foscat 2025.8.3__py3-none-any.whl → 2025.9.1__py3-none-any.whl

foscat/HealBili.py

@@ -0,0 +1,305 @@
+"""
+HealBili: Bilinear weights from a curvilinear (theta, phi) source grid to arbitrary HEALPix targets.
+
+This module provides a class `HealBili` that, given a *curvilinear* source grid of angular
+coordinates (theta[y,x], phi[y,x]) on the sphere (e.g., a tangent-plane grid built on an ellipsoid),
+computes **bilinear interpolation weights** to map values from that grid onto arbitrary target
+directions specified by HEALPix angles (heal_theta[n], heal_phi[n]).
+
+Key idea
+--------
+Because the source grid is not rectilinear in index-space, we cannot assume a simple affine mapping
+from (i,j) to angles. Instead, for each target direction (theta_h, phi_h), we:
+  1) locate a nearby source cell (seed) by nearest neighbor search on the unit sphere;
+  2) consider up to 4 candidate quads around the seed: [(i0,j0),(i0+1,j0),(i0,j0+1),(i0+1,j0+1)];
+  3) project the 4 corner unit vectors and the target onto a **local tangent plane** built at the
+     quad barycenter;
+  4) *invert* the bilinear mapping f(s,t) from the quad corners to the plane point using Newton,
+     retrieving (s,t) in [0,1]^2;
+  5) build the 4 bilinear weights [(1-s)(1-t), s(1-t), (1-s)t, st] and the 4 linear indices
+     into the source image (row-major, j*W + i).
+
+If no candidate quad cleanly contains the point, we choose the one with the smallest residual in the
+plane and clamp (s,t) to [0,1].
+
+The code is NumPy-only by default, but can optionally use `scipy.spatial.cKDTree` for a faster nearest
+neighbor seed search by setting `prefer_kdtree=True` (falls back automatically if SciPy is absent).
+
+Usage
+-----
+>>> hb = HealBili(src_theta, src_phi, prefer_kdtree=True)
+>>> I, W = hb.compute_weights(heal_theta, heal_phi)
+>>> # Apply to a source image `img` of shape (H,W):
+>>> vals = hb.apply_weights(img, I, W)  # shape (N,)
+
+All angles must be in **radians**. theta is colatitude (0 at north pole), phi is longitude in [0, 2*pi).
+"""
+from __future__ import annotations
+
+from typing import Tuple
+import numpy as np
+
+try:
+    from scipy.spatial import cKDTree  # optional
+    _HAVE_SCIPY = True
+except Exception:  # pragma: no cover
+    cKDTree = None
+    _HAVE_SCIPY = False
+
+
+class HealBili:
+    """Compute bilinear interpolation weights from a curvilinear (theta, phi) grid to HEALPix targets.
+
+    Parameters
+    ----------
+    src_theta : np.ndarray, shape (H, W)
+        Source **colatitude** (radians) at each grid node.
+    src_phi : np.ndarray, shape (H, W)
+        Source **longitude** (radians) at each grid node.
+    prefer_kdtree : bool, default False
+        If True and SciPy is available, use cKDTree on unit vectors for a faster nearest-neighbor seed.
+        Falls back to blocked brute-force dot-product search otherwise.
+    """
+
+    def __init__(self, src_theta: np.ndarray, src_phi: np.ndarray, *, prefer_kdtree: bool = False) -> None:
+        if src_theta.shape != src_phi.shape or src_theta.ndim != 2:
+            raise ValueError("src_theta and src_phi must have the same 2D shape (H, W)")
+        self.src_theta = np.asarray(src_theta, dtype=float)
+        self.src_phi = np.asarray(src_phi, dtype=float)
+        self.H, self.W = self.src_theta.shape
+        # Precompute unit vectors of source grid nodes
+        self._Vsrc = self._sph_to_vec(self.src_theta.ravel(), self.src_phi.ravel())  # (H*W, 3)
+        self.prefer_kdtree = bool(prefer_kdtree) and _HAVE_SCIPY
+        if self.prefer_kdtree:  # optional acceleration
+            self._kdtree = cKDTree(self._Vsrc)
+        else:
+            self._kdtree = None
+
+    # -----------------------------
+    # Public API
+    # -----------------------------
+    def compute_weights(
+        self,
+        heal_theta: np.ndarray,
+        heal_phi: np.ndarray,
+    ) -> Tuple[np.ndarray, np.ndarray]:
+        """Compute bilinear weights/indices for target HEALPix angles.
