical Physics.

Landau and Lifshitz introduce statistical mechanics top-down, starting from fundamental principles and from Hamiltonian classical mechanics: no applications, no definitions of ‘heat’ and ‘work’, nor historical references needed for motivation. Classical phenomenological thermodynamics is only introduced after their are done with the statistical foundations. Both entropy and temperature are defined – these are useful fundamental properties spotted in the mathematical derivations and thus deserve special names. They cover both classical and quantum statistics in small number of pages – LL’s style has been called terse or elegant.

The behaviour of a system with a large number of particles is encoded in a probability distribution function in phase space, a density. In the classical case this is a continuous function of phase-space co-ordinates. In the quantum case you consider distinct states – whose energy levels are densely packed together though. Moving from classical to quantum statistics means to count those states rather than to integrate the smooth density function over a volume. There are equivalent states created by permutations of identical particles – but factoring in that is postponed and not required for a first definition of entropy. A quasi-classical description is sufficient: using a minimum cell in phase space, whose dimensions are defined by Planck’s constant h that has a dimension of action – length times momentum.

Entropy as statistical weight

Entropy S is defined as the logarithm of the statistical weight $\Delta \Gamma$ – the number of quantum states associated with the part of phase phase used by the (sub)-system. (Landau and Lifshitz use the concept of a – still large – subsystem embedded in a larger volume most consequentially, in order to avoid reliance on the ergodic hypothesis as mentioned in the preface). In the quasi-classical view the statistical weight is the volume in phase space occupied by the system divided by the size of the minimum unit cell defined by Planck’s constant h. Denoting momenta by p, positions by q, using $\Delta p$ and $\Delta q$ as a shortcut applying multiple dimensions equivalent to s degrees of freedom…

$S = log \Delta \Gamma = log \frac {\Delta p \Delta q}{2 \pi \hbar^s}$

An example from solid state physics: if the system is considered a rectangular box in the physical world, possible quantum states related to vibrations can be visualized in terms of possible standing waves that ‘fit’ into the box. The statistical weight would then single out those bunch of states the system actually ‘has’ / ‘uses’ / ‘occupies’ in the long run.

Different sorts of statistical functions are introduced, and one reason for writing this article to emphasize the difference between them: The density function associates each point in phase space – each possible configuration of a system characterized by the momenta and locations of all particles – with a probability. These points are also called microstates. Taking into account the probabilities to find a system in any of these microstates gives you the so-called macrostate characterized by the statistical weight: How large or small a part of phase space the system will use when watched for a long time.

The canonical example is an ideal gas in a vessel: The most probable spacial distribution of particles is to find them spread out evenly, the most unlikely configuration is to have them concentrated in (nearly) the same location, like one corner of the box. The density function assigns probabilities to these configurations. As the even distribution is so much much more likely, the $\Delta q$ part of the statistical weight would cover all of the physical volume available. The statistical weight function has to obtain a maximum value in the most likely case, in equilibrium.

The significance of energies – and why there are logarithms everywhere.

Different sufficiently large subsystems of one big system are statistically independent – as their properties are defined by their bulk volume rather than their surfaces interfacing with other subsystems – and the larger the volume, the larger the ratio of volume and surface.  Thus the probability density function for the combined system – as a function of momenta and locations of all particles in the total phase phase – has to be equal to the product of the densities for each subsystem. Denoting the classical density with $\rho$ and adding a subscript for the set of momenta and positions referring to a subsystem:

$\rho(q,p) = \rho_1(q_1,p_1) \rho_2(q_2,p_2)$

(Since these are probability densities, the actual probability is always obtained by multiplying with the differential(s) $dqdp$).

This means that the logarithm of the composite density is equal to the sum of the logarithms of the individual densities. This the root cause of having logarithms show up everywhere in statistical mechanics.

A mechanical system of particles is characterized by only 7 ‘meaningful’ additive integrals: Energy, momentum and angular momentum – they add up when you combine systems, in contrast to all the other millions of integration constants that would appear when solving the equations of motions exactly. Momentum and angular momentum are not that interesting thermodynamically, as one can change to a frame moving and rotating with the system (LL also cover rotating systems). So energy remains as the integral of outstanding importance.

From counting states to energy intervals

What we want is to relate entropy to energy, so assertions about numbers of states covered need to be translated to statements about energy and energy ranges.

