Imagine a walker on squared paper, choosing at each step to move up, down, left, or right. An ordinary random walk may return to the same place again and again. A self-avoiding walk requires the entire trace to remain free of self-intersections at lattice sites. This small restriction gives the path a memory: the next move depends not only on the present position, but on all of the past.
Self-avoidance does not mean “never turn back.” It means “never forget.” A locally generated path is forced to remember its entire history.
A path that remembers where it has been
1. From random walk to self-avoidance
On the \(d\)-dimensional integer lattice \(\mathbb Z^d\), a nearest-neighbour path of length \(n\) is a sequence
If the vertices \(\omega_0,\ldots,\omega_n\) are all distinct, the path is a self-avoiding walk (SAW). It may turn and it may pass close to itself; the one forbidden act is to step on a site it has already visited.
This prohibition destroys the most convenient feature of ordinary random walk. The increments of a simple random walk are independent, whereas the next step of a self-avoiding path depends on its complete history. It is neither an independent-increment process nor a Markov chain described by the current position alone. The difficulty comes not from having many rules, but from one rule coupling the whole path.
The same number of steps, two geometries
In the experiment above, an ordinary walk repeatedly folds over and covers its old trace, while self-avoidance swells the path. The self-avoiding samples are produced by the pivot algorithm: choose a pivot vertex, rotate or reflect the remaining tail, and accept the transformation only if the resulting path is still self-avoiding.
2. Why polymers trace such paths
The classical physical interpretation comes from polymer science. Think of a polymer as a long chain of monomers, with lattice edges representing the bonds between consecutive monomers. Matter cannot occupy the same place twice. In the simplest lattice model, this excluded-volume effect is exactly the self-avoidance constraint.
For ordinary random walk, the typical end-to-end distance scales as \(R_n\asymp n^{1/2}\). Self-avoidance stretches the chain, and one expects
where \(\nu\) is a critical exponent encoding the geometry of the path. In two dimensions, theory predicts \(\nu=3/4\). In three dimensions, the best numerical estimates are close to \(0.5876\). In dimensions five and above, the mean-field value \(\nu=1/2\) holds, while dimension four is critical and carries logarithmic corrections.
FLORY’s scaling argument gives an elegant explanation for the swelling. If a chain of length \(n\) occupies a region of size \(R\), its elastic cost is of order \(R^2/n\), while the cost of self-collisions is of order \(n^2/R^d\). Balancing the two gives
This heuristic gives the expected value \(3/4\) in two dimensions and is remarkably close to the numerical value in three dimensions. It is not a proof, and in high dimensions it must give way to mean-field behaviour. Its strength is conceptual: geometry emerges from a balance between entropy and repulsion.
3. How many paths are there?
Fix the starting point at the origin and let \(c_n\) denote the number of \(n\)-step self-avoiding walks. On the square lattice, the first values are
| \(n\) | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 |
|---|---|---|---|---|---|---|---|---|---|
| \(c_n\) | 1 | 4 | 12 | 36 | 100 | 284 | 780 | 2172 | 5916 |
Cutting an \((n+m)\)-step walk after step \(n\) produces an \(n\)-step self-avoiding walk and a translated \(m\)-step self-avoiding walk. Hence
This submultiplicativity, together with FEKETE’s lemma, guarantees the existence of
The number \(\mu\) is the lattice’s connective constant; it measures the exponential growth of the number of available paths. A more refined two-dimensional prediction is \(c_n\sim A\mu^n n^{\gamma-1}\), with \(\gamma=43/32\), but this exponent too remains unproved for the standard lattice models.
4. When counting becomes a critical phenomenon
Put all lengths into a single generating function:
Since \(c_n\) grows roughly like \(\mu^n\), this series has radius of convergence \(z_c=1/\mu\). The parameter \(z\) may be read as an activity per step. When \(z<z_c\), long paths are suppressed. As \(z\) approaches \(z_c\), arbitrarily large scales become relevant and both typical length and spatial extent diverge. A combinatorial counting problem has acquired a critical point in the sense of statistical mechanics.