+
+        Parameters
+        ----------
+        heal_theta, heal_phi : np.ndarray, shape (N,)
+            Target **colatitude** and **longitude** in radians.
+
+        Returns
+        -------
+        I : np.ndarray, shape (4, N), dtype=int64
+            Linear indices of the 4 source corners per target (row-major: j*W + i). Invalid corners are -1.
+        W : np.ndarray, shape (4, N), dtype=float64
+            Bilinear weights aligned with `I`. Weights are set to 0.0 for invalid corners and normalized to sum to 1
+            when at least one corner is valid.
+        """
+        ht = np.asarray(heal_theta, dtype=float).ravel()
+        hp = np.asarray(heal_phi, dtype=float).ravel()
+        if ht.shape != hp.shape:
+            raise ValueError("heal_theta and heal_phi must have the same 1D shape (N,)")
+        N = ht.size
+
+        # Target unit vectors
+        Vtgt = self._sph_to_vec(ht, hp)  # (N,3)
+
+        # 1) Choose a seed node for each target (nearest source grid node on the sphere)
+        seed_flat = self._nearest_source_indices(Vtgt)
+        seed_j, seed_i = np.divmod(seed_flat, self.W)
+
+        # 2) For each target, test up to 4 candidate quads around the seed; pick the best
+        I = np.full((4, N), -1, dtype=np.int64)
+        W = np.zeros((4, N), dtype=float)
+        candidates = [(0, 0), (-1, 0), (0, -1), (-1, -1)]  # offsets for (i0,j0)
+
+        for n in range(N):
+            v = Vtgt[n]
+            best = None  # (score_tuple, s, t, i0, j0, (idx00, idx10, idx01, idx11))
+
+            for di, dj in candidates:
+                i0 = seed_i[n] + di
+                j0 = seed_j[n] + dj
+                if i0 < 0 or j0 < 0 or i0 + 1 >= self.W or j0 + 1 >= self.H:
+                    continue  # out of bounds
+
+                idx00 = j0 * self.W + i0
+                idx10 = j0 * self.W + (i0 + 1)
+                idx01 = (j0 + 1) * self.W + i0
+                idx11 = (j0 + 1) * self.W + (i0 + 1)
+
+                v00 = self._Vsrc[idx00]
+                v10 = self._Vsrc[idx10]
+                v01 = self._Vsrc[idx01]
+                v11 = self._Vsrc[idx11]
+
+                # Local tangent plane at the quad barycenter
+                vC = v00 + v10 + v01 + v11
+                vC /= np.linalg.norm(vC)
+                ex, ey, _ = self._tangent_axes_from_vec(vC)
+
+                # Project 4 corners + target onto (ex, ey)
+                P00 = np.array(self._project_to_plane(v00, ex, ey))
+                P10 = np.array(self._project_to_plane(v10, ex, ey))
+                P01 = np.array(self._project_to_plane(v01, ex, ey))
+                P11 = np.array(self._project_to_plane(v11, ex, ey))
+                P   = np.array(self._project_to_plane(v,   ex, ey))
+
+                # Invert bilinear mapping f(s,t) = P
+                s, t, ok, resid = self._invert_bilinear(P00, P10, P01, P11, P)
+
+                # Prefer in-bounds solutions; otherwise smallest residual
+                score = (0, resid) if ok else (1, resid)
+                if (best is None) or (score < best[0]):
+                    best = (score, s, t, i0, j0, (idx00, idx10, idx01, idx11))
+
+            if best is None:
+                continue  # leave weights at 0 and indices at -1
+
+            _, s, t, i0, j0, (idx00, idx10, idx01, idx11) = best
+
+            # Bilinear weights
+            w00 = (1.0 - s) * (1.0 - t)
+            w10 = s * (1.0 - t)
+            w01 = (1.0 - s) * t
+            w11 = s * t
+
+            I[:, n] = np.array([idx00, idx10, idx01, idx11], dtype=np.int64)
+            W[:, n] = np.array([w00, w10, w01, w11], dtype=float)
+
+            # Normalize for numerical safety
+            sW = W[:, n].sum()
+            if sW > 0:
+                W[:, n] /= sW
+
+        return I, W
+
+    def apply_weights(self, img: np.ndarray, I: np.ndarray, W: np.ndarray) -> np.ndarray:
+        """Apply precomputed (I, W) to a source image to obtain values at the HEALPix targets.