LL denote the probability to find a system in (micro-)state n with energy $E_n$ as $w_n$ – the quantum equivalent of density $\rho$. $w_n$ has to be a linear function of the energy of this micro-state $E_n$ as per the additivity just mentioned above, and thus LL omit the subscript n for w:

$w_n = w(E_n)$

(They omit any symbol ever if possible to keep their notation succinct ;-))

A thermodynamic system has an enormous number of (mechanical) degrees of freedom. Fluctuations are small as per the law of large numbers in statistics, and the probability to find a system with a certain energy can be approximated by a sharp delta-function-like peak at the system’s energy E. So in thermal equilibrium its energy has a very sharp peak. It occupies a very thin ‘layer’ of thickness $\Delta E$ in config space – around the hyperplane that characterizes its average energy E.

Statistical weight $\Delta \Gamma$ can be considered the width of the related function: Energy-wise broadening of the macroscopic state $\Delta E$ needs to be translated to a broadening related to the number of quantum states.

We change variables, so the connection between Γ and E is made via the derivative of Γ with respect to E. E is an integral, statistical property of the whole system, and the probability for the system to have energy E in equilibrium is $W(E)dE$. E is not discrete so this is again a  probability density. It is capital W now – in contrast to $w_n$ which says something about the ‘population’ of each quantum state with energy $E_n$.

A quasi-continuous number of states per energy Γ is related to E by the differential:

$d\Gamma = \frac{d\Gamma}{dE} dE$.

As E peaks so sharply and the energy levels are packed so densely it is reasonable to use the function (small) w but calculate it for an argument value E. Capital W(E) is a probability density as a function of total energy, small w(E) is a function of discrete energies denoting states – so it has to be multiplied by the number of states in the range in question:

$W(E)dE = w(E)d\Gamma$

Thus…

$W(E) = w(E)\frac{d\Gamma}{dE}$.

The delta-function-like functions (of energy or states) have to be normalized, and the widths ΔΓ and ΔE multiplied by the respective heights W and w taken at the average energy $E_\text{avg}$ have to be 1, respectively:

$W(E_\text{avg}) \Delta E = 1$
$w(E_\text{avg}) \Delta \Gamma = 1$

(… and the ‘average’ energy is what is simply called ‘the’ energy in classical thermodynamics).

So $\Delta \Gamma$ is inversely proportional to the probability of the most likely state (of average energy). This can also be concluded from the quasi-classical definition: If you imagine a box full of particles, the least possible state is equivalent to all particles occupying a single cell in phase space. The probability for that is (size of the unit cell) over (size of the box) times smaller than the probability to find the particles evenly distributed on the whole box … which is exactly the definition of $\Delta \Gamma$.

The statistical weight is finally:

$\Delta \Gamma = \frac{d\Gamma(E_\text{avg})}{dE} \Delta E$.

… the broadening in $\Gamma$, proportional to the broadening in $E$

The more familiar (?) definition of entropy

From that, you can recover another familiar definition of entropy, perhaps the more common one. Taking the logarithm…

$log S = log (\Delta \Gamma) = -log (w(E_\text{avg}))$.

As log w is linear in E, the averaging of E can be extended to the whole log function. Then the definition of ‘averaging over states n’ can be used: To multiply the value for each state n by probability $w_n$ and sum up:

$- \sum_{n} w_n log w_n$.

… which is the first statistical expression for entropy I had once learned.

LL do not introduce Boltzmann’s constant k here

It is effectively set to 1 – so entropy is defined without a reference to k. k is is only mentioned in passing later: In case one wishes to measure energy and temperature in different units. But there is no need to do so, if you defined entropy and temperature based on first principles.

Back to units

In a purely classical description based on the volume in phase space instead of the number of states there would be no cell of minimum size, and then instead of the statistical weight we had simply this volume: But then entropy would be calculated in a very awkward unit, the logarithm of action. Every change of the unit for measuring volumes in phase space would result in an additive constant – the deeper reason why entropy in a classical context is only defined up to such a constant.

So the natural unit called $R_0$ above should actually be Planck’s constant taken to the power defined by the number of particles.

