----------------------------------------------------------------
* pmzize by num chg / N, instead of # sweeps?  somehow?
* pseudocode to daniel
! discontinuity in num_chg at T=6 ... ?!?

----------------------------------------------------------------
Daniel Mon. 4-23:

k doubles everywhere! :)
k sync2rcm
k bubble around pi(x) instead of x.
k Delta_H after y == ix test.
k get dist sq ...
k inline zero_centered_mod ...
* du reader's guide ... :)
* lattice program ... :(
* ri=1 pixb=0,1
* cache sqdist in bubbles ... use a hash??
! box structure up front

* occ dens verbosity or some such.
! rho discontinuity is *not* OK.  This needs rethinking!

* tune bubble frac to get ~0.5 acceptance rate.  This is OK
  now that we have pix bubbles. :)

* E[other RVs]:  counts?
  k code up
  - to DU
* Noise in <rho[k]> -- law of small numbers?
k For rho_infty, locate the 45-degree outer corner. :)
* test Var[rho_infty] as function of T, L, and brad.
>> Point:  higher acceptance rate + less cornery determination of rho_infty_T.

----------------------------------------------------------------
Daniel Mon. 4-21:

! above T_c, 45-degree reduction gives a bogus answer (no x-intercept).
  How to fix this continuously?

* cache dist in metro windows?

k re-examine number of keeps and changes ... ?

k drift/corr plots for T=4.0, 6.4, 6.8
k subtract off slope=1 line segment from rho cut
k autocorrelation: vary spa and plot it.
  - NOT thermalization but correlation.
k tune spa and nr for worst-case parameter values.
  - check against ri=0/1 -- do they meet?

* need to rework metro_windows for each x move ...
* n2fit to get T_c's and gamma's.  Need to visually trim.
* Tplot

* Work through the algebra for L fit.
* Code up the L fit.
--- OR ---
* is extrapolation enough?
* collect data for L=36, L=50 ...

* need to code up point motion & reacquire all the data.

* T=8.0 negative rho fit ... when is and when is not the fit appropriate?
  - check k1 and k2.

----------------------------------------------------------------
* mmap v2db.dat

k tabulate two-cycle perf
k make a cheat sheet for alpha_GRU vs. T and a vs. alpha
k implement T option in mcrcm.c
k phi(alpha) -> rho_infty(T)

k re-do plots for T vs. rho_infty(T) (rather than alpha_GRU and phi(alpha)).
k make log-log plots for T vs. rho_infty(T)
* power-law fit without point motions

* metropolize point motions
* power-law fit with point motions
  - (Tc-T)^gamma for some gamma (critical exponents)
    for each L; as L to infty, TcL to Tc
  - power-law fit for TcL and then extrapolate

* longest cycles etc ...

* tabulate estimated CPU time to find T_c(a).
  - L = 10,12,14,16            (4)
  - a = 0.01, 0.05, 0.08, 0.10 (4)
  - T = 4.8 5.2 5.6 6.0 6.4    (5)
  - Total 80 sim runs.
  - 10,000 seconds per run (xxx time this!!) ~= 3 processor-hours.
  - So, 240 processor-hours = 10 processor-days.
  - Doable in a couple of days using four processors.

* plot3d for dots
  - http://www.scipy.org/Cookbook/Matplotlib/mplot3D

----------------------------------------------------------------
* Make a done.txt
!! points_t:  1:N, 0:N, 0:N-1.  Comment well! :)

!! send_xy_delta -> just delta_n2 :)

* doc:
  - P(pi):  along the lines of rc2.txt. :)
  k long slides to paper! :)
  ! loop fig to paper.

* Python:
  - gif concatenator
  - gif maker
  - stats/histo/pdf :)
  - xgr replacement?!?
  - plot maker for H and rho
  - dot-plots

----------------------------------------------------------------
Interactions:

* n2:
  ! quantify the right alphas
  - with randinit=0, it's hard to leave the identity
  - with randinit=1 (even in the non-interacting case), initial energy is
    *very* high and the default thermalization criterion is premature.
	The effect is more pronounced at higher temperatures.
  - I need to get the same phi(alpha) results regardless of initial state!!!
  - Maybe run simultaneously with randinit=0 *and* randinit=1, and consider
    thermalization to have occurred when the two cross?  Or, at least, run
	two separate progress with ri=0/1 and experiment to find the right
	thermalization criterion.
  - pretherm w/ unif perm on metro windows.  randinit=2.  :)
  - plot phi(alpha_GRU) curves against T for various a's.