+
+        Parameters
+        ----------
+        img : np.ndarray, shape (H, W)
+            Source image values defined on the same grid as (src_theta, src_phi).
+        I : np.ndarray, shape (4, N), dtype=int64
+            Linear indices (row-major) of corner samples; -1 for invalid corners.
+        W : np.ndarray, shape (4, N), dtype=float64
+            Bilinear weights aligned with I.
+
+        Returns
+        -------
+        vals : np.ndarray, shape (N,)
+            Interpolated values at the target directions.
+        """
+        if img.shape != (self.H, self.W):
+            raise ValueError(f"img must have shape {(self.H, self.W)}, got {img.shape}")
+        img_flat = img.reshape(-1)
+        N = I.shape[1]
+        vals = np.zeros(N, dtype=float)
+        for k in range(4):
+            idx = I[k]
+            w = W[k]
+            m = idx >= 0
+            vals[m] += w[m] * img_flat[idx[m]]
+        return vals
+
+    # -----------------------------
+    # Internal helpers (geometry)
+    # -----------------------------
+    @staticmethod
+    def _sph_to_vec(theta: np.ndarray, phi: np.ndarray) -> np.ndarray:
+        """(theta, phi) -> unit vectors (x,y,z). theta=colat, phi=lon, radians."""
+        st, ct = np.sin(theta), np.cos(theta)
+        sp, cp = np.sin(phi), np.cos(phi)
+        x = st * cp
+        y = st * sp
+        z = ct
+        return np.stack([x, y, z], axis=-1)
+
+    @staticmethod
+    def _tangent_axes_from_vec(v: np.ndarray) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+        """Return an orthonormal basis (ex, ey, ez=v_hat) for the tangent plane at v."""
+        ez = v / np.linalg.norm(v)
+        a = np.array([1.0, 0.0, 0.0], dtype=float)
+        if abs(np.dot(a, ez)) > 0.9:  # avoid near-colinearity
+            a = np.array([0.0, 1.0, 0.0], dtype=float)
+        ex = a - np.dot(a, ez) * ez
+        ex /= np.linalg.norm(ex)
+        ey = np.cross(ez, ex)
+        return ex, ey, ez
+
+    @staticmethod
+    def _project_to_plane(vec: np.ndarray, ex: np.ndarray, ey: np.ndarray) -> Tuple[float, float]:
+        """Project 3D unit vector `vec` onto plane spanned by (ex, ey): returns (x, y)."""
+        return float(np.dot(vec, ex)), float(np.dot(vec, ey))
+
+    @staticmethod
+    def _bilinear_f(s: float, t: float, P00: np.ndarray, P10: np.ndarray, P01: np.ndarray, P11: np.ndarray) -> np.ndarray:
+        """Bilinear blend of 4 points in R^2."""
+        return ((1 - s) * (1 - t)) * P00 + (s * (1 - t)) * P10 + ((1 - s) * t) * P01 + (s * t) * P11
+
+    @staticmethod
+    def _bilinear_jacobian(s: float, t: float, P00: np.ndarray, P10: np.ndarray, P01: np.ndarray, P11: np.ndarray) -> np.ndarray:
+        """2x2 Jacobian of the bilinear map at (s,t)."""
+        dFds = (-(1 - t)) * P00 + (1 - t) * P10 + (-t) * P01 + t * P11
+        dFdt = (-(1 - s)) * P00 + (-s) * P10 + (1 - s) * P01 + s * P11
+        return np.stack([dFds, dFdt], axis=-1)
+
+    @classmethod
+    def _invert_bilinear(cls, P00: np.ndarray, P10: np.ndarray, P01: np.ndarray, P11: np.ndarray, P: np.ndarray,
+                         max_iter: int = 10, tol: float = 1e-9) -> Tuple[float, float, bool, float]:
+        """Invert the bilinear map f(s,t)=P with a Newton loop; return (s,t, ok, residual)."""