Temperature

The first task to be solved in statistical mechanics is to find a general way of formulating a proper density function small $w_n$ as a function of energy $E_n$. You can either assume that the system has a clearly defined energy upfront – the system lives on a ‘energy-hyperplane in phase space’ – or you can consider it immersed in a larger system later identified with a ‘heat bath’ which causes the system to reach thermal equilibrium. These two concepts are called the micro-canonical and the canonical distribution (or Gibbs distribution) and the actual distribution functions don’t differ much because the energy peaks so sharply also in the canonical case. It’s that type of calculations where those hyperspheres are actually needed.

Temperature as a concept emerges from a closer look at these distributions, but LL introduce it upfront from simpler considerations: It is sufficient to know that 1) entropy only depends on energy, 2) both are additive functions of subsystems, and 3) entropy is a maximum in equilibrium. You divide one system in two subsystems. The total change in entropy has to be zero as this is a maximum (in equilibrium), and what energy $dE_1$ leaves one system has to be received as $dE_2$ by the other system. Taking a look at the total entropy S as a function of the energy of one subsystem:

$0 = \frac{dS}{dE_1} = \frac{dS_1}{dE_1} + \frac{dS_2}{dE_1} =$
$= \frac{dS_1}{dE_1} + \frac{dS_2}{dE_2} \frac{dE_2}{dE_1} =$
$= \frac{dS_1}{dE_1} + \frac{dS_2}{dE_2}$

So $\frac{dS_x}{dE_x}$ has to be the same for each subsystem x. Cutting one of the subsystems in two  you can use the same argument again. So there is one very interesting quantity that is the same for every subsystem – $\frac{dS}{dE}$. Let’s call it 1/T and let’s call T the temperature.

Spheres in a Space with Trillions of Dimensions

I don’t venture into speculative science writing – this is just about classical statistical mechanics; actually about a special mathematical aspect. It was one of the things I found particularly intriguing in my first encounters with statistical mechanics and thermodynamics a long time ago – a curious feature of volumes.

I was mulling upon how to ‘briefly motivate’ the calculation below in a comprehensible way, a task I might have failed at years ago already, when I tried to use illustrations and metaphors (Here and here). When introducing the ‘kinetic theory’ in thermodynamics often the pressure of an ideal gas is calculated first, by considering averages over momenta transferred from particles hitting the wall of a container. This is rather easy to understand but still sort of an intermediate view – between phenomenological thermodynamics that does not explain the microscopic origin of properties like energy, and ‘true’ statistical mechanics. The latter makes use of a phase space with with dimensions the number of particles. One cubic meter of gas contains ~1025 molecules. Each possible state of the system is depicted as a point in so-called phase space: A point in this abstract space represents one possible system state. For each (point-like) particle 6 numbers are added to a gigantic vector – 3 for its position and 3 for its momentum (mass times velocity), so the space has ~6 x 1025 dimensions. Thermodynamic properties are averages taken over the state of one system watched for a long time or over a lot of ‘comparable’ systems starting from different initial conditions. At the heart of statistical mechanics are distributions functions that describe how a set of systems described by such gigantic vectors evolves. This function is like a density of an incompressible fluid in hydrodynamics. I resorted to using the metaphor of a jelly in hyperspace before.

Taking averages means to multiply the ‘mechanical’ property by the density function and integrate it over the space where these functions live. The volume of interest is a  generalized N-ball defined as the volume within a generalized sphere. A ‘sphere’ is the surface of all points in a certain distance (‘radius’ R) from an origin

$x_1^2 + x_2^2 + ... + x_ {N}^2 = R^2$

($x_n$ being the co-ordinates in phase space and assuming that all co-ordinates of the origin are zero). Why a sphere? Because states are ordered or defined by energy, and larger energy means a greater ‘radius’ in phase space. It’s all about rounded surfaces enclosing each other. The simplest example for this is the ellipse of the phase diagram of the harmonic oscillator – more energy means a larger amplitude and a larger maximum velocity.

And here is finally the curious fact I actually want to talk about: Nearly all the volume of an N-ball with so many dimensions is concentrated in an extremely thin shell beneath its surface. Then an integral over a thin shell can be extended over the full volume of the sphere without adding much, while making integration simpler.