* n2 article:
  k reproduce GRU T_c(alpha)
  - write that in terms of beta.  Or T.
  - find BU T_c(alpha) for N2
  - Write it up! :)
  - find MH T_c(alpha) for TJI
  - Write it up.

* v2metro:
  !!! with V(*') fix, there are far more long-r table misses.  Damn!
  - average nnei printf
  n x,x',y,y' -> x_1, x_2, y_1, y_2.
  k check derivation of u<y iy x etc.
  k r and theta cutoffs?
  ! distribution of theta?
  k V(x,x',y,y') --> V(x,y,x',y').  CODE AND DOC!
  k decimate rho plots
  k v2db extents.  more r?
  k v2db dups?  enhance metro windows?
  k get_Delta_V:  restrict to metro windows?  MW radius?
  - v2db hit/miss hook for printfs at end.
  n separate r1/r2 meshes w/ r2 longer?  & r1<r2 lookup mod?
  - v2db nr+1 et al.?
  - layer the v2 and v2db.
  k smoothwin as arg
  k forcetherm=%d arg

* V2I:
  k check against special cases
  k exp(-A/s): matricize in Matlab
  k unloop test_* and plot_*
  k C code:  rename test_V2 to sth else ...
  k C code:  V2(r1, r2, theta).
  k C code:  V2r_exact.
  k C code:  one-shot xyT_bridge after all.
  k test_V2_IX.c: check exact against integral.
  k test_V2_IB.c: loose bridge, with tabulation.
  k test_V2_IB1.c: tight bridge for spec r1, r2, theta, to check dbV against duV
  k test_brdb_slice xyt:  r1, r2, theta?
  k xyt taps look too hoppy.
  ! xyT remap scale is T^(1/2) not T^(d/2)!?!
  k nb too small:  affects v2i variance?
    > test_V2_IB1 r1=1 r2=1 theta=180 np=?? nb=??
    > np too small:  v2i drift??
    > integral is 0.20000
    > np=100:   consistently about 0.03 regardless of nb
    > np=1000:  consistently about 0.13 regardless of nb
    > np=10000: consistently about 0.19 regardless of nb
    > for this r1/r2/theta, it takes about 500,000 np to get close to the
      integral.
    > ... so now I guess I trust the integral?
    k tell daniel about this

* Bridge:
  k check dbV against non-db V
  k move to r1/r2/theta

  k doc:  1/sqrt(n) in variance of sample mean.  ramifications
    for bridge method. :(
  k hard-core U -- re-write entirely!!
  k separate out Vpi, V4, V2.
  k random walk vs. normal?
  k test mu/var/cov of xyT bridges.
  k integral numpts & avgreps are still silly.  compute variances.
    90 sec vs. 47 for L=3 d=2 a=1
  - data file
  k reuse bridges;
  - V interpolation from brdbs.
  k split out remap function w/ separate src and dst.
  n write bessel bridges to disk.  no need to compute norm squared on the fly.
  k brdb_populate -> brdb_populate_zz1
  k test_brdb_slice leak!

k show only cycles of minimum length?

* Bridge database:
  k d=1,2,3:  write floats.
  k bin2text and text2bin exes.
  k d into hdr.

n Special function for V:
  - Yor
  - Handbook of Brownian Motion
  - Burdzy
  - Gikhman and Kiefer?  bridge?
  - Pitman 534.pdf?

k DPRINTF for DEBUG_V

k lib.mk

* Bridge:
  k test gasdev w/ histogram
  k test gasdev_ms w/ histogram
  k test gasdevd w/ plot?
  k test gasdev_dms w/ plot?
  k brownian
  k bridge
  k U() functions ... U-type enums ... ?
  k integrate U(w(s)) from 0 to 4beta.
  k interaction_type into get_H_of_pi
  k de-stub the metropolizer
  k only compute H in mcrcm if asked to.
  k track system energy in delta H and metropolizer.
  k code up / modify a track-H tester.
  k make a new test_delta_h.c derived from test_energy.c.

k 2-cycle model in section 5
k need alpha and beta separately now.
  ~ alpha -> 1/4beta; gru_alpha cmd-line opts.
  - new alpha
k optimize get_n2 -> get_n2_delta, avoiding the swap.
  Track the number of cycles throughout the simulation.
  # With L=7 d=3 alpha=0.8:
  # n0:       1.658 sec
  # n2 slow: 29.844 sec
  # n2 fast:  3.347 sec (9x speed-up)
k pmt_swap_xy is broken; remove it.
k test_n2
k cases 2a, 2b?  doc.
k two-cycle init: not 0 but compute, in case random.
n optionally safetyize n2 in mcrcm.c.