+        # Initial guess from parallelogram (ignore cross term)
+        A = np.column_stack([P10 - P00, P01 - P00])  # 2x2
+        b = P - P00
+        try:
+            st0 = np.linalg.lstsq(A, b, rcond=None)[0]
+            s, t = float(st0[0]), float(st0[1])
+        except np.linalg.LinAlgError:  # fallback
+            s, t = 0.5, 0.5
+
+        for _ in range(max_iter):
+            F = cls._bilinear_f(s, t, P00, P10, P01, P11)
+            r = P - F
+            if np.linalg.norm(r) < tol:
+                break
+            J = cls._bilinear_jacobian(s, t, P00, P10, P01, P11)
+            try:
+                delta = np.linalg.solve(J, r)
+            except np.linalg.LinAlgError:
+                break
+            s += float(delta[0])
+            t += float(delta[1])
+
+        # Clamp to [0,1] softly and compute residual
+        s_c = min(max(s, 0.0), 1.0)
+        t_c = min(max(t, 0.0), 1.0)
+        F_end = cls._bilinear_f(s_c, t_c, P00, P10, P01, P11)
+        resid = float(np.linalg.norm(P - F_end))
+        ok = (0.0 <= s <= 1.0) and (0.0 <= t <= 1.0)
+        return s_c, t_c, ok, resid
+
+    # -----------------------------
+    # Internal helpers (search)
+    # -----------------------------
+    def _nearest_source_indices(self, Vtgt: np.ndarray) -> np.ndarray:
+        """Return flat indices of nearest source nodes for each target unit vector."""
+        if self._kdtree is not None:  # fast path
+            _, nn = self._kdtree.query(Vtgt, k=1)
+            return nn.astype(np.int64)
+        # Brute-force in blocks to limit memory
+        N = Vtgt.shape[0]
+        out = np.empty(N, dtype=np.int64)
+        block = 20000
+        VsrcT = self._Vsrc.T  # (3, H*W)
+        for start in range(0, N, block):
+            end = min(N, start + block)
+            D = Vtgt[start:end] @ VsrcT  # cosine similarities
+            out[start:end] = np.argmax(D, axis=1)
+        return out
+
+
+__all__ = ["HealBili"]

foscat/Plot.py

@@ -0,0 +1,328 @@
+import numpy as np
+import matplotlib.pyplot as plt
+import healpy as hp
+
+def lgnomproject(
+    cell_ids,               # array-like (N,), HEALPix pixel indices of your samples
+    data,                   # array-like (N,), values per cell id
+    nside: int,
+    rot=None,               # (lon0_deg, lat0_deg, psi_deg). If None: auto-center from cell_ids (pix centers)
+    xsize: int = 800,
+    ysize: int = 800,
+    reso: float = None,     # deg/pixel on tangent plane; if None, use fov_deg
+    fov_deg=10.0,           # full FoV deg (scalar or (fx,fy))
+    nest: bool = True,      # True if your cell_ids are NESTED (and ang2pix to be done in NEST)
+    reduce: str = "mean",   # 'mean'|'median'|'sum'|'first' when duplicates in cell_ids
+    mask_outside: bool = True,
+    unseen_value=None,      # default to hp.UNSEEN
+    return_image_only: bool = False,
+    title: str = None, cmap: str = "viridis", vmin=None, vmax=None,
+    notext: bool = False,    # True to avoid tick marks
+    hold: bool = True,       # create a new figure if True otherwise use sub to split panel
+    sub=(1,1,1),             # declare sub plot
+    cbar: bool = False,      # plot colorbar if True
+
+):
+    """
+    Gnomonic projection from *sparse* HEALPix samples (cell_ids, data) to an image (ysize, xsize).
+
+    For each output image pixel (i,j):
+      plane (x,y) --inverse gnomonic--> (theta, phi) --HEALPix--> ipix
+      if ipix in `cell_ids`: assign aggregated value, else UNSEEN.
+
+    Parameters
+    ----------
+    cell_ids : (N,) int array
+        HEALPix pixel indices of your samples. Must correspond to `nside` and `nest`.
+    data : (N,) float array
+        Sample values for each cell id.
+    nside : int
+        HEALPix NSIDE used for both `cell_ids` and the image reprojection.
+    rot : (lon0_deg, lat0_deg, psi_deg) or None
+        Gnomonic center (lon, lat) and in-plane rotation psi (deg).
+        If None, we auto-center from the *centers of the provided pixels* (via hp.pix2ang).
+    xsize, ysize : int
+        Output image size (pixels).
+    reso : float or None
+        Pixel size (deg/pixel) on the tangent plane. If None, derived from `fov_deg`.