This can be seen immediately from plotting the volume of a sphere over radius: The volume of an N-ball is always equal to some numerical factor, times the radius to the power of the number of dimensions. In three dimensions the volume is the traditional, honest volume proportional to r3, in two dimensions the ‘ball’ is a circle, and its ‘volume’ is its area. In a realistic thermodynamic system, the volume is then proportional to rN with a very large N.

The power function rN turn more and more into an L-shaped function with increasing exponent N. The volume increases enormously just by adding a small additional layer to the ball. In order to compare the function for different exponents, both ‘radius’ and ‘volume’ are shown in relation to the respective maximum value, R and RN.

The interesting layer ‘with all the volume’ is certainly much smaller than the radius R, but of course it must not be too small to contain something. How thick the substantial shell has to be can be found by investigating the volume in more detail – using a ‘trick’ that is needed often in statistical mechanics: Taylor expanding in the exponent.

A function can be replaced by its tangent if it is sufficiently ‘straight’ at this point. Mathematically it means: If dx is added to the argument x, then the function at the new target is f(x + dx), which can be approximated by f(x) + [the slope df/dx] * dx. The next – higher-order term would be proportional to the curvature, the second derivation – then the function is replaced by a 2nd order polynomial. Joseph Nebus has recently published a more comprehensible and detailed post about how this works.

So the first terms of this so-called Taylor expansion are:

$f(x + dx) = f(x) + dx{\frac{df}{dx}} + {\frac{dx^2}{2}}{\frac{d^2f}{dx^2}} + ...$

If dx is small higher-order terms can be neglected.

In the curious case of the ball in hyperspace we are interested in the ‘remaining volume’ V(r – dr). This should be small compared to V(r) = arN (a being the uninteresting constant numerical factor) after we remove a layer of thickness dr with the substantial ‘bulk of the volume’.

However, trying to expand the volume V(r – dr) = a(r – dr)N, we get:

$V(r - dr) = V(r) - adrNr^{N-1} + a{\frac{dr^2}{2}}N(N-1)r^{N-2} + ...$
$= ar^N(1 - N{\frac{dr}{r}} + {\frac{N(N-1)}{2}}({\frac{dr}{r}})^2) + ...$

But this is not exactly what we want: It is finally not an expansion, a polynomial, in (the small) ratio of dr/r, but in Ndr/r, and N is enormous.

So here’s the trick: 1) Apply the definition of the natural logarithm ln:

$V(r - dr) = ae^{N\ln(r - dr)} = ae^{N\ln(r(1 - {\frac{dr}{r}}))}$
$= ae^{N(\ln(r) + ln(1 - {\frac{dr}{r}}))}$
$= ar^Ne^{\ln(1 - {\frac{dr}{r}}))} = V(r)e^{N(\ln(1 - {\frac{dr}{r}}))}$

2) Spot a function that can be safely expanded in the exponent: The natural logarithm of 1 plus something small, dr/r. So we can expand near 1: The derivative of ln(x) is 1/x (thus equal to 1/1 near x=1) and ln(1) = 0. So ln(1 – x) is about -x for small x:

$V(r - dr) = V(r)e^{N(0 - 1{\frac{dr}{r})}} \simeq V(r)e^{-N{\frac{dr}{r}}}$

3) Re-arrange fractions …

$V(r - dr) = V(r)e^{-\frac{dr}{(\frac{r}{N})}}$

This is now the remaining volume, after the thin layer dr has been removed. It is small in comparison with V(r) if the exponential function is small, thus if ${\frac{dr}{(\frac{r}{N})}}$ is large or if:

$dr \gg \frac{r}{N}$

Summarizing: The volume of the N-dimensional hyperball is contained mainly in a shell dr below the surface if the following inequalities hold:

${\frac{r}{N}} \ll dr \ll r$

The second one is needed to state that the shell is thin – and allow for expansion in the exponent, the first one is needed to make the shell thick enough so that it contains something.

This might help to ‘visualize’ a closely related non-intuitive fact about large numbers, like eN: If you multiply such a number by a factor ‘it does not get that much bigger’ in a sense – even if the factor is itself a large number:

Assuming N is about 1025  then its natural logarithm is about 58 and…

$Ne^N = e^{\ln(N)+N} = e^{58+10^{25}}$

… 58 can be neglected compared to N itself. So a multiplicative factor becomes something to be neglected in a sum!