This viewpoint also explains why the self-avoiding walk, the \(n\to0\) limit of the \(O(n)\) model, and polymer criticality share the same family of exponents. The path is no longer merely a path; it is a geometric observable of a critical statistical system.
5. An exact constant on the honeycomb lattice
Connective constants rarely have a known closed form. Even on the familiar square lattice, only high-precision numerical estimates are available. The celebrated exception is the honeycomb lattice. In 1982, Bernard NIENHUIS used Coulomb-gas arguments to predict
In 2012, Hugo DUMINIL-COPIN and Stanislav SMIRNOV proved the formula rigorously. They constructed a complex-valued parafermionic observable whose contributions from paths in different directions cancel through an exact local identity. The striking point is that a problem of integer enumeration is solved by a piece of discrete complex analysis.
Hugo DUMINIL-COPIN
Stanislav SMIRNOV
6. The planar limit and SLE\(_{8/3}\)
Shrink the lattice spacing while magnifying the path to a visible scale. Does the discrete polygonal line converge to a continuous random curve? The leading conjecture in two dimensions is that a suitably defined scaling limit of self-avoiding walk is Schramm–Loewner evolution with parameter \(\kappa=8/3\), or SLE\(_{8/3}\).
The Hausdorff dimension of SLE\(_{8/3}\) is
exactly matching the prediction \(1/\nu=4/3\). Conformal invariance, the restriction property, and the critical exponents therefore form a remarkably coherent picture. Yet, as of this writing, convergence of the standard two-dimensional self-avoiding walk to SLE\(_{8/3}\) remains open.
7. What dimension changes
Two dimensions: conjectural conformal geometry
The predictions \(\nu=3/4\), \(\gamma=43/32\), and SLE\(_{8/3}\) fit together precisely, but the central scaling-limit statement is still unproved.
Three dimensions: the physical polymer world
There is neither the conformal toolkit of the plane nor the weak-intersection simplification of high dimension. The sharpest information comes mainly from large-scale simulation and renormalisation-group methods.
High dimensions: a return to Gaussian behaviour
Different pieces of the path are less likely to meet. The lace expansion developed by HARA and SLADE rigorously yields mean-field critical behaviour in dimensions five and above.
Dimension four is the upper critical dimension. The leading exponent returns to \(\nu=1/2\), but logarithmic corrections remain. It is the watershed between the non-trivial geometry shaped by strong self-repulsion below and the Gaussian leading scale above.
Conclusion: the full memory of a path
The self-avoiding walk has a one-line definition, yet it contains three competing questions. Combinatorics asks how many paths there are; probability asks what a typical path looks like; statistical mechanics asks which features survive as the scale grows. The connective constant answers the first question, critical exponents describe the second, and a scaling limit would complete the third.
What makes the subject beautiful is that these answers do not arrive together. We know the exact exponential growth rate on the honeycomb lattice, yet still do not know whether a typical planar path converges to SLE\(_{8/3}\). We can prove Gaussian behaviour in high dimensions, yet the most visual cases—two and three dimensions—remain elusive. The theory resembles its own paths: the rule is clear, the direction is open, and every step carries the whole history behind it.
References and further reading
- J. M. HAMMERSLEY, Percolation processes II. The connective constant (1957).
- N. MADRAS and A. D. SOKAL, The pivot algorithm: A highly efficient Monte Carlo method for the self-avoiding walk (1988).
- T. HARA and G. SLADE, Self-avoiding walk in five or more dimensions I. The critical behaviour (1992).
- H. DUMINIL-COPIN and S. SMIRNOV, The connective constant of the honeycomb lattice equals \(\sqrt{2+\sqrt2}\) (2012).
- G. F. LAWLER, O. SCHRAMM and W. WERNER, On the scaling limit of planar self-avoiding walk (2004).
- N. CLISBY, Accurate estimate of the critical exponent \(\nu\) for self-avoiding walks (2010).
- N. MADRAS and G. SLADE, The Self-Avoiding Walk (1993).