+    fov_deg : float or (float,float)
+        Full field of view in degrees.
+    nest : bool
+        Use True if your `cell_ids` correspond to NESTED indexing.
+    reduce : str
+        How to combine duplicate cell ids: 'mean'|'median'|'sum'|'first'.
+    mask_outside : bool
+        Mask pixels outside the valid gnomonic hemisphere (cosc <= 0).
+    unseen_value : float or None
+        Value for invalid pixels (default hp.UNSEEN).
+    
+    return_image_only : bool
+        If True, return the 2D array only (no plotting).
+
+    Returns
+    -------
+    (fig, ax, img) or img
+        If return_image_only=True, returns img (ysize, xsize).
+    """
+    if unseen_value is None:
+        unseen_value = hp.UNSEEN
+
+    cell_ids = np.asarray(cell_ids, dtype=np.int64)
+    vals_in  = np.asarray(data, dtype=float)
+    if cell_ids.shape != vals_in.shape:
+        raise ValueError("cell_ids and data must have the same shape (N,)")
+
+    # -------- 1) Aggregate duplicates in cell_ids (if any) --------
+    uniq, inv = np.unique(cell_ids, return_inverse=True)  # uniq is sorted
+    if reduce == "first":
+        first_idx = np.full(uniq.size, -1, dtype=np.int64)
+        for i, g in enumerate(inv):
+            if first_idx[g] < 0:
+                first_idx[g] = i
+        agg_vals = vals_in[first_idx]
+    elif reduce == "sum":
+        agg_vals = np.zeros(uniq.size, dtype=float)
+        np.add.at(agg_vals, inv, vals_in)
+    elif reduce == "median":
+        agg_vals = np.empty(uniq.size, dtype=float)
+        for k, pix in enumerate(uniq):
+            agg_vals[k] = np.median(vals_in[cell_ids == pix])
+    elif reduce == "mean":
+        sums = np.zeros(uniq.size, dtype=float)
+        cnts = np.zeros(uniq.size, dtype=float)
+        np.add.at(sums, inv, vals_in)
+        np.add.at(cnts, inv, 1.0)
+        agg_vals = sums / np.maximum(cnts, 1.0)
+    else:
+        raise ValueError("reduce must be one of {'mean','median','sum','first'}")
+
+    # -------- 2) Choose gnomonic center (rot) --------
+    if rot is None:
+        # Center from pixel centers of provided ids
+        theta_c, phi_c = hp.pix2ang(nside, uniq, nest=nest)  # colat, lon (rad)
+        # circular mean for lon, median for colat
+        lon0_deg = np.degrees(np.angle(np.mean(np.exp(1j * phi_c))))
+        lat0_deg = 90.0 - np.degrees(np.median(theta_c))
+        psi_deg  = 0.0
+        rot = (lon0_deg % 360.0, float(lat0_deg), float(psi_deg))
+
+    lon0_deg, lat0_deg, psi_deg = rot
+    lon0 = np.deg2rad(lon0_deg)
+    lat0 = np.deg2rad(lat0_deg)
+    psi  = np.deg2rad(psi_deg)
+
+    # -------- 3) Tangent-plane grid (gnomonic) --------
+    if reso is not None:
+        dx = np.tan(np.deg2rad(reso))
+        dy = dx
+        half_x = 0.5 * xsize * dx
+        half_y = 0.5 * ysize * dy
+    else:
+        if np.isscalar(fov_deg):
+            fx, fy = float(fov_deg), float(fov_deg)
+        else:
+            fx, fy = float(fov_deg[0]), float(fov_deg[1])
+        ax = np.deg2rad(0.5 * fx)
+        ay = np.deg2rad(0.5 * fy)
+        half_x = np.tan(ax)
+        half_y = np.tan(ay)
+    
+    xs = np.linspace(-half_x, +half_x, xsize, endpoint=False) + (half_x / xsize)
+    ys = np.linspace(-half_y, +half_y, ysize, endpoint=False) + (half_y / ysize)
+    
+    X, Y = np.meshgrid(xs, ys)  # (ysize, xsize)
+
+    # in-plane rotation psi
+    c, s = np.cos(psi), np.sin(psi)
+    Xr =  c * X + s * Y
+    Yr = -s * X + c * Y
+
+    # -------- 4) Inverse gnomonic → sphere --------
+    rho  = np.hypot(Xr, Yr)
+    cang = np.arctan(rho)
+    sinc, cosc = np.sin(cang), np.cos(cang)
+    sinlat0, coslat0 = np.sin(lat0), np.cos(lat0)
+
+    with np.errstate(invalid="ignore", divide="ignore"):