----------------------------------------------------------------
Non-interacting:

k alpha=1.68 in table1:
  - kfrac 0.001?  Or what??  Maybe just vis?
  - L=25 to L=50?

k Reproduce figure 2.
* Reproduce figure 3.
* Reproduce figure 4.
* Reproduce figure 5.
k Reproduce figure 6.
k Reproduce figure 7.
k Reproduce table  1.

k phi(alpha)
  k phi(alpha) ~ 1 - <rho(N)>?  Just pick N close to 0.  Make it
    argified -- of course. :)

k Metro windows -- O(N^2) storage, or constant recomputation?  Is there a
  third choice?  bit-mask adjacency matrix?  print a warning if the incipient
  malloc exceeds a high-water mark?
k run on chivos?

k re-do code comments for rho.  He now uses an ell rather than a chi.
k Code up <lmax>

----------------------------------------------------------------
Math:

* Is DU interested in / aware of combinatorics of which cycle
  types predominate as N increases?

* For zero temp, <rho> ~= L/N?  Prove it.

* Constancy of rho along cycle types -- prove it and write it up.
  Also constancy of <rho> along cycle types at zero temperature.
  Follows by linearity of expectation?

k Side note:  with init=id, H *rises* during thermalization;
  with init=unifrand, H *falls* during thermalization.

----------------------------------------------------------------
Software as application:

* get/set:  use set for void functions.

k Downsample <rho> for xgr purposes.  One million is ridiculous.

* Make a tabulate option for P*, Z, P for N <= PMT_SN_MAX.

k Fix my thermalization-detection criterion to make it reasonable for high
  alpha ...

* print-pre-therm option

n Examine the fortran code?

k Then, code up the interacting case from that

* Well-comment the bounds of rho and <rho>: 0 .. N.

k Reproduce non-interacting graphs from condmat paper
k rho and <rho> plots
n bitwise adjacency matrix for metro windows?
k for non-trivial metro windows, implement windowing per DU comment
k remove/replace trivial metro windows.  This slows down thermalization.
k arg-ify N/L/d, temp/alpha
k code and test cycle-type computation & display
k averaged rho
k #ifdef for trivial include-all metro windows.
k Thermalization criterion -- ck w/ DU
k Sample size for <rho>.  Pre/post thermalization niter?
k graph delta rho
k spec psdes seed, rand48 seed:  sran32, srand48

----------------------------------------------------------------
Software management:

r groom the directories on gadsden, mathnet, and stick.
! burn a CD.

* GPL all the C and Python code

* rename?  pmt_free et al. vs. free_pmt et al.

k py/c report:
  - 11.2 sec vs. 0.240 sec for L=7 d=3 nsw=1000; 46x
  - prof shows 40% of C time in psdes_hash_64 & 40% in get_dist_sq.

k restore valgrind
k ctags -> gadsden
k make debug, opt, prof
k call deconstructors in all test proggies & re-valgrind
k fold psdes mods back into src/gen and re-publish.
k rcmrand.h with #define's
k profile and optimize

r replace psdes?
r valgrind for all test programs, w/ opt & debug, w/ all print options

(r means done, but can & should be re-done periodically)

----------------------------------------------------------------
Documentation:

* gprof how-to's, valgrind how-to's, packages needed, -Wall -Werror.
  Full details into rcm1.tex.

* Usage of random-number generator:
  - unit uniform float
  - integer mod a specified N

* make opt etc.

----------------------------------------------------------------
Reading:

* Martin & Rothen's BEC section

* Layman description:  "What are you working on?"  Make sure it's accessible
  AND accurate.

* Sims/diffs <--> superfluids, superconductivity?

* 541 stat mech stuff ...

* Feynman-Kac ...............
  - Faris notes?
  - Get Ginibre [5]

----------------------------------------------------------------