The answer to an intractable series is to stop using one. The lattice replaces the expansion with a computation nobody can check by hand, and that is the price of getting numbers at all.
🎯 Why this matters
Physics arriving as a computation fails differently from physics arriving as a formula. The uncertainties become discretisation, volume and extrapolation — engineering quantities, argued over the way engineering quantities are.Every section since §6.5 has ended in the same place: at low energy α_s is large, the perturbative series is useless, and the interesting quantities cannot be computed. This section is what you do about that.
§6.10 Computing without a series
Wilson’s 1974 idea, borrowed from the statistical physics of condensed matter: replace continuous infinite spacetime with a finite hypercubic grid — lattice QCD lattice QCD computing QCD by discretizing spacetime onto a finite grid and evaluating the path integral numerically. The only first-principles route to quantities where the coupling is too large for a perturbative series — hadron masses, the LO-HVP term in a_μ, the deconfinement transition. defined in §6.10-6.11 — open in glossary . The path integral becomes a very large but finite sum, and a finite sum can be evaluated numerically. No expansion in α_s appears anywhere — the method sums over all the relevant diagrams at once, which is exactly why it works where perturbation theory does not.
The data layout is worth stating precisely, because it is unusual and it explains where the cost goes:
- quark fields sit on the sites;
- gluon fields sit on the links between them, one 3 × 3 SU(3) matrix per link.
Gauge fields living on links rather than points is not an implementation convenience. It is what keeps the discretized theory exactly gauge invariant — a link matrix is precisely the object that transports colour from one site to its neighbour.
the lattice · two limits that pull against each other
Halve the spacing and the work goes up sixteenfold — the sites scale as (L/a)⁴ because spacetime has four dimensions, and the algorithms cost more than one operation per site on top of that. That single exponent is why lattice QCD is a story about hardware: sub-per-cent accuracy needs around 10¹⁰ operations, which arrived with petaflop machines, five orders of magnitude beyond a laptop.
Fig. 6.33, with the arithmetic live. The two sliders are the two limits that must be taken, and they pull against each other: a → 0 for the continuum, L → ∞ for the box. Push either and watch the site count — and the memory — move by four powers.
⚙️ Engineer’s bridge — an exponent of four, and why the physics waited for the hardware
Lattice QCD is the one place in this book where progress was gated by floating-point throughput, and the reason is a single exponent.
The number of sites is — four, because spacetime has four dimensions. Halve the spacing and you need sixteen times as many sites, before counting that the solvers cost more than one operation per site and that the condition number worsens as the quark masses fall towards their physical values. The book’s figure: sub-per-cent accuracy takes of order operations, and a petaflop machine is five orders of magnitude beyond a laptop.
Any engineer who has met a four-dimensional grid recognises the trap. It is the same reason a 3D FDTD simulation at half the cell size costs 16× (eight in space, two in time for the CFL condition), and the same reason people fight so hard for adaptive meshes. There is no algorithmic escape from a volume scaling law; you either buy the machine or you improve the discretization so a coarser grid suffices — which is precisely what “improved actions” are in this field.
There is also a correctness constraint that has no analogue in ordinary numerics. Discretizing a chiral theory naively produces spurious extra fermions — the doubling problem — and destroys the chiral symmetry that §6.9 showed is the reason the pion is light. Getting a discretization that keeps chirality was a decades-long problem, and the book calls it out as one of the major difficulties solved in the last decade. In software terms: the obvious discretization is fast and computes the wrong theory.
Where it breaks: you cannot refine until convergence and compare against the exact answer, because there is no exact answer to compare against — the continuum limit is an extrapolation, not a check. Two harder limits follow. The lattice works in Euclidean time, so quantities that are genuinely about real time — scattering phase shifts, transport coefficients — are not directly accessible and must be reached by analytic continuation. And at finite baryon density the integrand becomes complex, so importance sampling fails outright: the sign problem is why the interior of a neutron star is not a lattice calculation waiting on more CPU.
the cost of the continuum limit
hbarc, mpi = 0.1973269804, 0.13957 # GeV fm, GeV
Lmin = 4 * hbarc / mpi # the box must hold the pion cloud
print(f"the pion Compton wavelength is {hbarc/mpi:.3f} fm, so the box needs")
print(f"L >~ 4/m_pi = {Lmin:.2f} fm (the book quotes ~6.3 fm for the BMW lattice)")
print()
L = 6.3
print(f"at L = {L} fm, sites = (L/a)^4:")
a = 0.125
for k in range(5):
aa = a / 2**k
n = round(L / aa)
print(f" a = {aa:.4f} fm -> {n:4d} per side -> {n**4:.2e} sites"
f" ({16**k:>6d}x the a = 0.125 cost)")
print()
print("Four dimensions means halving a costs 16x. There is no algorithmic")
print("escape from a volume scaling law.")