+        lat = np.arcsin(cosc * sinlat0 + (Yr * sinc * coslat0) / np.where(rho == 0, 1.0, rho))
+        lon = lon0 + np.arctan2(Xr * sinc, rho * coslat0 * cosc - Yr * sinlat0 * sinc)
+
+    lon = (lon + 2*np.pi) % (2*np.pi)
+    theta_img = (np.pi / 2.0) - lat
+    outside = (cosc <= 0.0) if mask_outside else np.zeros_like(cosc, dtype=bool)
+
+    # -------- 5) Map image pixels to HEALPix ids --------
+    ip_img = hp.ang2pix(nside, theta_img.ravel(), lon.ravel(), nest=nest).astype(np.int64)
+
+    # -------- 6) Assign values by matching ip_img ∈ uniq (safe searchsorted) --------
+    # uniq is sorted; build insertion pos then check matches only where pos < len(uniq)
+    pos = np.searchsorted(uniq, ip_img, side="left")
+    valid = pos < uniq.size
+    match = np.zeros_like(valid, dtype=bool)
+    match[valid] = (uniq[pos[valid]] == ip_img[valid])
+
+    img_flat = np.full(ip_img.shape, unseen_value, dtype=float)
+    img_flat[match] = agg_vals[pos[match]]
+    img = img_flat.reshape(ysize, xsize)
+
+    # Mask out-of-hemisphere gnomonic region
+    if mask_outside:
+        img[outside] = unseen_value
+
+    # -------- 7) Return / plot --------
+    if return_image_only:
+        return img
+
+    # Axes in approx. "gnomonic degrees" (atan of plane coords)
+    x_deg = np.degrees(np.arctan(xs))
+    y_deg = np.degrees(np.arctan(ys))
+    
+    longitude_min=x_deg[0]/np.cos(np.deg2rad(lat0_deg))+lon0_deg
+    longitude_max=x_deg[-1]/np.cos(np.deg2rad(lat0_deg))+lon0_deg
+    
+    if longitude_min>180:
+        longitude_min-=360
+        longitude_max-=360
+    
+    extent = (longitude_min,longitude_max,
+              y_deg[0]+lat0_deg, y_deg[-1]+lat0_deg)
+
+    if hold:
+        fig, ax = plt.subplots(figsize=(xsize/100, ysize/100), dpi=100)
+    else:
+        ax=plt.subplot(sub[0],sub[1],sub[2])
+    
+    im = ax.imshow(
+        np.where(img == unseen_value, np.nan, img),
+        origin="lower",
+        extent=extent,
+        cmap=cmap,
+        vmin=vmin, vmax=vmax,
+        interpolation="nearest",
+        aspect="equal"
+    )
+    if not notext:
+        ax.set_xlabel("Longitude (deg)")
+        ax.set_ylabel("Latitude (deg)")
+    else:
+        ax.set_xticks([])
+        ax.set_yticks([])
+        
+    if title:
+        ax.set_title(title)
+        
+    if cbar:
+        if hold:
+            cb = fig.colorbar(im, ax=ax)
+            cb.set_label("value")
+        else:
+            plt.colorbar(im, ax=ax, orientation="vertical", label="value")
+        
+    plt.tight_layout()
+    if hold:
+        return fig, ax #, img
+    else:
+        return ax
+
+
+def plot_scat(s1,s2,s3,s4):
+
+    if not isinstance(s1,np.ndarray): # manage if Torch tensor
+        S1=s1.cpu().numpy()
+        S2=s2.cpu().numpy()
+        S3=s3.cpu().numpy()
+        S4=s4.cpu().numpy()
+    else:
+        S1=s1
+        S2=s2
+        S3=s3
+        S4=s4
+        
+    N_image=s1.shape[0]
+    J=s1.shape[1]
+    N_orient=s1.shape[2]
+
+    # compute index j1 and j2 for S3
+    j1_s3=np.zeros([s3.shape[1]],dtype='int')
+    j2_s3=np.zeros([s3.shape[1]],dtype='int')
+    
+    # compute index j1 and j2 for S4
+    j1_s4=np.zeros([s4.shape[1]],dtype='int')
+    j2_s4=np.zeros([s4.shape[1]],dtype='int')
+    j3_s4=np.zeros([s4.shape[1]],dtype='int')
+
+    n_s3=0
+    n_s4=0
+    for j3 in range(0,J):
+        for j2 in range(0, j3 + 1):
+            j1_s3[n_s3]=j2
+            j2_s3[n_s3]=j3
+            n_s3+=1
+            for j1 in range(0, j2 + 1):
+                j1_s4[n_s4]=j1
+                j2_s4[n_s4]=j2
+                j3_s4[n_s4]=j3
+                n_s4+=1
+
+
+    color=['b','r','orange','pink']
+    symbol=['',':','-','.']