I used a plain number – base e – deliberately as I am obsessed with units. ‘r’ in phase space would be associated with a unit incorporating lots of lengths and momenta. Note that I use the term ‘dimensions’ in two slightly different, but related ways here: One is the mathematical dimension of (an abstract) space, the other is about cross-checking the physical units in case a ‘number’ is something that can be measured – like meters. The co-ordinate  numbers in the vector refer to measurable physical quantities. Applying the definition of the logarithm just to rN would result in dimensionless number N side-by-side with something that has dimensions of a logarithm of the unit.

Using r – a number with dimensions of length – as base, it has to be expressed as a plain number, a multiple of the unit length $R_0$ (like ‘1 meter’). So comparing the original volume of the ball $a{(\frac{r}{R_0})}^N$ to one a factor of N bigger …

$aN{(\frac{r}{R_0})}^N = ae^{\ln{(N)} + N\ln{(\frac{r}{R_0})}}$

… then ln(N) can be neglected as long as $\frac{r}{R_0}$ is not extreeeemely tiny. Using the same argument as for base e above, we are on the safe side (and can neglect factors) if r is of about the same order of magnitude as the ‘unit length’ $R_0$. The argument about negligible factors is an argument about plain numbers – and those ‘don’t exist’ in the real world as one could always decide to measure the ‘radius’ in a units of, say, 10-30 ‘meters’, which would make the original absolute number small and thus the additional factor non-negligible. One might save the argument by saying that we would always use units that sort of match the typical dimensions (size) of a system.

Saying everything in another way: If the volume of a hyperball ~rN is multiplied by a factor, this corresponds to multiplying the radius r by a factor very, very close to 1 – the Nth root of the factor for the volume. Only because the number of dimensions is so large, the volume is increased so much by such a small increase in radius.

As the ‘bulk of the volume’ is contained in a thin shell, the total volume is about the product of the surface area and the thickness of the shell dr. The N-ball is bounded by a ‘sphere’ with one dimension less than the ball. Increasing the volume by a factor means that the surface area and/or the thickness have to be increased by factors so that the product of these factors yield the volume increase factor. dr scales with r, and does thus not change much – the two inequalities derived above do still hold. Most of the volume factor ‘goes into’ the factor for increasing the surface. ‘The surface becomes the volume’.

This was long-winded. My excuse: Also Richard Feynman took great pleasure in explaining the same phenomenon in different ways. In his lectures you can hear him speak to himself when he says something along the lines of: Now let’s see if we really understood this – let’s try to derive it in another way…

And above all, he says (in a lecture that is more about math than about physics)

Now you may ask, “What is mathematics doing in a physics lecture?” We have several possible excuses: first, of course, mathematics is an important tool, but that would only excuse us for giving the formula in two minutes. On the other hand, in theoretical physics we discover that all our laws can be written in mathematical form; and that this has a certain simplicity and beauty about it. So, ultimately, in order to understand nature it may be necessary to have a deeper understanding of mathematical relationships. But the real reason is that the subject is enjoyable, and although we humans cut nature up in different ways, and we have different courses in different departments, such compartmentalization is really artificial, and we should take our intellectual pleasures where we find them.

___________________________________

Further reading / sources: Any theoretical physics textbook on classical thermodynamics / statistical mechanics. I am just re-reading mine.

Learning General Relativity

Math blogger Joseph Nebus does another A – Z series of posts, explaining technical terms in mathematics. He asked readers for their favorite pick of things to be covered in this series, and I came up with General Covariance. Which he laid out in this post – in his signature style, using neither equations nor pop-science images like deformed rubber mattresses – but ‘just words’. As so often, he manages to explain things really well!

Actually, I asked for that term as I am in the middle of yet another physics (re-)learning project – in the spirit of my ventures into QFT a while back.

Since a while I have now tried (on this blog) to cover only the physics related to something I have both education in and hands-on experience with. Re General Relativity I have neither: My PhD was in applied condensed-matter physics – lasers, superconductors, optics – and this article by physicist Chad Orzel about What Math Do You Need For Physics? covers well what sort of math you need in that case. Quote:

I moved into the lab, and was concerned more with technical details of vacuum pumps and lasers and electronic circuits and computer data acquisition and analysis.