print()
n = round(L / a)
links = 4 * n**4
print(f"and the memory: {links:.3e} links x one SU(3) matrix (18 doubles) each")
print(f" = {links*18*8/1e9:.1f} GB for ONE gauge configuration,")
print(" and a calculation needs thousands of them.")the pion Compton wavelength is 1.414 fm, so the box needs L >~ 4/m_pi = 5.66 fm (the book quotes ~6.3 fm for the BMW lattice) at L = 6.3 fm, sites = (L/a)^4: a = 0.1250 fm -> 50 per side -> 6.25e+06 sites ( 1x the a = 0.125 cost) a = 0.0625 fm -> 101 per side -> 1.04e+08 sites ( 16x the a = 0.125 cost) a = 0.0312 fm -> 202 per side -> 1.66e+09 sites ( 256x the a = 0.125 cost) a = 0.0156 fm -> 403 per side -> 2.64e+10 sites ( 4096x the a = 0.125 cost) a = 0.0078 fm -> 806 per side -> 4.22e+11 sites ( 65536x the a = 0.125 cost) Four dimensions means halving a costs 16x. There is no algorithmic escape from a volume scaling law. and the memory: 2.500e+07 links x one SU(3) matrix (18 doubles) each = 3.6 GB for ONE gauge configuration, and a calculation needs thousands of them.
🔢 Worked example — three inputs, and everything else predicted
The Budapest–Marseille–Wuppertal calculation (Dürr et al. 2008) computed the light hadron spectrum ab initio using exactly three experimental numbers.
QCD cannot predict masses in physical units at all — it is a theory with one dimensionful parameter and no metre stick. Only dimensionless combinations, i.e. mass ratios, are calculable. So three inputs are needed and no more:
- the pion mass → fixes the average u, d mass;
- the kaon mass → fixes the s quark mass;
- the Ξ mass → sets the overall physical scale.
The third choice is the interesting one. Why the Ξ? Because it is in the octet, where the calculation is more precise than for the decuplet; and within the octet it has the most strange quarks, which makes its mass least sensitive to the delicate chiral behaviour of the light quarks. It is chosen to be the least fragile available yardstick, not the most familiar one.
Everything else in Fig. 6.34 — the ρ, K*, N, Λ, Σ, Δ, Σ*, Ξ*, Ω — is then a prediction with no further freedom, and they land on the measured values within a few per cent.
- measured
- lattice QCD (BMW 2008)
💡 What this really says — the debt from §5.9b, paid here
§5.9b left an open account. The muon’s magnetic anomaly is measured to 0.35 ppm, but the comparison with theory is limited by one number — the leading-order hadronic vacuum polarization, , which is 95 % of the theoretical error bar and cannot be computed perturbatively because it lives at low energy where α_s is large.
This is where it is computed. The same BMW collaboration, with the same method, produced — Eq. (5.77), quoted in Chapter 5 and derived here.
And the tension noted there stands: the older data-driven route, integrating the measured hadrons cross-section of §5.7 through a dispersion relation, gives a smaller value, which would push the muon discrepancy from 1.5σ to roughly 4σ. The two methods disagree by more than either claims. Whether there is new physics in the muon anomaly currently turns on which of these two calculations of a hadronic loop is right — one of them a lattice computation of the kind this section describes.
It is a good illustration of what lattice QCD is for, and of what it costs: the frontier of the most sensitive test of the Standard Model is a numerical disagreement about a low-energy QCD quantity.
§6.11 Melting a hadron
Squeeze nuclear matter hard enough and the quarks stop belonging to individual nucleons. Above a critical temperature they move freely over distances much larger than a hadron: the quark–gluon plasma quark–gluon plasma the deconfined phase of quarks and gluons that filled the first microsecond of the universe, recreated in heavy-ion collisions at RHIC and the LHC; behaves as a near-perfect liquid, not a gas. defined in §6.10-6.11 — open in glossary , the state of the universe for its first microsecond.
The mechanism is §6.5 run in reverse. Compress the matter, the inter-quark distances fall, α_s falls with them, and confinement — which is a long-distance phenomenon — simply stops applying. Deconfinement is asymptotic freedom achieved by crowding rather than by collision energy.
how dense, and how flat
rho_n = 0.29e45 # nucleons per m^3 in a Pb nucleus
m_N = 1.6749e-27 # kg
print("nuclear matter against water, as MASS density:")
print(f" Pb nucleus: {rho_n*m_N:.2e} kg/m^3")
print(f" water: {1.0e3:.2e} kg/m^3")
print(f" ratio = {rho_n*m_N/1e3:.1e} -- 14 orders of magnitude, as the book says")
print()
# Question 6.6: a Pb nucleus of 5.5 fm radius, Lorentz-contracted along the beam
mN, R = 0.9315, 5.5
for E, lab in [(2.5e3, "2.5 TeV PER NUCLEON -- the LHC convention"),
(2.5e3/208, "2.5 TeV for the whole nucleus")]:
g = E / mN
print(f"Question 6.6, reading {lab}:")
print(f" gamma = {g:8.1f}, contracted diameter = {2*R:.0f} fm / gamma = {2*R/g:.4f} fm")
print()
print("Heavy-ion beams are always quoted per nucleon, so the first is meant:")
print(f" an 11 fm nucleus becomes {2*R/(2.5e3/mN)*1000:.1f} attometres thick --")
print(f" about {2*0.84/(2*R/(2.5e3/mN)):.0f} times thinner than a single proton is wide.")