+    plt.figure(figsize=(16,12))
+    
+    plt.subplot(2,2,1)
+    for k in range(4):
+        plt.plot(S1[0,:,k],color=color[k%len(color)],label=r'$\Theta = %d$'%(k))
+    plt.legend(frameon=0,ncol=2)
+    plt.xlabel(r'$J_1$')
+    plt.ylabel(r'$S_1$')
+    plt.yscale('log')
+    
+    plt.subplot(2,2,2)
+    for k in range(4):
+        plt.plot(S2[0,:,k],color=color[k%len(color)],label=r'$\Theta = %d$'%(k))
+    plt.xlabel(r'$J_1$')
+    plt.ylabel(r'$S_2$')
+    plt.yscale('log')
+    
+    plt.subplot(2,2,3)
+    nidx=np.concatenate([np.zeros([1]),np.cumsum(np.bincount(j1_s3))],0)
+    l_pos=[]
+    l_name=[]
+    for i in np.unique(j1_s3):
+        idx=np.where(j1_s3==i)[0]
+        for k in range(4):
+            for l in range(4):
+                if i==0:
+                    plt.plot(j2_s3[idx]+nidx[i],S3[0,idx,k,l],symbol[l%len(symbol)],color=color[k%len(color)],label=r'$\Theta = %d,%d$'%(k,l))
+                else:
+                    plt.plot(j2_s3[idx]+nidx[i],S3[0,idx,k,l],symbol[l%len(symbol)],color=color[k%len(color)])
+        l_pos=l_pos+list(j2_s3[idx]+nidx[i])
+        l_name=l_name+["%d,%d"%(j1_s3[m],j2_s3[m]) for m in idx]
+    plt.legend(frameon=0,ncol=2)
+    
+    plt.xticks(l_pos,l_name, fontsize=6)
+    plt.xlabel(r"$j_{1},j_{2}$", fontsize=9)
+    plt.ylabel(r"$S_{3}$", fontsize=9)
+    
+    plt.subplot(2,2,4)
+    nidx=0
+    l_pos=[]
+    l_name=[]
+    for i in np.unique(j1_s4):
+        for j in np.unique(j2_s4):
+            idx=np.where((j1_s4==i)*(j2_s4==j))[0]
+            for k in range(4):
+                for l in range(4):
+                    for m in range(4):
+                        if i==0 and j==0 and m==0:
+                            plt.plot(j2_s4[idx]+j3_s4[idx]+nidx,S4[0,idx,k,l,m],symbol[l%len(symbol)],color=color[k%len(color)],label=r'$\Theta = %d,%d,%d$'%(k,l,m))
+                        else:
+                            plt.plot(j2_s4[idx]+j3_s4[idx]+nidx,S4[0,idx,k,l,m],symbol[l%len(symbol)],color=color[k%len(color)])
+            l_pos=l_pos+list(j2_s4[idx]+j3_s4[idx]+nidx)
+            l_name=l_name+["%d,%d,%d"%(j1_s4[m],j2_s4[m],j3_s4[m]) for m in idx]
+        nidx+=np.max(j2_s4[j1_s4==i]+j3_s4[j1_s4==i])-np.min(j2_s4[j1_s4==i]+j3_s4[j1_s4==i])+1
+    plt.legend(frameon=0,ncol=2)
+    
+    plt.xticks(l_pos,l_name, fontsize=6, rotation=90)
+    plt.xlabel(r"$j_{1},j_{2},j_{3}$", fontsize=9)
+    plt.ylabel(r"$S_{4}$", fontsize=9)