So I cannot find the remotest way to justify why I would need General Relativity on a daily basis – insider jokes about very peculiarly torus-shaped underground water/ice tanks for heat pumps aside.

My motivation is what I described in this post: Math-heavy physics is – for me, that means a statistical sample of 1 – the best way of brazing myself for any type of tech / IT / engineering work. This positive effect is not even directly related to math/physics aspects of that work.

But I also noticed ‘on the internet’ that there is a community of science and math enthusiasts, who indulge in self-studying theoretical physics seriously as a hobby. Often these are physics majors who ended up in very different industry sectors or in management / ‘non-tech’ jobs and who want to reconnect with what they once learned.

For those fellow learners I’d like to publish links to my favorite learning resources.

There seem to be two ways to start a course or book on GR, and sometimes authors toggle between both modes. You can start from the ‘tangible’ physics of our flat space (spacetime) plus special relativity and then gradually ‘add a bit of curvature’ and related concepts. In this way the introduction sounds familiar, and less daunting. Or you could try to introduce the mathematical concepts at a most rigorous abstract level, and return to the actual physics of our 4D spacetime and matter as late as possible.

The latter makes a lot of sense as you better unlearn some things you took for granted about vector and tensor calculus in flat space. A vector must no longer be visualized as an arrow that can be moved around carelessly in space, and one must be very careful in visualizing what transforming coordinates really means.

For motivation or as an ‘upper level pop-sci intro’…

Richard Feynman’s lecture on curved space might be a very good primer. Feynman explains what curved space and curved spacetime actually mean. Yes, he is using that infamous beetle on a balloon, but he also gives some numbers obtained by back-of-the-envelope calculations that explain important concepts.

For learning about the mathematical foundations …

I cannot praise these Lectures given at the Heraeus International Winter School Gravity and Light 2015 enough. Award-winning lecturer Frederic P. Schuller goes to great lengths to introduce concepts carefully and precisely. His goal is to make all implicit assumptions explicit and avoid allusions to misguided ‘intuitions’ one might got have used to when working with vector analysis, tensors, gradients, derivatives etc. in our tangible 3D world – covered by what he calls ‘undergraduate analysis’. Only in lecture 9 the first connection is made back to Newtonian gravity. Then, back to math only for some more lectures, until finally our 4D spacetime is discussed in lecture 13.

Schuller mentions in passing that Einstein himself struggled with the advanced math of his own theory, e.g. in the sense of not yet distinguishing clearly between the mathematical structure that represents the real world (a topological manifold) and the multi-dimensional chart we project our world onto when using an atlas. It is interesting to pair these lectures with this paper on the history and philosophy of general relativity – a link Joseph Nebus has pointed to in his post on covariance.

Learning physics or math from videos you need to be much more disciplined than with plowing through textbooks – in the sense that you absolutely have to do every single step in a derivation on your own. It is easy to delude oneself that you understood something by following a derivation passively, without calculating anything yourself. So what makes these lectures so useful is that tutorial sessions have been recorded as well: Tutorial sheets and videos can be found here.
(Edit: The Youtube channel of the event has not all the recordings of the tutorial sessions, only this conference website has. It seems the former domain does not work any more, but the content is perserved at gravity-and-light.herokuapp.com)

You also find brief notes for these lectures here.

For a ‘physics-only’ introduction …

… I picked a classical, ‘legendary’ resource: Landau and Lifshitz give an introduction to General Relativity in the last third of the second volume in their Course of Theoretical Physics, The Classical Theory of Fields. Landau and Lifshitz’s text is terse, perhaps similar in style to Dirac’s classical introduction to quantum mechanics. No humor, but sublime and elegant.

Landau and Lifshitz don’t need manifolds nor tangent bundles, and they use the 3D curvature tensor of space a lot in addition to the metric tensor of 4D spacetime. They introduce concepts of differences in space and time right from the start, plus the notion of simultaneity. Mathematicians might be shocked by a ‘typical physicist’s’ way to deal with differentials, the way vectors on different points in space are related, etc. – neglecting (at first sight, explore every footnote in detail!) the tower of mathematical structures you actually need to do this precisely.

But I would regard Lev Landau sort of a Richard Feynman of The East, so it takes his genius not make any silly mistakes by taking the seemingly intuitive notions too literally. And I recommend this book only when combined with a most rigorous introduction.