print(" Two of these pass through each other; that is the collision.")nuclear matter against water, as MASS density: Pb nucleus: 4.86e+17 kg/m^3 water: 1.00e+03 kg/m^3 ratio = 4.9e+14 -- 14 orders of magnitude, as the book says Question 6.6, reading 2.5 TeV PER NUCLEON -- the LHC convention: gamma = 2683.8, contracted diameter = 11 fm / gamma = 0.0041 fm Question 6.6, reading 2.5 TeV for the whole nucleus: gamma = 12.9, contracted diameter = 11 fm / gamma = 0.8525 fm Heavy-ion beams are always quoted per nucleon, so the first is meant: an 11 fm nucleus becomes 4.1 attometres thick -- about 410 times thinner than a single proton is wide. Two of these pass through each other; that is the collision.
💡 What this really says — a liquid, not a gas — which nobody expected
The natural guess for deconfined matter is a gas: asymptotic freedom says the constituents barely interact, so they should behave like free particles rattling around.
They do not. RHIC and then the LHC found the QGP behaves as a liquid with remarkably low viscosity — close to the lowest any known substance achieves, and close to a bound conjectured from string theory. The constituents are strongly coupled, not weakly.
The resolution is the one this whole chapter has been circling: asymptotic freedom applies at short distance, and the plasma’s collective behaviour is governed by what happens at the inter-particle spacing, which is not short enough. A deconfined quark is free of its own hadron; it is not free of the medium.
| what it is | why it is needed | |
|---|---|---|
| heavy ions | Au+Au at RHIC (√s_NN = 200 GeV); Pb+Pb, Xe+Xe at the LHC (2.76 then 5 TeV) | a single pp collision makes a tiny fireball; you need a large nucleus to reach the volume and energy density where a plasma phase makes sense |
| centrality | selected by multiplicity — the number of detected particles | collisions happen at every impact parameter, and only the head-on ones make plasma. Multiplicity is the proxy you actually have |
| pseudorapidity η = −ln tan(θ/2) | a remapping of the polar angle, Eq. (6.70) | its distribution is much flatter than θ's in the central region, so equal bins of η carry comparable statistics — a variable chosen for the shape of its distribution, not its physical meaning |
| the central region | |η| ≲ 1.4 | where roughly equal numbers of quarks and antiquarks are produced, so the matter resembles the early universe rather than a beam remnant |
| the timescale | a few × 10⁻²³ s | the fireball expands and cools within that, so everything detected is a hadron that froze out afterwards. The plasma is never observed directly |
The last row is the honest one: nobody has ever detected a quark–gluon plasma. What is detected is hadrons, long after the fact, and the plasma is inferred from their statistical properties — the same epistemic position as the jets of §6.1, one level further removed.
🔑 If you remember only three things
-
It pays a debt left open in Chapter 5. The hadronic term limiting the muon’s anomaly is precisely the kind of number no other method can supply.
-
A prediction can be gated by hardware. Four powers of the lattice spacing meant this physics waited on throughput rather than on anybody’s insight.
-
Deconfined matter turned out to be a liquid. Asymptotic freedom suggested a gas of nearly free particles, and the measurement said otherwise.
Where this goes next
- §5.9b is where this section’s lattice number is used, and where the disagreement between the lattice and data-driven evaluations decides whether the muon anomaly is new physics.
- §6.7 estimated the proton mass to a factor; Fig. 6.34 computes it to a per cent, by the method described here.
- §6.9’s chiral symmetry is what makes the discretization hard — the naive lattice destroys it, and keeping it took decades.
- §9.14 is the LHC, whose heavy-ion programme and the ALICE experiment are where the plasma is made.
✅ Check yourself — lattice QCD and the quark–gluon plasma
0/5 answered · 0 correct
1.Why must gluon fields live on the links of the lattice rather than on the sites, where the quark fields sit?
2.Halving the lattice spacing multiplies the work by sixteen. What is the exponent, and is there a way around it?
3.The BMW calculation uses exactly three experimental inputs. Why three, and why is the Ξ one of them?
4.What debt from Chapter 5 does §6.10 settle, and is it settled?
5.The quark–gluon plasma behaves as a low-viscosity liquid, not a gas. Why is that surprising, and what resolves it?